[{"data":1,"prerenderedAt":2154},["ShallowReactive",2],{"blog-post-agents-md-best-practices":3,"related-posts-agents-md-best-practices":808},{"id":4,"title":5,"author":6,"body":10,"category":788,"date":789,"description":790,"extension":452,"featured":791,"image":792,"imageHeight":793,"imageWidth":793,"meta":794,"navigation":230,"path":795,"readingTime":796,"seo":797,"seoTitle":798,"stem":799,"tags":800,"updatedDate":789,"__hash__":807},"blog/blog/agents-md-best-practices.md","AGENTS.md Best Practices: Write the File That Makes Your Agent Actually Follow Instructions",{"name":7,"role":8,"avatar":9},"Shabnam Katoch","Growth Head","/img/avatars/shabnam-profile.jpeg",{"type":11,"value":12,"toc":770},"minimark",[13,17,20,23,26,29,32,35,38,43,46,58,61,64,71,77,80,83,86,91,94,112,115,120,123,128,131,136,139,144,160,164,170,175,178,181,184,188,191,201,204,325,328,332,335,338,342,345,351,357,367,373,376,379,383,386,389,395,435,438,441,445,448,703,706,709,715,718,721,724,728,732,735,739,745,749,752,756,759,763,766],[14,15,16],"p",{},"60,000+ repos have one. 30+ AI tools read it. But most AGENTS.md files are either too long, too vague, or generated by an LLM (which makes your agent worse). Here's how to write one that works.",[14,18,19],{},"Our coding agent kept generating class components. The entire codebase was functional React with hooks. The README said \"React 19, functional components.\" The contributing guide explained it in detail.",[14,21,22],{},"The agent didn't read the README. It doesn't know to.",[14,24,25],{},"Then we added an AGENTS.md file with one line in the style section: Always use functional components with hooks. Never use class components. Plus one code example showing the pattern.",[14,27,28],{},"The problem disappeared instantly.",[14,30,31],{},"That's what AGENTS.md does. It's a Markdown file at the root of your repository that gives AI agents the project-specific context they need to work correctly. Not documentation for humans. Instructions for machines. And as of 2026, it's read natively by 30+ AI tools including Claude Code, GitHub Copilot, Cursor, OpenAI Codex, Gemini CLI, Windsurf, Devin, Aider, and Amazon Q.",[14,33,34],{},"The convention has been adopted by over 60,000 repositories and is now stewarded by the Agentic AI Foundation under the Linux Foundation. If you build with AI agents and don't have an AGENTS.md, your agents are guessing at conventions they should know.",[14,36,37],{},"Here's how to write one that actually improves agent behavior instead of wasting context tokens.",[39,40,42],"h2",{"id":41},"what-goes-in-an-agentsmd-and-what-doesnt","What goes in an AGENTS.md (and what doesn't)",[14,44,45],{},"The common mistake is treating AGENTS.md like documentation. Long explanations. Architecture philosophy. Design rationale. History of the project.",[14,47,48,49,53,54,57],{},"Your agent doesn't need to know why you chose Next.js over Remix. It needs to know that you use Next.js 15 with App Router, and the test command is ",[50,51,52],"code",{},"pnpm test",", not ",[50,55,56],{},"npm test",".",[14,59,60],{},"Research across 2,500+ repositories found that AGENTS.md files over 150 lines deliver diminishing returns and can increase inference costs 20-23% without improving agent performance. The technical maximum is 32 KiB, but aim for under 150 lines.",[14,62,63],{},"Here are the sections that matter, in the order they should appear:",[14,65,66],{},[67,68],"img",{"alt":69,"src":70},"The AGENTS.md nutrition label: serving size under 150 lines, with project stack, build and test commands, code style with examples, architecture constraints, boundaries, and a short git workflow section.","/img/blog/agents-md-best-practices-nutrition-label.jpg",[14,72,73],{},[74,75,76],"strong",{},"Project stack (5-10 lines)",[14,78,79],{},"Language, framework, major dependencies, runtime. Be specific.",[14,81,82],{},"Good: Next.js 15 App Router, React 19, TypeScript 5.4, Tailwind CSS, Drizzle ORM, Bun",[14,84,85],{},"Bad: This is a modern web application built with React and various supporting libraries.",[14,87,88],{},[74,89,90],{},"Build and test commands (5-10 lines)",[14,92,93],{},"Exact commands with exact flags. Package manager matters. Build targets matter.",[14,95,96,97,100,101,104,105,108,109],{},"Good: Build: ",[50,98,99],{},"bun run build",". Test: ",[50,102,103],{},"bun test",". Lint: ",[50,106,107],{},"bun run lint --fix",". Deploy: ",[50,110,111],{},"bun run deploy:staging",[14,113,114],{},"Bad: Run the standard build and test commands as described in package.json.",[14,116,117],{},[74,118,119],{},"Code style conventions (10-20 lines, with examples)",[14,121,122],{},"One code example per convention beats three paragraphs describing it. This is the single most important insight from studying effective AGENTS.md files. Show, don't tell.",[14,124,125],{},[74,126,127],{},"Architecture constraints (5-15 lines)",[14,129,130],{},"Directory structure rules. Module boundaries. Import restrictions. Data flow patterns.",[14,132,133],{},[74,134,135],{},"Boundaries (5-10 lines)",[14,137,138],{},"What the agent should never touch. \"Never commit secrets\" was the most common helpful constraint across the 2,500-repository analysis. Add: generated files, legacy modules, configuration files that shouldn't change, directories managed by other tools.",[14,140,141],{},[74,142,143],{},"Git workflow (5 lines)",[14,145,146,147,150,151,150,154,150,157,57],{},"Branch naming, commit message format, PR conventions. Squash merge only. Conventional commits: ",[50,148,149],{},"feat:",", ",[50,152,153],{},"fix:",[50,155,156],{},"chore:",[50,158,159],{},"docs:",[39,161,163],{"id":162},"the-three-mistakes-that-make-agentsmd-useless","The three mistakes that make AGENTS.md useless",[14,165,166],{},[67,167],{"alt":168,"src":169},"The AGENTS.md hall of shame: the three mistakes that make a file useless — letting an LLM generate it, describing conventions instead of showing them with code, and not updating it.","/img/blog/agents-md-best-practices-hall-of-shame.jpg",[171,172,174],"h3",{"id":173},"mistake-1-letting-an-llm-generate-it","Mistake 1: Letting an LLM generate it",[14,176,177],{},"This is the most common mistake and the most damaging. Research found that LLM-generated AGENTS.md files reduce task success in 5 out of 8 tested settings and add 2.45-3.92 extra steps per task.",[14,179,180],{},"Why? LLMs generate generic, verbose instructions that state the obvious. \"Follow best practices for error handling.\" \"Write clean, maintainable code.\" \"Use descriptive variable names.\" Your agent already knows this. You're burning context tokens on advice the model was trained to follow by default.",[14,182,183],{},"Write your AGENTS.md by hand. Include only the conventions that are specific to YOUR project and that an agent wouldn't know without being told.",[171,185,187],{"id":186},"mistake-2-describing-conventions-instead-of-showing-them","Mistake 2: Describing conventions instead of showing them",[14,189,190],{},"Bad:",[192,193,198],"pre",{"className":194,"code":196,"language":197},[195],"language-text","Use our custom error handling approach where errors bubble up through\nthe service layer and are caught by the global error handler rather\nthan being caught at individual function level.\n","text",[50,199,196],{"__ignoreMap":200},"",[14,202,203],{},"Good:",[192,205,209],{"className":206,"code":207,"language":208,"meta":200,"style":200},"language-js shiki shiki-themes github-light","// Error handling: Let errors propagate. Do not wrap individual calls\n// in try/catch. The global handler catches everything.\n\n// CORRECT\nasync function getUser(id) {\n  const user = await db.users.findById(id);\n  return user;\n}\n\n// WRONG\nasync function getUser(id) {\n  try {\n    const user = await db.users.findById(id);\n    return user;\n  } catch (e) {\n    logger.error(e);\n    throw e;\n  }\n}\n","js",[50,210,211,219,225,232,238,244,250,256,262,267,273,278,284,290,296,302,308,314,320],{"__ignoreMap":200},[212,213,216],"span",{"class":214,"line":215},"line",1,[212,217,218],{},"// Error handling: Let errors propagate. Do not wrap individual calls\n",[212,220,222],{"class":214,"line":221},2,[212,223,224],{},"// in try/catch. The global handler catches everything.\n",[212,226,228],{"class":214,"line":227},3,[212,229,231],{"emptyLinePlaceholder":230},true,"\n",[212,233,235],{"class":214,"line":234},4,[212,236,237],{},"// CORRECT\n",[212,239,241],{"class":214,"line":240},5,[212,242,243],{},"async function getUser(id) {\n",[212,245,247],{"class":214,"line":246},6,[212,248,249],{},"  const user = await db.users.findById(id);\n",[212,251,253],{"class":214,"line":252},7,[212,254,255],{},"  return user;\n",[212,257,259],{"class":214,"line":258},8,[212,260,261],{},"}\n",[212,263,265],{"class":214,"line":264},9,[212,266,231],{"emptyLinePlaceholder":230},[212,268,270],{"class":214,"line":269},10,[212,271,272],{},"// WRONG\n",[212,274,276],{"class":214,"line":275},11,[212,277,243],{},[212,279,281],{"class":214,"line":280},12,[212,282,283],{},"  try {\n",[212,285,287],{"class":214,"line":286},13,[212,288,289],{},"    const user = await db.users.findById(id);\n",[212,291,293],{"class":214,"line":292},14,[212,294,295],{},"    return user;\n",[212,297,299],{"class":214,"line":298},15,[212,300,301],{},"  } catch (e) {\n",[212,303,305],{"class":214,"line":304},16,[212,306,307],{},"    logger.error(e);\n",[212,309,311],{"class":214,"line":310},17,[212,312,313],{},"    throw e;\n",[212,315,317],{"class":214,"line":316},18,[212,318,319],{},"  }\n",[212,321,323],{"class":214,"line":322},19,[212,324,261],{},[14,326,327],{},"The code example communicates instantly what three paragraphs of description fail to convey.",[171,329,331],{"id":330},"mistake-3-not-updating-it","Mistake 3: Not updating it",[14,333,334],{},"An AGENTS.md with stale table names, deprecated commands, or outdated architecture constraints is worse than no AGENTS.md. It actively misleads the agent.",[14,336,337],{},"Update your AGENTS.md in the same PR where conventions change. Set a quarterly review to remove stale guidance. A useful signal: when an agent repeatedly produces incorrect output, that's a gap in your AGENTS.md.",[39,339,341],{"id":340},"agentsmd-vs-claudemd-vs-cursorrules-which-do-you-need","AGENTS.md vs CLAUDE.md vs .cursorrules (which do you need?)",[14,343,344],{},"Here's the honest answer: start with AGENTS.md, add tool-specific files only if you need their unique features.",[14,346,347],{},[67,348],{"alt":349,"src":350},"The universal file and its specialist cousins: AGENTS.md is the universal core read by 30+ tools, while CLAUDE.md adds @imports for Claude Code and .cursorrules adds glob-based rules for Cursor.","/img/blog/agents-md-best-practices-file-comparison.jpg",[14,352,353,356],{},[74,354,355],{},"AGENTS.md"," has the broadest compatibility. 30+ tools read it natively. It's the universal format.",[14,358,359,362,363,366],{},[74,360,361],{},"CLAUDE.md"," is Claude Code-specific. Its unique feature is ",[50,364,365],{},"@imports"," that compose instructions from multiple files. If your team uses Claude Code exclusively and needs modular instruction files, add CLAUDE.md alongside AGENTS.md.",[14,368,369,372],{},[74,370,371],{},".cursorrules"," is Cursor-specific. Its unique feature is MDC frontmatter with glob patterns that activate different instructions for different file types. If your team uses Cursor and needs per-filetype rules, add .cursorrules.",[14,374,375],{},"90% of the content across all three files is identical. Build commands, architecture rules, testing conventions, and boundaries don't change per tool. A converter tool called rule-porter can translate between formats if needed.",[14,377,378],{},"Write AGENTS.md first. It's the single source of truth that every tool reads. Add CLAUDE.md or .cursorrules only for features that AGENTS.md can't express.",[39,380,382],{"id":381},"beyond-coding-agentsmd-for-business-agents","Beyond coding: AGENTS.md for business agents",[14,384,385],{},"Here's where things get interesting. The AGENTS.md convention was born in coding repositories. But the same principle applies to any AI agent: give the agent structured, project-specific context at session start, and it performs better.",[14,387,388],{},"For business agents (support, sales, operations), the equivalent configuration includes:",[14,390,391],{},[67,392],{"alt":393,"src":394},"Same passport, two different languages: a coding agent's AGENTS.md (stack, commands, code style, boundaries) maps to a business agent's config (identity, capabilities, constraints, trust level, escalation, output format).","/img/blog/agents-md-best-practices-business-agents.jpg",[396,397,398,405,411,417,423,429],"ul",{},[399,400,401,404],"li",{},[74,402,403],{},"Identity:"," Who the agent is. Name, role, company, tone of voice.",[399,406,407,410],{},[74,408,409],{},"Capabilities:"," What the agent can do. Which tools it has access to. Which integrations it can call.",[399,412,413,416],{},[74,414,415],{},"Constraints:"," What the agent cannot do. Actions that require human approval. Topics it should escalate. Data it should never share.",[399,418,419,422],{},[74,420,421],{},"Trust level:"," How much autonomy the agent has. Draft-only? Execute with approval? Fully autonomous within boundaries?",[399,424,425,428],{},[74,426,427],{},"Escalation rules:"," When to hand off to a human. Sentiment triggers, confidence thresholds, sensitive categories.",[399,430,431,434],{},[74,432,433],{},"Output format:"," How the agent should respond. Structured JSON? Natural language? Specific templates?",[14,436,437],{},"For coding agents, AGENTS.md is a file you write and maintain manually. For business agents, this configuration should be visual and managed through a platform.",[14,439,440],{},"This is exactly how BetterClaw's agent builder works. Every field that would go into a business AGENTS.md is a visual input in the builder. Identity, capabilities, constraints, trust levels (Intern, Specialist, Lead), escalation rules, output format. Version-controlled by the platform. No file management. No forgetting to update. Free plan with every feature. $19/month per agent on Pro. BYOK with zero markup.",[39,442,444],{"id":443},"the-annotated-template-copy-and-adapt","The annotated template (copy and adapt)",[14,446,447],{},"Here's a minimal, effective AGENTS.md template. Delete sections that don't apply. A shorter, accurate file outperforms a comprehensive, generic one.",[192,449,453],{"className":450,"code":451,"language":452,"meta":200,"style":200},"language-md shiki shiki-themes github-light","# Project Name\n\nNext.js 15 App Router, React 19, TypeScript 5.4, Tailwind CSS, Drizzle ORM, Bun.\n\n## Commands\n\nBuild: `bun run build`\nTest: `bun test`\nLint: `bun run lint --fix`\nSingle test: `bun test path/to/file.test.ts`\n\n## Code Style\n\nFunctional components only. Never class components.\nUse `const` exclusively. Never `var`, never `let` unless reassignment is needed.\nNamed exports only. Never default exports.\n\n// Component pattern:\nexport const UserCard = ({ name, email }: UserCardProps) => {\n  return \u003Cdiv className=\"p-4\">{name}\u003C/div>;\n};\n\n## Error Handling\n\nLet errors propagate. Do not wrap individual calls in try/catch.\nThe global error handler in middleware.ts catches everything.\n\n## Architecture\n\n/app         -> Routes and page components\n/components  -> Shared UI components\n/lib         -> Business logic and utilities\n/db          -> Database schema and migrations\n\nNever import from /app into /lib. Data flows one direction.\n\n## Boundaries\n\nNever modify files in /generated/.\nNever commit .env or any file containing secrets.\nThe /legacy/ module uses sync patterns. Do not convert to async.\n\n## Git\n\nSquash merge only.\nConventional commits: feat:, fix:, chore:, docs:.\nBranch format: type/short-description (e.g., feat/user-auth).\n","md",[50,454,455,460,464,469,473,478,482,487,492,497,502,506,511,515,520,525,530,534,539,544,550,556,561,567,572,578,584,589,595,600,606,612,618,624,629,635,640,646,651,657,663,669,674,680,685,691,697],{"__ignoreMap":200},[212,456,457],{"class":214,"line":215},[212,458,459],{},"# Project Name\n",[212,461,462],{"class":214,"line":221},[212,463,231],{"emptyLinePlaceholder":230},[212,465,466],{"class":214,"line":227},[212,467,468],{},"Next.js 15 App Router, React 19, TypeScript 5.4, Tailwind CSS, Drizzle ORM, Bun.\n",[212,470,471],{"class":214,"line":234},[212,472,231],{"emptyLinePlaceholder":230},[212,474,475],{"class":214,"line":240},[212,476,477],{},"## Commands\n",[212,479,480],{"class":214,"line":246},[212,481,231],{"emptyLinePlaceholder":230},[212,483,484],{"class":214,"line":252},[212,485,486],{},"Build: `bun run build`\n",[212,488,489],{"class":214,"line":258},[212,490,491],{},"Test: `bun test`\n",[212,493,494],{"class":214,"line":264},[212,495,496],{},"Lint: `bun run lint --fix`\n",[212,498,499],{"class":214,"line":269},[212,500,501],{},"Single test: `bun test path/to/file.test.ts`\n",[212,503,504],{"class":214,"line":275},[212,505,231],{"emptyLinePlaceholder":230},[212,507,508],{"class":214,"line":280},[212,509,510],{},"## Code Style\n",[212,512,513],{"class":214,"line":286},[212,514,231],{"emptyLinePlaceholder":230},[212,516,517],{"class":214,"line":292},[212,518,519],{},"Functional components only. Never class components.\n",[212,521,522],{"class":214,"line":298},[212,523,524],{},"Use `const` exclusively. Never `var`, never `let` unless reassignment is needed.\n",[212,526,527],{"class":214,"line":304},[212,528,529],{},"Named exports only. Never default exports.\n",[212,531,532],{"class":214,"line":310},[212,533,231],{"emptyLinePlaceholder":230},[212,535,536],{"class":214,"line":316},[212,537,538],{},"// Component pattern:\n",[212,540,541],{"class":214,"line":322},[212,542,543],{},"export const UserCard = ({ name, email }: UserCardProps) => {\n",[212,545,547],{"class":214,"line":546},20,[212,548,549],{},"  return \u003Cdiv className=\"p-4\">{name}\u003C/div>;\n",[212,551,553],{"class":214,"line":552},21,[212,554,555],{},"};\n",[212,557,559],{"class":214,"line":558},22,[212,560,231],{"emptyLinePlaceholder":230},[212,562,564],{"class":214,"line":563},23,[212,565,566],{},"## Error Handling\n",[212,568,570],{"class":214,"line":569},24,[212,571,231],{"emptyLinePlaceholder":230},[212,573,575],{"class":214,"line":574},25,[212,576,577],{},"Let errors propagate. Do not wrap individual calls in try/catch.\n",[212,579,581],{"class":214,"line":580},26,[212,582,583],{},"The global error handler in middleware.ts catches everything.\n",[212,585,587],{"class":214,"line":586},27,[212,588,231],{"emptyLinePlaceholder":230},[212,590,592],{"class":214,"line":591},28,[212,593,594],{},"## Architecture\n",[212,596,598],{"class":214,"line":597},29,[212,599,231],{"emptyLinePlaceholder":230},[212,601,603],{"class":214,"line":602},30,[212,604,605],{},"/app         -> Routes and page components\n",[212,607,609],{"class":214,"line":608},31,[212,610,611],{},"/components  -> Shared UI components\n",[212,613,615],{"class":214,"line":614},32,[212,616,617],{},"/lib         -> Business logic and utilities\n",[212,619,621],{"class":214,"line":620},33,[212,622,623],{},"/db          -> Database schema and migrations\n",[212,625,627],{"class":214,"line":626},34,[212,628,231],{"emptyLinePlaceholder":230},[212,630,632],{"class":214,"line":631},35,[212,633,634],{},"Never import from /app into /lib. Data flows one direction.\n",[212,636,638],{"class":214,"line":637},36,[212,639,231],{"emptyLinePlaceholder":230},[212,641,643],{"class":214,"line":642},37,[212,644,645],{},"## Boundaries\n",[212,647,649],{"class":214,"line":648},38,[212,650,231],{"emptyLinePlaceholder":230},[212,652,654],{"class":214,"line":653},39,[212,655,656],{},"Never modify files in /generated/.\n",[212,658,660],{"class":214,"line":659},40,[212,661,662],{},"Never commit .env or any file containing secrets.\n",[212,664,666],{"class":214,"line":665},41,[212,667,668],{},"The /legacy/ module uses sync patterns. Do not convert to async.\n",[212,670,672],{"class":214,"line":671},42,[212,673,231],{"emptyLinePlaceholder":230},[212,675,677],{"class":214,"line":676},43,[212,678,679],{},"## Git\n",[212,681,683],{"class":214,"line":682},44,[212,684,231],{"emptyLinePlaceholder":230},[212,686,688],{"class":214,"line":687},45,[212,689,690],{},"Squash merge only.\n",[212,692,694],{"class":214,"line":693},46,[212,695,696],{},"Conventional commits: feat:, fix:, chore:, docs:.\n",[212,698,700],{"class":214,"line":699},47,[212,701,702],{},"Branch format: type/short-description (e.g., feat/user-auth).\n",[14,704,705],{},"That's under 50 lines. It communicates everything an agent needs to work correctly in this codebase. Every line is specific. Every section has a reason to exist.",[14,707,708],{},"Gartner projects 40% of enterprise applications will embed AI agents by end of 2026. The teams that invest 30 minutes in a good AGENTS.md today will save hours of agent correction every week. The teams that skip it will keep wondering why their agent \"doesn't follow instructions.\"",[14,710,711],{},[67,712],{"alt":713,"src":714},"Treat it like a bonsai, not an encyclopedia: start small, prune stale guidance, and add a section only when an agent consistently gets something wrong. A shorter, accurate file beats a comprehensive, generic one.","/img/blog/agents-md-best-practices-bonsai.jpg",[14,716,717],{},"The best AGENTS.md file is the one your team actually maintains. Not the one that comprehensively documents every edge case and is outdated by next sprint.",[14,719,720],{},"Start with 30 lines. Add a section when an agent consistently gets something wrong. Remove a section when the convention changes. Treat it like code, not documentation.",[14,722,723],{},"And if your agents are business agents, not coding agents, this entire configuration belongs in a visual builder, not a markdown file. Give BetterClaw a look. Free plan with 1 agent and every feature. $19/month per agent for Pro. Agent configuration through the UI. No files to forget to update.",[39,725,727],{"id":726},"frequently-asked-questions","Frequently Asked Questions",[171,729,731],{"id":730},"what-is-an-agentsmd-file","What is an AGENTS.md file?",[14,733,734],{},"AGENTS.md is a Markdown file placed at the root of a code repository that provides AI coding agents with project-specific instructions: build commands, code style conventions, architecture constraints, testing procedures, and boundaries. It's read natively by 30+ AI tools including Claude Code, GitHub Copilot, Cursor, OpenAI Codex, and Gemini CLI. Over 60,000 repositories have adopted it, and it's now stewarded by the Agentic AI Foundation under the Linux Foundation.",[171,736,738],{"id":737},"how-does-agentsmd-compare-to-claudemd-and-cursorrules","How does AGENTS.md compare to CLAUDE.md and .cursorrules?",[14,740,741,742,744],{},"AGENTS.md has the broadest compatibility (30+ tools read it). CLAUDE.md adds Claude-specific features like ",[50,743,365],{}," for modular instruction files. .cursorrules adds Cursor-specific features like glob-based auto-attach rules. 90% of the content is identical across all three. Start with AGENTS.md as your universal source of truth, and add tool-specific files only for features AGENTS.md can't express. A converter tool (rule-porter) can translate between formats.",[171,746,748],{"id":747},"how-long-should-an-agentsmd-file-be","How long should an AGENTS.md file be?",[14,750,751],{},"Under 150 lines. Research across 2,500+ repositories found that files beyond 150 lines deliver diminishing returns and can increase inference costs 20-23% without improving agent performance. The technical maximum is 32 KiB. Shorter, accurate files consistently outperform comprehensive, generic ones. Start with 30-50 lines covering stack, build commands, code style (with examples), and boundaries. Add sections only when an agent consistently makes a specific mistake.",[171,753,755],{"id":754},"should-i-use-an-llm-to-generate-my-agentsmd","Should I use an LLM to generate my AGENTS.md?",[14,757,758],{},"No. Research found that LLM-generated AGENTS.md files reduce task success in 5 out of 8 tested settings and add 2.45-3.92 extra steps per task. LLMs generate generic, verbose instructions that state things agents already know (\"write clean code,\" \"follow best practices\"). Write your AGENTS.md by hand with only the conventions specific to your project that an agent wouldn't know without being told. One real code snippet is worth more than three paragraphs of description.",[171,760,762],{"id":761},"can-agentsmd-principles-apply-to-business-agents-not-just-coding-agents","Can AGENTS.md principles apply to business agents, not just coding agents?",[14,764,765],{},"Yes. The core principle (give the agent structured, specific context about identity, capabilities, constraints, and boundaries) applies to any AI agent. For business agents, the equivalent sections are identity, capabilities, constraints, trust levels, escalation rules, and output format. The difference is that business agent configuration belongs in a visual builder (like BetterClaw) rather than a markdown file, since business agents don't operate from a code repository.",[767,768,769],"style",{},"html .default .shiki span {color: var(--shiki-default);background: var(--shiki-default-bg);font-style: var(--shiki-default-font-style);font-weight: var(--shiki-default-font-weight);text-decoration: var(--shiki-default-text-decoration);}html .shiki span {color: var(--shiki-default);background: var(--shiki-default-bg);font-style: var(--shiki-default-font-style);font-weight: var(--shiki-default-font-weight);text-decoration: var(--shiki-default-text-decoration);}",{"title":200,"searchDepth":221,"depth":221,"links":771},[772,773,778,779,780,781],{"id":41,"depth":221,"text":42},{"id":162,"depth":221,"text":163,"children":774},[775,776,777],{"id":173,"depth":227,"text":174},{"id":186,"depth":227,"text":187},{"id":330,"depth":227,"text":331},{"id":340,"depth":221,"text":341},{"id":381,"depth":221,"text":382},{"id":443,"depth":221,"text":444},{"id":726,"depth":221,"text":727,"children":782},[783,784,785,786,787],{"id":730,"depth":227,"text":731},{"id":737,"depth":227,"text":738},{"id":747,"depth":227,"text":748},{"id":754,"depth":227,"text":755},{"id":761,"depth":227,"text":762},"Best Practices","2026-06-11","60K+ repos use AGENTS.md. Most do it wrong. Under 150 lines, hand-written, with code examples. Template and annotated best practices.",false,"/img/blog/agents-md-best-practices.jpg",null,{},"/blog/agents-md-best-practices","10 min read",{"title":5,"description":790},"AGENTS.md Best Practices: Template and Guide (2026)","blog/agents-md-best-practices",[801,802,803,804,805,806],"agents.md best practices","agents md file","agents md template","agent configuration file","agents md format","agents md guide","l0tsD3Xto_iUXYLnstT4PJhf79saRo2pNBztXXua51w",[809,1151,1658],{"id":810,"title":811,"author":812,"body":813,"category":788,"date":789,"description":1136,"extension":452,"featured":791,"image":1137,"imageHeight":793,"imageWidth":793,"meta":1138,"navigation":230,"path":1139,"readingTime":796,"seo":1140,"seoTitle":1141,"stem":1142,"tags":1143,"updatedDate":789,"__hash__":1150},"blog/blog/ai-agent-human-approval-guardrails.md","AI Agent Guardrails: How to Add Human Approval Without Killing Speed",{"name":7,"role":8,"avatar":9},{"type":11,"value":814,"toc":1116},[815,818,821,824,827,830,833,836,840,846,852,855,861,867,871,874,877,880,883,887,890,893,896,900,903,909,915,921,927,930,934,937,940,943,947,950,953,959,965,971,974,978,981,987,993,999,1005,1008,1011,1015,1018,1024,1030,1036,1042,1045,1048,1052,1058,1061,1064,1067,1070,1073,1076,1079,1081,1085,1088,1092,1095,1099,1102,1106,1109,1113],[14,816,817],{},"Fully autonomous agents are fast and terrifying. Fully supervised agents are safe and useless. Here's the architecture that gives you both.",[14,819,820],{},"Our support agent sent a refund email to a customer. Without asking anyone. $487. Gone.",[14,822,823],{},"The agent was correct. The customer qualified for the refund. The policy was clear. The response was well-written.",[14,825,826],{},"But nobody on the team knew it happened until the customer replied with a thank-you email. Our finance lead walked into Monday morning with a transaction she didn't authorize, processed by a system she didn't know could authorize transactions.",[14,828,829],{},"That's the moment we added human approval gates.",[14,831,832],{},"Not because the agent was wrong. Because the right action, taken without oversight, eroded trust more than the wrong action would have. AI agent human approval isn't about catching mistakes. It's about maintaining the trust that keeps your team willing to let agents do more over time.",[14,834,835],{},"Here's how to add approval gates without turning your agent into a chatbot that asks permission to breathe.",[39,837,839],{"id":838},"the-speed-vs-safety-tradeoff-and-why-both-extremes-fail","The speed vs safety tradeoff (and why both extremes fail)",[14,841,842,845],{},[74,843,844],{},"Fully autonomous:"," The agent acts on everything without asking. Fast. Efficient. Also terrifying. One hallucinated tool call, one misinterpreted instruction, one edge case nobody anticipated, and the agent takes an action you can't undo. Meta's Summer Yue watched her agent mass-delete emails while ignoring stop commands. That's the fully autonomous failure mode.",[14,847,848,851],{},[74,849,850],{},"Fully supervised:"," The agent drafts everything and waits for human approval before every action. Safe. Predictable. Also useless. If a human has to review and approve every single agent action, you haven't automated anything. You've added a middleman between the human and the task.",[14,853,854],{},"The answer is neither extreme. It's tiered autonomy: the agent acts autonomously on routine, low-risk tasks and pauses for human approval on high-risk actions. The boundary between \"auto-approve\" and \"wait for human\" is the entire engineering problem.",[856,857,858],"blockquote",{},[14,859,860],{},"The goal isn't \"should this agent need approval?\" The goal is \"which specific actions need approval, and which can the agent handle alone?\" Draw the line at the action level, not the agent level.",[14,862,863],{},[67,864],{"alt":865,"src":866},"The three-tier approval highway: Tier 1 auto-approves routine read-only actions, Tier 2 queues medium-risk actions for human review, and Tier 3 blocks irreversible high-impact actions until a human approves.","/img/blog/ai-agent-human-approval-guardrails-three-tiers.jpg",[39,868,870],{"id":869},"tier-1-auto-approve-the-80-that-should-just-happen","Tier 1: Auto-approve (the 80% that should just happen)",[14,872,873],{},"Most of what your agent does is read-only or low-impact. Reading emails. Classifying tickets. Looking up customer data. Drafting responses. Summarizing documents. Querying databases.",[14,875,876],{},"These actions should execute immediately with zero approval gate. If you make your agent ask permission to read an email, you've killed the value proposition.",[14,878,879],{},"The rule: if the action is reversible, read-only, or internal-only, auto-approve it. Nobody needs to approve a CRM lookup. Nobody needs to approve a draft that hasn't been sent yet.",[14,881,882],{},"Most teams start by requiring approval on everything and gradually moving actions to auto-approve as trust builds. This is backwards. Start by auto-approving everything that's obviously safe and add approval gates only to the specific actions that need them.",[39,884,886],{"id":885},"tier-2-queue-for-review-the-15-that-needs-a-quick-check","Tier 2: Queue for review (the 15% that needs a quick check)",[14,888,889],{},"These are actions with moderate impact that the agent handles correctly 95% of the time but where the 5% failure case matters.",[14,891,892],{},"Sending an email to a customer. Updating a CRM record. Applying a discount code. Posting to a public channel. Scheduling a meeting on someone's calendar.",[14,894,895],{},"The agent drafts the action and queues it for human review. The human gets a notification (Slack, email, or dashboard), reviews the draft, and approves or rejects. If approved, the agent executes. If rejected, the agent logs the feedback.",[171,897,899],{"id":898},"designing-the-approval-channel","Designing the approval channel",[14,901,902],{},"Where the approval notification goes matters more than most people think.",[14,904,905],{},[67,906],{"alt":907,"src":908},"Three ways to deliver approval requests: Slack/Teams notifications for real-time teams (2-15 min), email digests for batch review (30-60 min), and a dashboard queue for high-volume teams (1-4 hours).","/img/blog/ai-agent-human-approval-guardrails-approval-channels.jpg",[14,910,911,914],{},[74,912,913],{},"Slack/Teams notification:"," Best for real-time teams. The agent posts a message with the proposed action, a summary of why, and approve/reject buttons. Approval latency: 2-15 minutes during business hours.",[14,916,917,920],{},[74,918,919],{},"Email digest:"," Best for batch review. The agent collects queued actions and sends a summary every 30 minutes or every hour. The reviewer approves or rejects in bulk. Approval latency: 30-60 minutes.",[14,922,923,926],{},[74,924,925],{},"Dashboard queue:"," Best for high-volume teams. A dedicated approval dashboard shows all pending actions, sorted by priority. Reviewers work through the queue during scheduled review windows. Approval latency: 1-4 hours.",[14,928,929],{},"Pick based on your team's workflow. A 5-person startup that lives in Slack should use Slack notifications. A 50-person operations team should use a dashboard queue.",[39,931,933],{"id":932},"tier-3-block-until-approved-the-5-you-cant-undo","Tier 3: Block until approved (the 5% you can't undo)",[14,935,936],{},"These are actions with irreversible or high-financial-impact consequences. Processing refunds. Deleting records. Sending external communications with legal implications. Making API calls that trigger financial transactions.",[14,938,939],{},"For Tier 3, the agent stops completely and waits for explicit human approval. No timeout. No auto-approve after 30 minutes. The action does not happen until a human says yes.",[14,941,942],{},"This sounds restrictive. In practice, Tier 3 actions are rare. If your agent handles 200 tasks per day and 5% are Tier 3, that's 10 approval requests. Your team reviews 10 items per day. That's manageable. The other 190 tasks happen instantly.",[171,944,946],{"id":945},"the-timeout-problem","The timeout problem",[14,948,949],{},"Tier 2 needs a timeout. What happens if the reviewer is in a meeting, on vacation, or simply doesn't see the notification?",[14,951,952],{},"Three options:",[14,954,955,958],{},[74,956,957],{},"Auto-approve after timeout."," The action executes if nobody rejects it within 30 minutes. This is the fastest but riskiest. Only appropriate for Tier 2 actions where the failure case is mild.",[14,960,961,964],{},[74,962,963],{},"Escalate after timeout."," If the primary reviewer doesn't respond within 30 minutes, the approval request escalates to the next person in the chain (team lead, then manager). This is the most common pattern for business-critical agents.",[14,966,967,970],{},[74,968,969],{},"Auto-reject after timeout."," The action is cancelled if nobody approves it. The customer or process waits. This is the safest but creates the most friction. Use for actions where doing nothing is better than doing the wrong thing.",[14,972,973],{},"On most teams, escalation is the right default. Auto-approve timeouts sound efficient until the one time it auto-approves a $5,000 refund because the reviewer was at lunch.",[39,975,977],{"id":976},"the-escalation-chain-who-approves-what-and-when","The escalation chain (who approves what, and when)",[14,979,980],{},"A well-designed escalation chain has three levels:",[14,982,983],{},[67,984],{"alt":985,"src":986},"The approval escalation plan — know this before 2 AM: Level 1 is the agent's assigned reviewer, Level 2 is the team lead after a timeout, and Level 3 is a catch-all admin or on-call rotation.","/img/blog/ai-agent-human-approval-guardrails-escalation.jpg",[14,988,989,992],{},[74,990,991],{},"Level 1: The agent's assigned reviewer."," This is the person or team responsible for the agent's domain. A support agent's reviewer is the support lead. A sales agent's reviewer is the sales manager.",[14,994,995,998],{},[74,996,997],{},"Level 2: The team lead or department head."," If Level 1 doesn't respond within the timeout window (15-30 minutes), the request escalates.",[14,1000,1001,1004],{},[74,1002,1003],{},"Level 3: A catch-all admin or on-call."," If Level 2 doesn't respond, the request goes to someone who is always monitoring. For critical systems, this should be a PagerDuty-style rotation.",[14,1006,1007],{},"Map each agent to an escalation chain when you deploy it. Don't assume \"someone will see it.\" Assumption is the enemy of reliable agent operations.",[14,1009,1010],{},"This is exactly the kind of guardrail architecture that should be built into your platform, not bolted on after the fact. On BetterClaw, trust levels (Intern, Specialist, Lead) implement tiered autonomy natively. An \"Intern\" agent drafts but never executes. A \"Specialist\" executes within defined boundaries. A \"Lead\" acts autonomously with the one-click kill switch as the emergency brake. 200+ verified skills with built-in action approval. Free plan with every feature. $19/month per agent on Pro. BYOK with zero markup.",[39,1012,1014],{"id":1013},"the-metrics-that-tell-you-if-your-guardrails-are-working","The metrics that tell you if your guardrails are working",[14,1016,1017],{},"Deploy approval gates and then measure whether they're calibrated correctly. Three metrics matter:",[14,1019,1020],{},[67,1021],{"alt":1022,"src":1023},"The guardrail calibration dashboard: three dials to watch — approval latency (target under 30 min), override rate (move under 2% to auto-approve, fix over 20%), and false-positive rate (tighten if over 15% are rubber-stamped).","/img/blog/ai-agent-human-approval-guardrails-metrics.jpg",[14,1025,1026,1029],{},[74,1027,1028],{},"Approval latency."," How long does it take from the agent queuing an action to a human approving it? If average latency exceeds 30 minutes for Tier 2 actions, your agent is spending more time waiting than working. Either the notification channel is wrong, the escalation chain is too slow, or you have too many actions in Tier 2 that should be in Tier 1.",[14,1031,1032,1035],{},[74,1033,1034],{},"Override rate."," What percentage of queued actions does the human reject or modify? If the override rate is below 2%, the actions in that tier are likely safe to auto-approve. Move them to Tier 1. If the override rate is above 20%, the agent is making too many errors and needs prompt or model improvements before being trusted with those actions.",[14,1037,1038,1041],{},[74,1039,1040],{},"False-positive rate."," How often does the guardrail fire on an action that didn't need review? If 90% of your Tier 2 approvals are rubber-stamped without changes, you've created busywork. Tighten the criteria for what triggers review.",[14,1043,1044],{},"The ideal approval system has high Tier 1 volume (the agent works fast on routine tasks), low Tier 2 volume (only genuinely ambiguous actions need review), and near-zero Tier 3 volume (high-risk actions are rare by design).",[14,1046,1047],{},"Gartner projects 40% of enterprise applications will embed AI agents by end of 2026. The organizations that get adoption right will be the ones whose teams trust their agents. And trust is built by demonstrating that the agent asks before acting on the things that matter.",[39,1049,1051],{"id":1050},"the-kill-switch-for-when-everything-else-fails","The kill switch (for when everything else fails)",[14,1053,1054],{},[67,1055],{"alt":1056,"src":1057},"The kill switch: build it first, hope to never use it. One button immediately halts all agent activity — no pending actions execute, no queued tasks complete, everything stops.","/img/blog/ai-agent-human-approval-guardrails-kill-switch.jpg",[14,1059,1060],{},"Every agent needs an emergency stop. Not a \"politely wind down\" mechanism. A button that immediately halts all agent activity. No pending actions execute. No queued tasks complete. Everything stops.",[14,1062,1063],{},"This isn't for routine operations. It's for the scenario where the agent is doing something unexpected at scale and you need it to stop now. The Meta email deletion incident is the canonical example. The agent was acting autonomously, the user couldn't stop it, and the damage compounded with every second.",[14,1065,1066],{},"Build the kill switch before you deploy the agent. Not after. Test it regularly. Know where it is at 2 AM when your phone buzzes.",[14,1068,1069],{},"The hardest part of AI agent guardrails isn't the technology. It's the organizational discipline. Defining which actions are Tier 1, 2, and 3. Assigning reviewers. Setting timeouts. Reviewing the metrics. Updating the tiers as the agent proves itself.",[14,1071,1072],{},"The teams that do this well gradually move actions from Tier 2 to Tier 1 as the agent demonstrates reliability. The boundary between \"needs approval\" and \"auto-approve\" shifts over time. That's the whole point. You're not building a permanent approval wall. You're building a trust gradient that widens as the agent earns it.",[14,1074,1075],{},"Start with more oversight than you think you need. Remove it as the data tells you to. Never the other way around.",[14,1077,1078],{},"Give BetterClaw a look if you want trust levels built in from day one. Intern, Specialist, Lead. Action approval on every skill. One-click kill switch. Free plan with 1 agent and every feature. $19/month per agent for Pro. We handle the guardrail infrastructure. You decide where the lines go.",[39,1080,727],{"id":726},[171,1082,1084],{"id":1083},"what-are-ai-agent-guardrails","What are AI agent guardrails?",[14,1086,1087],{},"AI agent guardrails are safety controls that define what an autonomous agent can do without supervision and what requires human approval before execution. The most effective approach is tiered autonomy: routine, low-risk actions (reading data, drafting responses) are auto-approved, medium-risk actions (sending emails, updating records) are queued for human review, and high-risk actions (processing payments, deleting data) are blocked until a human explicitly approves.",[171,1089,1091],{"id":1090},"how-does-human-approval-compare-to-fully-autonomous-agents","How does human approval compare to fully autonomous agents?",[14,1093,1094],{},"Fully autonomous agents are faster but riskier. A single hallucinated action can cause real damage (unauthorized refunds, deleted data, incorrect external communications). Fully supervised agents are safe but negate the value of automation. Tiered autonomy gives you 80-90% of the speed (routine actions execute instantly) with 95%+ of the safety (risky actions wait for human review). The overhead is typically 10-15 approval reviews per day for a well-calibrated system.",[171,1096,1098],{"id":1097},"how-do-i-set-up-approval-workflows-for-my-ai-agent","How do I set up approval workflows for my AI agent?",[14,1100,1101],{},"Define three tiers: Tier 1 (auto-approve: read-only, reversible, internal actions), Tier 2 (queue for review: customer-facing, data-modifying actions), Tier 3 (block until approved: financial, irreversible, legal-impact actions). Choose an approval channel (Slack for real-time, email for batch, dashboard for high-volume). Set timeout rules for Tier 2 (escalate after 30 minutes). Assign an escalation chain per agent. On BetterClaw, this maps directly to trust levels: Intern (Tier 2/3), Specialist (Tier 1/2), Lead (mostly Tier 1).",[171,1103,1105],{"id":1104},"does-adding-approval-steps-slow-down-my-ai-agent","Does adding approval steps slow down my AI agent?",[14,1107,1108],{},"Only for the 15-20% of actions that genuinely need review. Tier 1 actions (80%+ of typical workloads) execute with zero delay. Tier 2 actions add 2-60 minutes depending on your approval channel and reviewer availability. Tier 3 actions wait for explicit approval with no timeout. If your approval latency exceeds 30 minutes on average, either your channel is wrong, your escalation chain is too slow, or too many actions are in Tier 2 that should be in Tier 1.",[171,1110,1112],{"id":1111},"how-do-i-know-if-my-agent-guardrails-are-calibrated-correctly","How do I know if my agent guardrails are calibrated correctly?",[14,1114,1115],{},"Track three metrics: approval latency (target: under 30 minutes average for Tier 2), override rate (if under 2%, move those actions to auto-approve; if over 20%, improve the agent before trusting it), and false-positive rate (if over 15% of approvals are rubber-stamped, tighten the review criteria). Review these metrics monthly and adjust tier assignments. The goal is to gradually move actions from Tier 2 to Tier 1 as the agent demonstrates reliability over time.",{"title":200,"searchDepth":221,"depth":221,"links":1117},[1118,1119,1120,1123,1126,1127,1128,1129],{"id":838,"depth":221,"text":839},{"id":869,"depth":221,"text":870},{"id":885,"depth":221,"text":886,"children":1121},[1122],{"id":898,"depth":227,"text":899},{"id":932,"depth":221,"text":933,"children":1124},[1125],{"id":945,"depth":227,"text":946},{"id":976,"depth":221,"text":977},{"id":1013,"depth":221,"text":1014},{"id":1050,"depth":221,"text":1051},{"id":726,"depth":221,"text":727,"children":1130},[1131,1132,1133,1134,1135],{"id":1083,"depth":227,"text":1084},{"id":1090,"depth":227,"text":1091},{"id":1097,"depth":227,"text":1098},{"id":1104,"depth":227,"text":1105},{"id":1111,"depth":227,"text":1112},"Your agent shouldn't approve refunds alone. But it shouldn't ask permission to read email. Three-tier approval architecture inside.","/img/blog/ai-agent-human-approval-guardrails.jpg",{},"/blog/ai-agent-human-approval-guardrails",{"title":811,"description":1136},"AI Agent Guardrails: Human Approval Without the Lag","blog/ai-agent-human-approval-guardrails",[1144,1145,1146,1147,1148,1149],"ai agent human approval","agent guardrails","human in the loop agent","agent approval workflow","ai agent safety controls","agent trust levels","ybSc6xSKgKh1Sh3OfgUCMc-BLEBeAjeo30CrjQyfHX8",{"id":1152,"title":1153,"author":1154,"body":1155,"category":788,"date":1632,"description":1633,"extension":452,"featured":230,"image":1634,"imageHeight":793,"imageWidth":793,"meta":1635,"navigation":230,"path":1636,"readingTime":1637,"seo":1638,"seoTitle":1639,"stem":1640,"tags":1641,"updatedDate":1656,"__hash__":1657},"blog/blog/best-openclaw-skills.md","15+ Best OpenClaw ClawHub Skills (Tested & Security-Vetted, 2026)",{"name":7,"role":8,"avatar":9},{"type":11,"value":1156,"toc":1620},[1157,1162,1165,1168,1171,1177,1184,1187,1190,1193,1197,1200,1210,1216,1221,1224,1227,1233,1237,1240,1246,1252,1258,1264,1270,1276,1280,1283,1289,1295,1301,1307,1313,1318,1324,1328,1331,1337,1343,1349,1354,1360,1366,1370,1373,1379,1385,1391,1397,1401,1404,1410,1416,1422,1428,1433,1452,1456,1459,1465,1471,1476,1496,1499,1505,1509,1515,1518,1521,1529,1536,1540,1543,1546,1549,1552,1555,1557,1562,1573,1578,1581,1586,1601,1606,1609,1614,1617],[14,1158,1159],{},[74,1160,1161],{},"With 5,700+ skills on ClawHub, most people install the wrong ones first. Here are the ones that actually matter, organized by what you're trying to get done. Last verified and updated: March 2026.",[14,1163,1164],{},"The first skill I ever installed on OpenClaw nearly leaked my Google credentials.",[14,1166,1167],{},"It had good documentation. Decent stars on ClawHub. The description sounded exactly like what I needed. But buried in the install flow was a dependency pull from an unverified mirror. Nothing flagged it. No warning. I only caught it because I read the source code before running it.",[14,1169,1170],{},"Most people don't do that.",[14,1172,1173,1174],{},"And here's the uncomfortable truth about ClawHub in March 2026: there are over 5,700 community-built skills on the registry. Security researchers have flagged at least 341 malicious ones. Semgrep's analysis estimates the registry is roughly 10% compromised. That's not a typo. ",[74,1175,1176],{},"One in ten skills on the most popular AI agent marketplace might be trying to steal your data.",[14,1178,1179,1180],{},"So when you search \"best OpenClaw skills,\" what you're really asking is: ",[1181,1182,1183],"em",{},"which ones can I actually trust, and which ones will make my agent genuinely useful?",[14,1185,1186],{},"That's what this guide is for.",[14,1188,1189],{},"We've spent weeks testing, vetting, and running OpenClaw skills across real workflows. Not just poking at them in a sandbox for five minutes. Actually running them in production agent deployments. What follows is our curated, opinionated list organized by what you're actually trying to accomplish.",[14,1191,1192],{},"But first, a quick refresher on something most guides get wrong.",[39,1194,1196],{"id":1195},"skills-vs-tools-the-distinction-that-saves-you-from-yourself","Skills vs. Tools: The Distinction That Saves You From Yourself",[14,1198,1199],{},"Before you install anything, understand this:",[14,1201,1202,1205,1206,1209],{},[74,1203,1204],{},"Tools are the muscles."," They determine what your agent can do. Read files. Execute commands. Browse the web. These are controlled by the ",[50,1207,1208],{},"tools.allow"," configuration.",[14,1211,1212,1215],{},[74,1213,1214],{},"Skills are the playbook."," They teach your agent how to combine tools for specific tasks. The github skill teaches your agent how to manage repos. The obsidian skill teaches it how to organize notes. But without the right tools enabled, skills are just instructions with no hands.",[14,1217,1218],{},[74,1219,1220],{},"Key takeaway: Installing a skill does NOT automatically give your agent new permissions. You still control what tools are enabled. This is your primary safety lever. Use it.",[14,1222,1223],{},"Three conditions must be met for any skill to actually work: the tool must be allowed in config, the required software must be installed on your machine (or in the sandbox), and the skill must be loaded in your workspace. Miss any one of these, and nothing happens.",[14,1225,1226],{},"Now, let's get into the picks.",[14,1228,1229],{},[67,1230],{"alt":1231,"src":1232},"OpenClaw skills vs tools diagram showing the distinction between tool permissions and skill playbooks","/img/blog/openclaw-skills-vs-tools.jpg",[39,1234,1236],{"id":1235},"the-productivity-stack-your-agents-daily-operating-system","The Productivity Stack: Your Agent's Daily Operating System",[14,1238,1239],{},"These are the skills that turn OpenClaw from \"interesting experiment\" into \"I can't work without this.\"",[14,1241,1242],{},[67,1243],{"alt":1244,"src":1245},"Productivity skills stack overview showing Google Workspace, Notion, Meeting Prep, and Task Prioritizer integrations","/img/blog/openclaw-productivity-stack.jpg",[14,1247,1248,1251],{},[74,1249,1250],{},"Google Workspace (gog)"," This is the foundational productivity skill and probably the first one you should install. It gives your agent access to Gmail, Google Calendar, Google Docs, and Sheets. The real power shows up when you combine it with the heartbeat scheduler. Set your agent to check your calendar every morning and send you a briefing via WhatsApp before you've had coffee.",[14,1253,1254,1257],{},[1181,1255,1256],{},"Security note:"," This skill gets deep access to your Google account. Scope it carefully. Give read access to your calendar but limit write access to specific documents. Never give blanket Drive access.",[14,1259,1260,1263],{},[74,1261,1262],{},"Notion Integration"," If your team runs on Notion (and in 2026, who doesn't?), this skill lets your agent create pages, update databases, query project boards, and manage documentation. The sweet spot is pairing it with meeting notes. Your agent joins a call summary, extracts action items, and drops them into your Notion project board. Automatically.",[14,1265,1266,1269],{},[74,1267,1268],{},"Meeting Prep Agent"," This one changed my workflow more than any other. Before every meeting, it gathers relevant context: calendar details, past notes, related documents, email threads. It assembles a briefing you can skim in 90 seconds. No more scrambling to remember what you discussed last week.",[14,1271,1272,1275],{},[74,1273,1274],{},"Task Prioritizer"," Uses AI to rank your to-do list based on deadlines, dependencies, and context from your other skills. It's not magic, but it's surprisingly good at surfacing the thing you should be doing right now instead of the thing that feels urgent.",[39,1277,1279],{"id":1278},"the-developer-stack-skills-that-actually-ship-code","The Developer Stack: Skills That Actually Ship Code",[14,1281,1282],{},"If you're a developer, these are the skills that earn their keep.",[14,1284,1285],{},[67,1286],{"alt":1287,"src":1288},"Developer skills stack showing GitHub, Cursor CLI, Docker, Vercel, and Sentry integrations for coding workflows","/img/blog/openclaw-developer-stack.jpg",[14,1290,1291,1294],{},[74,1292,1293],{},"GitHub Integration"," Non-negotiable if you write code. Manage issues, pull requests, repos, and webhooks directly through your agent. The real unlock: set up a webhook listener so your agent gets notified on new PRs and can summarize changes before you review them. Pair it with the heartbeat to get a daily digest of repo activity.",[14,1296,1297,1300],{},[74,1298,1299],{},"Cursor CLI Agent"," This skill bridges your OpenClaw agent to the Cursor AI coding assistant. If you're already using Cursor for development, this lets you trigger code generation, refactoring, and analysis tasks from any chat channel. Text your agent from Telegram, and it kicks off a Cursor session in the background. Updated for 2026 features with tmux automation support.",[14,1302,1303,1306],{},[74,1304,1305],{},"Docker Manager"," For DevOps workflows, this skill lets your agent manage Docker containers, images, and compose stacks. Start, stop, inspect, and clean up containers through chat. Particularly useful if you're managing multiple environments and don't want to SSH into a server every time something needs a restart.",[14,1308,1309,1312],{},[74,1310,1311],{},"Vercel Deployment"," If you deploy to Vercel, this skill turns deployments into conversational commands. Manage environment variables, configure domains, trigger releases. You go from \"I deploy when I decide to\" to \"the system deploys when conditions are met.\"",[14,1314,1315,1317],{},[1181,1316,1256],{}," This gives your agent production deployment rights. Start in a staging environment. Always.",[14,1319,1320,1323],{},[74,1321,1322],{},"Sentry CLI"," Connects your agent to Sentry for error monitoring. Get notified about new errors through your messaging channels, query error details, and even trigger resolutions. When combined with the GitHub skill, your agent can spot an error, find the relevant PR, and create an issue with full context.",[39,1325,1327],{"id":1326},"the-automation-stack-making-your-agent-proactive","The Automation Stack: Making Your Agent Proactive",[14,1329,1330],{},"These skills move your agent from reactive (\"do this when I ask\") to proactive (\"do this because you noticed something\").",[14,1332,1333],{},[67,1334],{"alt":1335,"src":1336},"Automation skills stack showing Cron Job Manager, Web Browser, Tavily Search, and n8n workflow integrations","/img/blog/openclaw-automation-stack.jpg",[14,1338,1339,1342],{},[74,1340,1341],{},"Cron Job Manager"," Create scheduled tasks using natural language. \"Remind me every Monday at 9 AM to review the sprint board.\" \"Check Hacker News every morning and send me the top 5 AI stories.\" The cron system is one of OpenClaw's most powerful features, and this skill makes it accessible without touching terminal syntax.",[14,1344,1345,1348],{},[74,1346,1347],{},"Web Browser Automation"," A Rust-based headless browser skill that lets your agent navigate pages, click elements, fill forms, and capture screenshots. This is the backbone of any monitoring or scraping workflow. Want your agent to check competitor pricing every day? This is how.",[14,1350,1351,1353],{},[1181,1352,1256],{}," Browser automation skills can visit any URL your agent encounters. This is a significant prompt injection surface. Sandbox this aggressively.",[14,1355,1356,1359],{},[74,1357,1358],{},"Tavily Search"," AI-optimized web search that's far more useful than having your agent use a basic search tool. Tavily returns structured, AI-ready results with summaries. Perfect for research tasks, competitive analysis, and keeping your agent informed about topics that matter to you.",[14,1361,1362,1365],{},[74,1363,1364],{},"n8n Workflow Manager"," If you're running n8n for workflow automation, this skill connects your OpenClaw agent to your n8n instance. Activate workflows, check execution status, trigger manual runs. It turns your agent into a control panel for your entire automation stack.",[39,1367,1369],{"id":1368},"the-smart-home-and-personal-stack","The Smart Home and Personal Stack",[14,1371,1372],{},"These are the skills that make OpenClaw feel less like a dev tool and more like an actual assistant.",[14,1374,1375],{},[67,1376],{"alt":1377,"src":1378},"Smart home and personal skills showing Home Assistant, Sonos, and Weather integrations for everyday use","/img/blog/openclaw-smarthome-stack.jpg",[14,1380,1381,1384],{},[74,1382,1383],{},"Home Assistant Integration"," Control lights, locks, thermostats, and other smart devices through your chat channels. The home automation community has embraced OpenClaw hard, and this skill is one of the most polished in the entire ecosystem. Text your agent to turn off the lights from bed. Or set up a heartbeat that adjusts your thermostat based on your calendar (leaving for work? Lower the heat).",[14,1386,1387,1390],{},[74,1388,1389],{},"Sonos Control"," Manage your Sonos speakers through your agent. Play, pause, adjust volume, switch rooms. It's simple, but it's also the kind of thing that makes you realize you're living in the future when you text \"play lo-fi in the office\" from the other room.",[14,1392,1393,1396],{},[74,1394,1395],{},"Weather + Solar"," Real-time weather data and solar weather monitoring. Useful on its own, but powerful when combined with heartbeats. \"If it's going to rain tomorrow, remind me tonight to bring an umbrella.\" Small quality-of-life automation that adds up.",[39,1398,1400],{"id":1399},"the-skills-you-should-not-install-yet","The Skills You Should NOT Install (Yet)",[14,1402,1403],{},"Here's where we get opinionated.",[14,1405,1406],{},[67,1407],{"alt":1408,"src":1409},"Warning signs for unsafe OpenClaw skills showing red flags to watch for on ClawHub","/img/blog/openclaw-skills-to-avoid.jpg",[14,1411,1412,1415],{},[74,1413,1414],{},"Avoid skills from unverified authors with fewer than 100 installs."," The ClawHub registry's vetting process is still immature. Three independent reports can auto-hide a skill, but the removal process is slow. Stick to skills published in the official github.com/openclaw/skills repository or from authors you can verify.",[14,1417,1418,1421],{},[74,1419,1420],{},"Be cautious with \"self-improving\" or \"auto-evolution\" skills."," Several highly-starred skills claim to make your agent \"continuously enhance its own capabilities.\" That sounds exciting. It's also exactly the kind of recursive, autonomous behavior that's hardest to audit and most likely to surprise you in production.",[14,1423,1424,1427],{},[74,1425,1426],{},"Skip any skill that asks for broader permissions than its stated purpose."," If a calendar skill wants terminal access, that's a red flag. If a weather skill wants to read your files, walk away. Apply the principle of least privilege to every skill you install.",[14,1429,1430],{},[74,1431,1432],{},"Our rule of thumb: if you can't read and understand a skill's SKILL.md and source code in under five minutes, it's either too complex for its stated purpose or doing more than it claims.",[14,1434,1435,1436,1441,1442,1446,1447,1451],{},"For a full breakdown of every documented security incident, see our ",[1437,1438,1440],"a",{"href":1439},"/blog/openclaw-security-risks","OpenClaw security risks guide",". If you're running skills on ",[1437,1443,1445],{"href":1444},"/pricing","BetterClaw's managed OpenClaw platform",", this risk is significantly lower. Every agent runs in a Docker-sandboxed environment with AES-256 encrypted credentials, workspace scoping, and ",[1437,1448,1450],{"href":1449},"/#features","real-time health monitoring that auto-pauses on anomalies",". You still choose your skills, but the blast radius of a bad one is contained by default.",[39,1453,1455],{"id":1454},"how-to-install-openclaw-skills-the-right-way","How to Install OpenClaw Skills (The Right Way)",[14,1457,1458],{},"The process is simple. Doing it safely takes a few extra steps.",[14,1460,1461,1464],{},[74,1462,1463],{},"Step 1: Search before you install."," Use ClawHub's vector search to describe what you need in plain English. \"I need something that summarizes my emails every morning\" will return better results than keyword searching \"email summarizer.\"",[14,1466,1467,1470],{},[74,1468,1469],{},"Step 2: Vet before you trust."," Check the skill's install count, last update date, and author. Read the source code. Check the VirusTotal report on the skill's ClawHub page. If anything looks off, skip it.",[14,1472,1473],{},[74,1474,1475],{},"Step 3: Install with one command.",[192,1477,1481],{"className":1478,"code":1479,"language":1480,"meta":200,"style":200},"language-bash shiki shiki-themes github-light","clawhub install skill-name\n","bash",[50,1482,1483],{"__ignoreMap":200},[212,1484,1485,1489,1493],{"class":214,"line":215},[212,1486,1488],{"class":1487},"s7eDp","clawhub",[212,1490,1492],{"class":1491},"sYBdl"," install",[212,1494,1495],{"class":1491}," skill-name\n",[14,1497,1498],{},"The skill downloads, validates, and activates. Start a new OpenClaw session to pick it up.",[14,1500,1501,1504],{},[74,1502,1503],{},"Step 4: Scope your permissions."," After installing, review what tools the skill needs and only enable the minimum required. Don't give write access when read access will do. Don't enable exec when the skill only needs web access.",[39,1506,1508],{"id":1507},"the-easier-path-skills-on-betterclaw","The Easier Path: Skills on BetterClaw",[14,1510,1511],{},[67,1512],{"alt":1513,"src":1514},"BetterClaw managed platform showing secure skill deployment with sandboxed execution and encrypted credentials","/img/blog/betterclaw-skills-deployment.jpg",[14,1516,1517],{},"Everything we've covered in this article, the vetting, the permission scoping, the sandbox configuration, the tool management, is work you have to do yourself when self-hosting OpenClaw.",[14,1519,1520],{},"And it's worth doing if you want to learn the system deeply.",[14,1522,1523,1524,1528],{},"But if your goal is a production-ready OpenClaw agent with the best skills running securely across your team's chat channels, ",[1437,1525,1527],{"href":1526},"/","BetterClaw handles the infrastructure"," so you can focus on choosing the right skills for your workflow. One-click deploy. Sandboxed execution. Encrypted credentials. $19/month per agent, BYOK.",[14,1530,1531,1532],{},"You pick the skills. We make sure they run safely. Already on self-hosted OpenClaw? ",[1437,1533,1535],{"href":1534},"/migrate","Migrate to BetterClaw in under an hour →",[39,1537,1539],{"id":1538},"start-with-three-then-expand","Start With Three, Then Expand",[14,1541,1542],{},"The biggest mistake I see new OpenClaw users make is installing 20 skills on day one. Don't do that.",[14,1544,1545],{},"Start with three. Pick the ones that solve a problem you actually have today. The Google Workspace skill for calendar and email. The GitHub integration if you're a developer. The cron job manager to make your agent proactive.",[14,1547,1548],{},"Run those for a week. Watch how your agent uses them. Get comfortable with the permission model and the heartbeat system. Then expand from there.",[14,1550,1551],{},"The best OpenClaw skills aren't the ones with the most stars. They're the ones you use every day without thinking about them. The ones that quietly handle the work you used to do manually. The ones that make you forget your agent is software and start treating it like a teammate.",[14,1553,1554],{},"That's when things get interesting.",[39,1556,727],{"id":726},[14,1558,1559],{},[74,1560,1561],{},"What are OpenClaw skills and how do they work?",[14,1563,1564,1565,1568,1569,1572],{},"OpenClaw skills are modular text-based extensions (a ",[50,1566,1567],{},"SKILL.md"," file plus supporting files) that teach your AI agent how to perform specific tasks. They don't grant new permissions on their own. Skills work by combining the tools already enabled in your agent's configuration. You install them via the ClawHub registry using a single CLI command (",[50,1570,1571],{},"clawhub install skill-name","), and they activate on your next agent session.",[14,1574,1575],{},[74,1576,1577],{},"How do OpenClaw skills compare to ChatGPT plugins or Claude tools?",[14,1579,1580],{},"The key difference is that OpenClaw skills run locally on your machine and have access to your actual files, apps, and system. ChatGPT plugins and Claude's tools run server-side with limited, sandboxed capabilities. OpenClaw skills can chain together (GitHub webhook triggers a Docker build which triggers a Discord notification), while cloud-based plugins typically operate in isolation. The tradeoff is more power but more security responsibility.",[14,1582,1583],{},[74,1584,1585],{},"How do I install OpenClaw skills from ClawHub safely?",[14,1587,1588,1589,1592,1593,1595,1596,1600],{},"Search ClawHub using the vector search or CLI (",[50,1590,1591],{},"clawhub search \"what you need\"","), then vet the skill by checking its install count, author, last update, and VirusTotal scan. Install with ",[50,1594,1571],{},". After installation, scope permissions to the minimum required. For maximum safety, run new skills in a sandbox first. On managed platforms like ",[1437,1597,1599],{"href":1598},"/compare/openclaw","BetterClaw",", sandbox isolation is built in by default.",[14,1602,1603],{},[74,1604,1605],{},"Is it worth paying for managed OpenClaw skill deployment?",[14,1607,1608],{},"If you're running OpenClaw for personal experimentation, self-hosting is fine and free. If you're running it for a team or business, the time spent on security auditing, permission management, Docker configuration, and monitoring adds up fast. BetterClaw at $19/month per agent includes sandboxed execution, encrypted credentials, and auto-pause monitoring, which effectively replaces hours of weekly ops work.",[14,1610,1611],{},[74,1612,1613],{},"Are OpenClaw ClawHub skills secure enough for business use?",[14,1615,1616],{},"Not all of them. Security researchers have identified hundreds of malicious skills on ClawHub, and the vetting process is still maturing. For business use, stick to official bundled skills and well-known community skills with high install counts and recent updates. Always review source code, apply least-privilege permissions, and run skills in sandboxed environments. Managed platforms like BetterClaw add enterprise-grade security layers (AES-256 encryption, Docker isolation, workspace scoping) that significantly reduce risk.",[767,1618,1619],{},"html pre.shiki code .s7eDp, html code.shiki .s7eDp{--shiki-default:#6F42C1}html pre.shiki code .sYBdl, html code.shiki .sYBdl{--shiki-default:#032F62}html .default .shiki span {color: var(--shiki-default);background: var(--shiki-default-bg);font-style: var(--shiki-default-font-style);font-weight: var(--shiki-default-font-weight);text-decoration: var(--shiki-default-text-decoration);}html .shiki span {color: var(--shiki-default);background: var(--shiki-default-bg);font-style: var(--shiki-default-font-style);font-weight: var(--shiki-default-font-weight);text-decoration: var(--shiki-default-text-decoration);}",{"title":200,"searchDepth":221,"depth":221,"links":1621},[1622,1623,1624,1625,1626,1627,1628,1629,1630,1631],{"id":1195,"depth":221,"text":1196},{"id":1235,"depth":221,"text":1236},{"id":1278,"depth":221,"text":1279},{"id":1326,"depth":221,"text":1327},{"id":1368,"depth":221,"text":1369},{"id":1399,"depth":221,"text":1400},{"id":1454,"depth":221,"text":1455},{"id":1507,"depth":221,"text":1508},{"id":1538,"depth":221,"text":1539},{"id":726,"depth":221,"text":727},"2026-03-27","The top ClawHub skills for OpenClaw ranked by actual usefulness: browser automation, code execution, memory plugins, Slack, GitHub integrations and more. Updated May 2026 with new community picks.","/img/blog/best-openclaw-skills.jpg",{},"/blog/best-openclaw-skills","15 min read",{"title":1153,"description":1633},"Best ClawHub Skills for OpenClaw in 2026 — Browser, Code, Memory Ranked","blog/best-openclaw-skills",[1642,1643,1644,1645,1646,1647,1648,1649,1650,1651,1652,1653,1654,1655],"best OpenClaw skills","best OpenClaw skills ClawHub 2026","best ClawHub skills 2026","OpenClaw skills to install","top OpenClaw ClawHub skills","popular OpenClaw skills","recommended OpenClaw skills","OpenClaw developer skills","OpenClaw productivity skills","OpenClaw skills list March 2026","safest OpenClaw skills","OpenClaw skills security vetted","OpenClaw GitHub skill","OpenClaw Google Workspace skill","2026-04-02","KaKfuRYqMRLHUB9ZrcgwOSZFR4sr-RMu9s19_kFFYow",{"id":1659,"title":1660,"author":1661,"body":1662,"category":788,"date":2136,"description":2137,"extension":452,"featured":791,"image":2138,"imageHeight":793,"imageWidth":793,"meta":2139,"navigation":230,"path":2140,"readingTime":1637,"seo":2141,"seoTitle":2142,"stem":2143,"tags":2144,"updatedDate":2136,"__hash__":2153},"blog/blog/cheapest-openclaw-ai-providers.md","Cheapest OpenClaw AI Providers: 5 Alternatives to OpenAI That Cut Costs 80%",{"name":7,"role":8,"avatar":9},{"type":11,"value":1663,"toc":2124},[1664,1669,1672,1675,1678,1681,1684,1691,1694,1698,1701,1704,1707,1710,1713,1721,1727,1731,1737,1743,1746,1749,1759,1762,1765,1771,1775,1780,1783,1786,1789,1792,1795,1801,1805,1810,1813,1816,1819,1822,1828,1834,1842,1846,1851,1854,1857,1872,1881,1888,1891,1897,1908,1912,1917,1920,1926,1937,1943,1950,1956,1960,1963,1966,1970,1973,1980,1983,1990,2000,2005,2013,2017,2020,2026,2032,2041,2053,2056,2062,2071,2073,2078,2081,2086,2089,2094,2108,2113,2116,2121],[14,1665,1666],{},[1181,1667,1668],{},"Your OpenClaw agent doesn't need GPT-4o for everything. Here are the providers that cost a fraction and work just as well.",[14,1670,1671],{},"My OpenAI dashboard showed $147. Fourteen days. One agent.",[14,1673,1674],{},"I'd set up my OpenClaw instance on a Friday, pointed it at GPT-4o because that's what every tutorial recommended, and let it run. Morning briefings. Email triage. Calendar management. A few research tasks. Nothing exotic.",[14,1676,1677],{},"Two weeks later, $147. For an AI assistant that mostly checked my calendar and summarized emails.",[14,1679,1680],{},"I pulled up the token logs and did the math. GPT-4o at $2.50 per million input tokens and $10 per million output tokens sounds reasonable in isolation. But OpenClaw agents are hungry. Heartbeats every 30 minutes. Sub-agents spawning for parallel tasks. Context windows that grow silently as cron jobs accumulate history.",[14,1682,1683],{},"The tokens add up. Fast.",[14,1685,1686,1687,1690],{},"Here's the thing: the ",[74,1688,1689],{},"cheapest OpenClaw AI provider isn't always the worst one",". In 2026, there are models that cost 90% less than GPT-4o and perform just as well for the kind of work most agents actually do. Some of them are better at tool calling. Some have larger context windows. One of them is literally free.",[14,1692,1693],{},"This is the guide I wish I'd read before handing OpenAI $147 for two weeks of calendar checks.",[39,1695,1697],{"id":1696},"why-openai-is-the-default-and-why-thats-costing-you","Why OpenAI is the default (and why that's costing you)",[14,1699,1700],{},"OpenAI is the default recommendation in most OpenClaw tutorials for a simple reason: familiarity. Everyone has an OpenAI account. The API is well-documented. GPT-4o is genuinely good.",[14,1702,1703],{},"But \"good\" and \"cost-effective for an always-on agent\" are very different things.",[14,1705,1706],{},"OpenClaw agents don't work like a ChatGPT conversation. They run continuously. They process heartbeats (periodic status checks) every 30 minutes using your primary model. They spawn sub-agents for parallel work. They execute skills that require multiple model calls per task.",[14,1708,1709],{},"A single browser automation task can consume 50-200+ steps, with each step using 500-2,000 tokens. At GPT-4o pricing, that's $0.50-2.00 per complex task. Run a few of those daily and your monthly bill climbs past $100 easily.",[14,1711,1712],{},"The viral Medium post \"I Spent $178 on AI Agents in a Week\" captured this pain perfectly. Most of that spend was GPT-4o running tasks that didn't need GPT-4o.",[14,1714,1715,1716,1720],{},"For a deeper look at where OpenClaw API costs actually come from (and how they compound faster than you'd expect), we wrote a ",[1437,1717,1719],{"href":1718},"/blog/openclaw-api-costs","complete breakdown of OpenClaw API costs"," with real monthly projections.",[14,1722,1723],{},[67,1724],{"alt":1725,"src":1726},"OpenClaw API cost breakdown showing GPT-4o token usage across heartbeats, sub-agents, and daily tasks","/img/blog/openclaw-136k-token-overhead-1.jpg",[39,1728,1730],{"id":1729},"_1-anthropic-claude-the-agent-first-provider","1. Anthropic Claude: The agent-first provider",[14,1732,1733,1736],{},[74,1734,1735],{},"Pricing:"," Haiku 4.5: $1/$5 | Sonnet 4.6: $3/$15 | Opus 4.6: $5/$25 (per million tokens, input/output)",[14,1738,1739,1740,57],{},"Claude isn't cheaper than GPT-4o across the board. Sonnet at $3/$15 is actually more expensive per output token. But here's why it's on this list: ",[74,1741,1742],{},"Claude is better at the specific things OpenClaw agents need to do",[14,1744,1745],{},"Tool calling reliability. Long-context accuracy. Prompt injection resistance. Multi-step instruction following. These are the areas where OpenClaw community benchmarks consistently rank Claude above GPT-4o.",[14,1747,1748],{},"The real savings come from Haiku 4.5 at $1/$5. That's 60% cheaper than GPT-4o on input and 50% cheaper on output. And for heartbeats, calendar lookups, simple queries, and sub-agent tasks, Haiku handles them beautifully.",[14,1750,1751,1754,1755,1758],{},[74,1752,1753],{},"The smart setup:"," Sonnet as your primary model, Haiku for heartbeats and sub-agents, Opus available via ",[50,1756,1757],{},"/model opus"," for complex reasoning when you need it. This tiered approach typically costs $40-70/month compared to $100-200 with GPT-4o for everything.",[14,1760,1761],{},"Claude isn't the cheapest option. It's the option where you get the most capability per dollar on agent-specific tasks.",[14,1763,1764],{},"OpenClaw's founder, Peter Steinberger, recommended Anthropic models before joining OpenAI. That recommendation still holds for most serious agent workloads.",[14,1766,1767],{},[67,1768],{"alt":1769,"src":1770},"Claude model tiers showing Haiku, Sonnet, and Opus pricing with recommended OpenClaw task assignments","/img/blog/openclaw-routing-tiers.jpg",[39,1772,1774],{"id":1773},"_2-deepseek-the-028-option-that-actually-works","2. DeepSeek: The $0.28 option that actually works",[14,1776,1777,1779],{},[74,1778,1735],{}," DeepSeek V3.2: $0.28/$0.42 per million tokens (input/output)",[14,1781,1782],{},"This is where the cost math gets wild.",[14,1784,1785],{},"DeepSeek V3.2 costs roughly 10x less than GPT-4o on input tokens and 24x less on output tokens. For an always-on OpenClaw agent, that difference compounds dramatically. A workload that costs $150/month on GPT-4o drops to approximately $15-20/month on DeepSeek.",[14,1787,1788],{},"And it's not a toy model. Community reports from the OpenClaw GitHub discussions consistently mention DeepSeek alongside Claude as the two providers that work best for agent tasks. It's particularly strong at code generation and debugging.",[14,1790,1791],{},"The tradeoffs are real though. DeepSeek's tool calling is less reliable than Claude's on complex multi-step chains. Context tracking over very long conversations can degrade. And if you're processing sensitive data, the provider routes through Chinese infrastructure, which matters for some use cases.",[14,1793,1794],{},"For pure cost optimization on non-sensitive tasks, DeepSeek is hard to beat. Set it as your heartbeat and sub-agent model while keeping a more capable model as your primary, and your bill drops by 70-80%.",[14,1796,1797],{},[67,1798],{"alt":1799,"src":1800},"DeepSeek V3.2 cost comparison against GPT-4o and Claude showing 10-24x savings per million tokens","/img/blog/cheapest-openclaw-deepseek-comparison.jpg",[39,1802,1804],{"id":1803},"_3-google-gemini-free-tier-thats-surprisingly-capable","3. Google Gemini: Free tier that's surprisingly capable",[14,1806,1807,1809],{},[74,1808,1735],{}," Gemini 2.5 Flash free tier: $0 (1,500 requests/day) | Paid: $0.075/$0.30 per million tokens",[14,1811,1812],{},"Yes, free. Google AI Studio offers a free tier for Gemini 2.5 Flash with 1,500 requests per day and a 1 million token context window. No credit card required.",[14,1814,1815],{},"For personal OpenClaw use (morning briefings, calendar management, basic research), the free tier is often enough. 1,500 requests per day is surprisingly generous for a single-user agent.",[14,1817,1818],{},"Even the paid tier at $0.075 per million input tokens is absurdly cheap. That's 33x cheaper than GPT-4o. A moderate usage pattern that costs $100/month on OpenAI costs roughly $3 on Gemini Flash.",[14,1820,1821],{},"The limitation: Gemini's tool calling isn't as reliable as Claude or even GPT-4o for complex chains. It handles straightforward tasks well but can stumble on multi-step reasoning that requires precise instruction following.",[14,1823,1824,1827],{},[74,1825,1826],{},"Best used for:"," heartbeats, simple lookups, data parsing, and as a fallback model. Not recommended as your sole primary model for complex agent workflows.",[14,1829,1830],{},[67,1831],{"alt":1832,"src":1833},"Google Gemini free tier details showing 1500 daily requests and 1M token context window for OpenClaw","/img/blog/cheapest-openclaw-gemini-free.jpg",[14,1835,1836,1837,1841],{},"To understand which tasks need a powerful model versus which tasks can run on something cheap, our guide to ",[1437,1838,1840],{"href":1839},"/blog/how-does-openclaw-work","how OpenClaw works under the hood"," explains the agent architecture and where model calls actually happen.",[39,1843,1845],{"id":1844},"_4-openrouter-one-api-key-200-models-automatic-routing","4. OpenRouter: One API key, 200+ models, automatic routing",[14,1847,1848,1850],{},[74,1849,1735],{}," Varies by model (typically 0-5% markup over direct provider pricing)",[14,1852,1853],{},"OpenRouter isn't a model provider. It's a routing layer. One API key gives you access to 200+ models across every major provider, and you can switch between them without managing separate API keys for each.",[14,1855,1856],{},"Here's why that matters for OpenClaw.",[14,1858,1859,1860,1863,1864,1867,1868,1871],{},"The ",[50,1861,1862],{},"/model"," command lets you switch models mid-conversation. With OpenRouter, you type ",[50,1865,1866],{},"/model deepseek/deepseek-v3.2"," and you're on DeepSeek. ",[50,1869,1870],{},"/model anthropic/claude-sonnet-4.6"," switches to Claude. No config file edits. No gateway restarts.",[14,1873,1874,1880],{},[1437,1875,1879],{"href":1876,"rel":1877},"https://www.youtube.com/results?search_query=openclaw+openrouter+setup+model+switching+2026",[1878],"nofollow","Watch on YouTube: OpenClaw Multi-Model Setup with OpenRouter"," (Community content)\nIf you want to see how OpenRouter's model switching works in practice with OpenClaw (including the auto-routing feature that selects the cheapest capable model per request), this community walkthrough covers the full configuration and real-time cost comparison.",[14,1882,1883,1884,1887],{},"But the real savings feature is ",[50,1885,1886],{},"openrouter/auto",". Set this as your model and OpenRouter automatically routes each request to the most cost-effective model based on the complexity of the prompt. Simple heartbeats go to cheap models. Complex reasoning gets routed to capable ones. You save money without manually managing model tiers.",[14,1889,1890],{},"The tradeoff: a small markup on token prices (typically under 5%), and you're adding a routing layer which occasionally introduces latency. For most users, the convenience of one API key and automatic cost optimization is worth it.",[14,1892,1893],{},[67,1894],{"alt":1895,"src":1896},"OpenRouter auto-routing diagram showing automatic model selection based on task complexity","/img/blog/cheapest-openclaw-openrouter-routing.jpg",[14,1898,1899,1900,1903,1904,1907],{},"If you don't want to think about model routing at all, if you want automatic cost optimization with zero configuration and built-in anomaly detection that pauses your agent before costs spiral, ",[1437,1901,1902],{"href":1526},"Better Claw handles all of this"," at ",[1437,1905,1906],{"href":1444},"$19/month per agent",". BYOK, 60-second deploy, and you can point it at any of these providers.",[39,1909,1911],{"id":1910},"_5-ollama-local-models-0-per-month-forever","5. Ollama (local models): $0 per month, forever",[14,1913,1914,1916],{},[74,1915,1735],{}," $0 API cost. Hardware and electricity only.",[14,1918,1919],{},"Running models locally through Ollama eliminates API costs entirely. Llama 3.3 70B, Mistral, Qwen 2.5: they all run on your machine, fully private, with no token charges.",[14,1921,1922,1925],{},[74,1923,1924],{},"The math:"," A Mac Mini M4 with 16GB RAM runs 7-8B models at 15-20 tokens per second. That's fast enough for most agent tasks. Larger models (30B+) need more RAM or a dedicated GPU.",[14,1927,1928,1929,1932,1933,1936],{},"For OpenClaw specifically, the ",[50,1930,1931],{},"hermes-2-pro"," and ",[50,1934,1935],{},"mistral:7b"," models are recommended for tool calling reliability. They're not Claude or GPT-4o, but for heartbeats, simple queries, and privacy-sensitive operations, they're genuinely useful.",[14,1938,1939,1942],{},[74,1940,1941],{},"The honest reality:"," local models in 2026 still can't match cloud providers on complex multi-step reasoning, long-context accuracy, or sophisticated tool use. The community consensus in OpenClaw's GitHub discussions is clear: local models work for experimentation and privacy-first setups, but cloud models are better for production agent workflows.",[14,1944,1945,1946,57],{},"The sweet spot is hybrid: local models for heartbeats and simple tasks, cloud models for complex reasoning. OpenClaw supports this natively through its ",[1437,1947,1949],{"href":1948},"/blog/openclaw-model-routing","model routing configuration",[14,1951,1952],{},[67,1953],{"alt":1954,"src":1955},"Ollama local model setup showing zero API cost with hardware requirements for different model sizes","/img/blog/cheapest-openclaw-ollama-local.jpg",[39,1957,1959],{"id":1958},"the-provider-nobody-talks-about-minimax","The provider nobody talks about: MiniMax",[14,1961,1962],{},"Quick honorable mention. MiniMax offers a $10/month plan with 100 prompts every 5 hours. Peter Steinberger himself recommended it during community discussions. It's not on the level of Opus, but community members describe it as \"competent enough for most tasks.\"",[14,1964,1965],{},"For budget-conscious users who want a flat monthly rate instead of per-token billing, it's worth testing. The predictability alone can be valuable when you're worried about runaway agent costs.",[39,1967,1969],{"id":1968},"the-real-problem-isnt-the-provider-its-the-architecture","The real problem isn't the provider. It's the architecture.",[14,1971,1972],{},"Here's what I've learned after months of optimizing OpenClaw costs across different providers.",[14,1974,1975,1976,1979],{},"Switching from GPT-4o to DeepSeek saves you money. Setting up ",[1437,1977,1978],{"href":1948},"model routing"," (different models for different task types) saves you more. But the biggest cost driver in OpenClaw isn't the per-token price. It's uncontrolled context growth.",[14,1981,1982],{},"Cron jobs accumulate context indefinitely. A task scheduled to check emails every 5 minutes eventually builds a 100,000-token context window. What starts at $0.02 per execution grows to $2.00 per execution regardless of which provider you use.",[14,1984,1859,1985,1989],{},[1437,1986,1988],{"href":1987},"/blog/openclaw-memory-fix","memory compaction bug in OpenClaw"," makes this worse. Context compaction can kill active work mid-session, and the workarounds require manual token limits in every skill config.",[14,1991,1992,1993,1932,1996,1999],{},"Set ",[50,1994,1995],{},"maxContextTokens",[50,1997,1998],{},"maxIterations"," in your skill configurations. Set daily spending caps on OpenRouter or your provider's dashboard. Monitor your token usage weekly. These operational habits matter more than which provider you choose.",[14,2001,2002],{},[74,2003,2004],{},"The cheapest provider in the world can't save you from a runaway agent loop burning tokens at 3 AM.",[14,2006,2007,2008,2012],{},"For a look at what tasks are worth running through a premium model versus which ones can safely run on the cheapest option available, our guide to the ",[1437,2009,2011],{"href":2010},"/blog/best-openclaw-use-cases","best OpenClaw use cases"," ranks workflows by complexity and cost.",[39,2014,2016],{"id":2015},"pick-your-fighter-a-practical-recommendation","Pick your fighter (a practical recommendation)",[14,2018,2019],{},"For most people reading this, here's what I'd actually recommend:",[14,2021,2022,2025],{},[74,2023,2024],{},"If you're just starting out:"," Gemini 2.5 Flash free tier. Zero risk. Learn how OpenClaw works without spending anything. Upgrade to a paid provider when you outgrow the free limits.",[14,2027,2028,2031],{},[74,2029,2030],{},"If you want the best quality-to-cost ratio:"," Claude Sonnet 4.6 as primary, Haiku 4.5 for heartbeats and sub-agents. This is what most serious OpenClaw users run. Expect $40-70/month.",[14,2033,2034,2037,2038,2040],{},[74,2035,2036],{},"If cost is the priority:"," DeepSeek V3.2 for everything except complex reasoning. Use Claude or GPT-4o on-demand via ",[50,2039,1862],{}," for the hard stuff. Expect $15-30/month.",[14,2042,2043,2046,2047,1903,2050,2052],{},[74,2044,2045],{},"If you don't want to think about any of this:"," OpenRouter auto-routing, or ",[1437,2048,2049],{"href":1526},"Better Claw",[1437,2051,1906],{"href":1444}," with BYOK and zero-config deployment.",[14,2054,2055],{},"The AI model market is getting cheaper every quarter. Opus 4.5 at $5/$25 is 66% cheaper than Opus 4.1 was at $15/$75. The trend is clear. But until prices hit zero (they won't), smart provider selection and model routing are the most impactful cost levers you have.",[14,2057,2058,2061],{},[74,2059,2060],{},"Stop paying GPT-4o prices for calendar checks."," Your agent will work just as well. Your wallet will thank you.",[14,2063,2064,2065,2070],{},"If you've been wrestling with API costs, config files, and model routing, and you'd rather just deploy an agent that works, ",[1437,2066,2069],{"href":2067,"rel":2068},"https://app.betterclaw.io/sign-in",[1878],"give Better Claw a try",". It's $19/month per agent, BYOK with any of the providers above, and your first agent deploys in about 60 seconds. We handle the infrastructure, the model routing, and the cost monitoring. You focus on building workflows.",[39,2072,727],{"id":726},[14,2074,2075],{},[74,2076,2077],{},"What are the cheapest AI providers for OpenClaw agents?",[14,2079,2080],{},"The cheapest cloud providers for OpenClaw in 2026 are DeepSeek V3.2 at $0.28/$0.42 per million tokens and Google Gemini 2.5 Flash at $0.075/$0.30 (with a free tier offering 1,500 requests per day). For zero-cost operation, Ollama lets you run local models like Llama 3.3 and Mistral with no API charges. Claude Haiku 4.5 at $1/$5 offers the best balance of low cost and agent-specific reliability.",[14,2082,2083],{},[74,2084,2085],{},"How does Claude compare to GPT-4o for OpenClaw?",[14,2087,2088],{},"Claude models (particularly Sonnet and Haiku) consistently outperform GPT-4o on the tasks that matter most for OpenClaw: tool calling reliability, long-context accuracy, and prompt injection resistance. GPT-4o is faster on simple tasks and has broader community support. Claude Sonnet 4.6 at $3/$15 is more expensive per output token than GPT-4o at $2.50/$10, but the improved agent performance often means fewer retries and lower total cost.",[14,2090,2091],{},[74,2092,2093],{},"How do I switch AI providers in OpenClaw?",[14,2095,2096,2097,2100,2101,2103,2104,2107],{},"Edit your ",[50,2098,2099],{},"~/.openclaw/openclaw.json"," file to change the model provider and API key, then restart your gateway. For quick switching mid-conversation, use the ",[50,2102,1862],{}," command (for example, ",[50,2105,2106],{},"/model anthropic/claude-sonnet-4-6","). OpenRouter simplifies this further by giving you one API key for 200+ models. The switch takes seconds and doesn't require reinstallation.",[14,2109,2110],{},[74,2111,2112],{},"How much does it cost to run an OpenClaw agent per month?",[14,2114,2115],{},"Monthly costs vary by provider and usage: $80-200 with GPT-4o for everything, $40-70 with Claude Sonnet plus Haiku routing, $15-30 with DeepSeek for most tasks, or $0-5 with Gemini free tier or local models. These are API costs only. Hosting adds $5-29/month depending on whether you self-host on a VPS or use a managed platform like Better Claw. BYOK means you control the API spend regardless of hosting.",[14,2117,2118],{},[74,2119,2120],{},"Is DeepSeek reliable enough for production OpenClaw agents?",[14,2122,2123],{},"DeepSeek V3.2 is reliable for most standard agent tasks and excels at code generation. Community reports confirm it works well for daily operations. The tradeoffs: tool calling can be less precise than Claude on complex multi-step chains, and data routes through Chinese infrastructure, which matters for sensitive workloads. For heartbeats, sub-agents, and non-sensitive tasks, it's a solid budget choice. For critical workflows, pair it with a more capable model as your primary.",{"title":200,"searchDepth":221,"depth":221,"links":2125},[2126,2127,2128,2129,2130,2131,2132,2133,2134,2135],{"id":1696,"depth":221,"text":1697},{"id":1729,"depth":221,"text":1730},{"id":1773,"depth":221,"text":1774},{"id":1803,"depth":221,"text":1804},{"id":1844,"depth":221,"text":1845},{"id":1910,"depth":221,"text":1911},{"id":1958,"depth":221,"text":1959},{"id":1968,"depth":221,"text":1969},{"id":2015,"depth":221,"text":2016},{"id":726,"depth":221,"text":727},"2026-03-10","Stop overpaying for OpenClaw. DeepSeek at $0.28, Gemini free tier, Claude Haiku at $1. Five providers that cut your agent costs 50-90%.","/img/blog/cheapest-openclaw-ai-providers.jpg",{},"/blog/cheapest-openclaw-ai-providers",{"title":1660,"description":2137},"5 Cheapest OpenClaw AI Providers (Save 80% vs OpenAI)","blog/cheapest-openclaw-ai-providers",[2145,2146,2147,2148,2149,2150,2151,2152],"cheapest OpenClaw AI provider","OpenClaw API costs","OpenClaw DeepSeek","OpenClaw Claude vs GPT","OpenRouter OpenClaw","reduce OpenClaw spending","OpenClaw model pricing","cheap AI agent hosting","LWLTptPQgUuPwnDBYYYMj_utp8Jk5GxRac8hPd7yetM",1781183475856]