# BetterClaw

> No-code AI agent platform with a free plan. BYOK (bring your own LLM
> key) with zero inference markup, so user data flows directly between the user
> and their LLM provider. Visual agent builder — no YAML, no Docker, no config
> files. Every skill passes a 4-layer security audit.

This file is a curated map for LLMs and answer engines. For the full machine
index, see the sitemap: https://www.betterclaw.io/sitemap.xml

BetterClaw is a managed, no-code AI agent platform: build agents in minutes, use any of 28+ model providers with your own key at zero markup, and connect 120+ apps. (Updated October 2026)

Plans are tiered by agents and monthly credits. One credit is one minute of your agent's machine running. Managed LLM usage is also converted into credits, but with your own key (BYOK) model usage is billed by your provider and never draws credits. Residential proxy usage also draws from credits. Credits reset every month.
Annual billing saves 20% on both paid plans.

## Free Plan
BetterClaw's free plan is a hosted AI agent you can keep forever: 1 agent, 100 credits a month and 3 connectors, with no credit card and no expiry. (Updated October 23, 2026)
- Price: $0 per month, no credit card, no expiry
- Agents: 1 autonomous AI agent, hosted in the cloud
- Credits: 100 per month, reset monthly (work pauses, never bills, when they run out)
- Connectors: 3 of 120+ apps (one account each)
- Models: Any of 28+ providers with your own key (BYOK only), zero markup
- Runs: 24/7 on schedules you set; no server, no Docker, no code
- Memory: 7 days
- Security: Sandboxed, AES-256, secrets auto-purge after 5 minutes, action approval
- Upgrade: Basic $19/month, Pro $49/month (optional, never forced)
- Skills available: Basic
- Team seats: None
- Support: Community (Discord, Reddit)
- Not included: Telegram and Slack (Basic and up), ad integrations, proxy, multi-account, custom webhooks, managed keys
- URL: https://www.betterclaw.io/free-plan
- Also in Spanish: https://www.betterclaw.io/es/agente-ia-gratis · Arabic: https://www.betterclaw.io/ar/wakil-dhakaa-istinaei-majani · Indonesian: https://www.betterclaw.io/id/agen-ai-gratis
- Sign up: https://app.betterclaw.io/sign-in

## Common Questions — Free AI Agents
- Is there a truly unlimited free AI agent? No. Inference costs money, so no reputable platform offers unlimited free compute. BetterClaw's limit is 100 credits per month, resetting monthly, forever, with no card on file. What is unlimited is time: no trial countdown, no expiry.
- Is Claude free? Claude is a language model, not an agent platform. Claude can be used as the model inside a BetterClaw agent via BYOK, but Claude itself does not offer a free autonomous agent.
- Are there any free AI agents? Yes. Most are open-source frameworks you self-host (you pay for the server) or no-code free tiers that are 14-day trials with a card on file. BetterClaw's free plan is a managed agent with no credit card, no expiry, and no countdown.
- Is there a free AI coding agent? Yes, in three senses: free open-source coding agents you run yourself (OpenCode, Pi, Hermes Agent; you still pay for the model unless you use a free tier or a local model), free tiers of editor-based coding tools, and BetterClaw's free plan, which runs an agent on a schedule to review pull requests, draft changelog entries and triage issues. It is not an in-editor assistant. No hosted coding agent is free and unlimited, because every model call costs someone money.
- Is there a free AI agent builder? Yes. BetterClaw is a no-code AI agent builder and the free plan includes the builder itself: describe the job or start from a ready-made workflow, pick apps, choose a model, set a schedule. Open-source builders such as n8n, Flowise and Langflow are free to download but need a server and Docker.
- Can I run a free AI agent 24/7 without a VPS? Yes. Agents run on managed cloud infrastructure. No VPS, no Docker, no server. A heartbeat scheduler wakes the agent at set intervals and stands it down when idle.
- How much does BetterClaw cost per month? Free is $0/month forever with no credit card (1 agent, 100 credits). Basic is $19/month (1 agent, 500 credits, 10 connectors, 2 vCPU / 4 GB RAM server). Pro is $49/month ($39/month billed annually; 5 agents, 12,000 credits). Business is $149/month (25 agents). Inference is BYOK with zero markup on every plan.

## Pro Plan
- Price: $49 per month ($39 per month billed annually, $468 per year)
- Hosted agents: 5
- Credits per month: 12,000
- Connectors: Unlimited, multi-account
- LLM inference: Bring your own key (BYOK) or managed keys, zero markup
- Memory retention: 90 days
- Team seats: 2
- Server: 4 vCPU / 8 GB RAM
- Skills available: Unlimited skills (security-audited curated library + custom skills)
- Ad integrations: Meta, Google, X
- Webhooks: Telegram, Slack
- Support: 24h email
- Add-ons: extra agent $12/mo, 1,000-credit pack $5, extra seat $10/mo
- URL: https://www.betterclaw.io/pricing

## Business Plan
- Price: $149 per month ($119 per month billed annually, $1,428 per year)
- Hosted agents: 25
- Credits per month: 40,000
- Connectors: Unlimited, multi-account
- Memory retention: Unlimited
- Team seats: 5
- Dedicated server: 8 vCPU / 16 GB RAM
- Webhooks: Telegram, Slack, custom
- Support: 12h priority
- Add-ons: extra agent $8/mo, 1,000-credit pack $4, extra seat $8/mo
- URL: https://www.betterclaw.io/pricing

## Enterprise Plan
- Price: Custom, negotiated
- Hosted agents: Unlimited
- Credits per month: Custom
- Support: Dedicated account manager
- Includes: custom infrastructure, data residency, custom SLA
- Contact: https://www.betterclaw.io/pricing

## Key Differentiators
- No-code AI agent platform with a free plan (as of April 2026)
- BYOK with zero inference markup on both free and paid plans
- 4-layer security audit on every skill (malicious code review, exfiltration analysis, prompt injection testing, permission scope verification)
- Visual agent builder (no YAML, no Docker, no config files)
- Built-in action approval guardrails and LLM cost controls

## AppSumo Lifetime Deal
- [BetterClaw AppSumo lifetime deal](https://www.betterclaw.io/appsumo): one-time payment on AppSumo for lifetime access. Tier 1 $79 (2 Agents, 2,000 Credits per month); Tier 2 $149 (4 Agents, 6,000 Credits per month); Tier 3 $269 (8 Agents, 12,000 Credits per month); Tier 4 $399 (15 Agents, 20,000 Credits per month); Tier 5 $549 (30 Agents, 30,000 Credits per month). Every tier: BYOK, Unlimited chat, Dedicated agent instance, Skill installer & ad-free integration, Integrations, Hourly cron minimum, Built-in workflows. Credits refresh monthly; one credit is used each time an agent runs a task; model costs are billed by your own provider (BYOK). Activate the AppSumo license within 60 days; 60-day refund through AppSumo. Deal listing: https://appsumo.com/products/betterclaw/
- [BetterClaw reviews](https://www.betterclaw.io/reviews): AppSumo and Product Hunt reviews quoted word for word with links to the originals (AppSumo-incentivized reviews labelled; team reviews excluded).

## Desktop app
- [Download the BetterClaw desktop app](https://www.betterclaw.io/downloads): BetterClaw Desktop lets BetterClaw agents use AI models running on your own computer through Ollama, LM Studio, llama.cpp or any custom OpenAI-compatible address (on your computer or local network), so you don't pay per token to an AI provider. No OLLAMA_HOST change, port forwarding or tunnel needed; it shows each model's context size and tool support, and runs in the menu bar or system tray. Chat, agent setup, workflows, connectors and billing stay in the web app. Runs on macOS 13+ (Apple Silicon only; no Intel Mac build; signed and notarized, auto-updates), Windows 10 and 11 (64-bit Intel or AMD; no Windows on ARM; no WSL, Docker or Virtual Machine Platform needed) and Linux (64-bit AppImage; no .deb or Flatpak yet). Sign-in stored in Keychain, Credential Manager or Secret Service; no local database; prompts and answers never written to disk; it never asks for an AI provider API key. Not open source. Version 0.1.0, released October 5, 2026.

## MCP
- [BetterClaw MCP](https://www.betterclaw.io/mcp): BetterClaw is an MCP client. Attach any remote MCP server (Streamable HTTP or SSE) to an agent from the dashboard by pasting its URL and signing in, with no config file or restart; its tools sit next to the 120+ built-in connectors in the same run. OAuth callback and token refresh handled by BetterClaw; Streamable HTTP vs SSE detected automatically; tools picked per agent with checkboxes; tokens AES-256 encrypted. Stdio servers not supported yet. Works with local models through BetterClaw Desktop when the model supports tool calling. Available on paid plans (from $19/mo); MCP tool calls use credits like built-in tools; no limit on servers. Includes a side-by-side of how OpenClaw (`mcp.servers`, `openclaw mcp` CLI) and Hermes Agent (`mcp_servers` in `~/.hermes/config.yaml`) set up MCP servers. BetterClaw cannot yet be used as an MCP server from Claude or Cursor (planned).

## Integrations
- [All integrations](https://www.betterclaw.io/integrations): 120+ apps an agent can connect to, plus any remote MCP server.
- [Gmail AI agent](https://www.betterclaw.io/integrations/gmail): Connector (Sign in with Google). It sorts your inbox, drafts replies in your voice, chases quiet threads and logs receipts. You stay in charge.
- [Google Sheets AI agent](https://www.betterclaw.io/integrations/google-sheets): Connector (Sign in with Google). It logs leads, receipts and applicants, updates rows, builds charts and keeps reports current, in plain English.
- [Google Calendar AI agent](https://www.betterclaw.io/integrations/google-calendar): Connector (Sign in with Google). Get a morning briefing, a prep note before outside meetings, open slots found and events added in plain English.
- [Google Docs AI agent](https://www.betterclaw.io/integrations/google-docs): Connector (Sign in with Google). It drafts articles, proposals, meeting notes and reports in a Google Doc for you to edit. It never publishes.
- [Slack AI agent](https://www.betterclaw.io/integrations/slack): Slack bot (Add to Slack) + connector. Message it in a DM or @mention it in a channel, and let it post standups, error summaries and alerts. Set up in one click.
- [HubSpot AI agent](https://www.betterclaw.io/integrations/hubspot): Connector (Sign in with HubSpot or private app token). 281 HubSpot actions, set up in plain English.
- [Notion AI agent](https://www.betterclaw.io/integrations/notion): Official MCP server (sign in). Search pages, update databases and file meeting notes, with per-tool control.
- [Stripe AI agent](https://www.betterclaw.io/integrations/stripe): Official MCP server (sign in, live/sandbox). Recover failed payments, answer billing questions and report revenue, with approvals.
- [Linear AI agent](https://www.betterclaw.io/integrations/linear): Official MCP server (sign in). Auto standups, weekly reports and error tickets, with read-only mode.
- [Google Search Console AI agent](https://www.betterclaw.io/integrations/google-search-console): Connector (Sign in with Google). Weekly CTR fixes, ranking-drop alerts, sitemap checks and article ideas from your own search data.

## Tools
- [Agent Error Decoder](https://www.betterclaw.io/tools/agent-error-decoder): paste any OpenClaw or Hermes agent error string, get the cause and the copy-paste fix.

## Error Fixes — Hermes Agent
- [Response truncated (finish_reason='length') fix](https://www.betterclaw.io/blog/hermes-response-truncated-fix): Hermes output truncation on v0.21.5 — model-server output defaults (the max_tokens settings were removed in v0.21.1), Ollama's VRAM-based context size, a full context window, and reasoning models using up the budget; plus Pi agent's "Response was truncated before completion".
- [Hermes auth error fix](https://www.betterclaw.io/blog/hermes-auth-error-fix): "auth handler error authenticating" checked against Hermes v0.21.7 (13 Oct 2026) — the top cause is not auth at all (quota exhaustion 429 reported as "Provider authentication failed", #89401), plus BSM placeholder secrets clobbered by dotenv override (#74265), the two xAI Grok 403s (stale OAuth token #108122, subscription-tier gate #26847), provider-side intermittent 401s on MiniMax/GLM (#8283), and the April bugs that are now closed. Also provider-specific 401s: Anthropic (`hermes auth add anthropic --type api-key|oauth`; OAuth only on Claude Max with extra-usage credits) and Qwen/DashScope region mismatch (`alibaba` vs `alibaba-cn`).
- [Hermes installation errors fix](https://www.betterclaw.io/blog/hermes-agent-installation-errors-fix): the six install errors still live — PEP 668 "externally-managed-environment" on Ubuntu 24.04 and Debian 12+, the real Python range (>=3.11,<3.14, so 3.12 and 3.13 are fine and the "3.11 only" claim is wrong), "hermes: command not found" and PATH, native-Windows path and encoding failures (#16201), Docker permission denied because the container runs as UID 10000 not 1000, and PyPI serving 0.19.0 while the GitHub release is v0.21.3 — plus five errors already patched.
- [Hermes error 400: every cause and fix](https://www.betterclaw.io/blog/hermes-agent-error-400): "Non-retryable client error (HTTP 400)" — model mismatch, models without tool support, Gemini dual auth, missing reasoning_content, OpenRouter credits, oversized payload, wrong endpoint, dots-vs-hyphens, with a copy-paste fix for each.
- [Hermes agent not working: every error and fix](https://www.betterclaw.io/blog/hermes-agent-not-working): the Hermes error index, led by the three most-searched errors — "Hermes couldn't start" (the Desktop startup screen), "backend exited before it became ready" (a broken venv after a partial update), and "Chat unavailable: 1" (the dashboard chat process exiting) — then debugged by layer: "command not found: hermes" and PATH, install script hangs, Python version conflict, WSL2 on Windows, auth and HTTP 400 failures, empty tool_calls, JSON parse cascades, web tools failing on local models, local Ollama models hanging when tools are defined (fix: `model.streaming false`, not the mythical `api_mode: chat`), the 2,048-token context trap, memory lost between sessions, ghost sessions and `hermes sessions prune` (there is no `hermes sessions clean`), Telegram/Discord gateway crashes, Tirith blocking (Tirith was removed in v0.22.0), the v0.21 Pantheon bugs through v0.21.7, what changed in v0.22.0 to v0.22.4 (current release, 21 Oct 2026; gateway starts after a reboot again from v0.22.2), and what `hermes doctor --fix` actually repairs.
- [Hermes context length: minimums, limits, and the fix](https://www.betterclaw.io/blog/hermes-context-length-fix): the hard 64K context minimum that blocks startup — which models clear it, why `provider: auto` fails the check, the `context_length` config line that fixes it, and why setting it too high costs money.
- [Hermes Agent slow or forgetting after compaction: 6 fixes](https://www.betterclaw.io/blog/hermes-agent-compaction-slow-fix): why a compaction pauses Hermes for minutes on a local model (the summary runs on the main model; at 300 tokens/s just reading 80K tokens takes ~4.5 min) and why it loses detail. How compression triggers (`compression.threshold` 0.50, raised to at least 75% for windows under 512K; `threshold_tokens`), what's pinned and summarised, and six fixes: a faster `auxiliary.compression` model, `proactive_prune_tokens`, `idle_compact_after_seconds`, the VCC Engine plugin (`hermes plugins install danbiocchi/hermes-vcc-engine --enable`), updating to v0.22.2+, and fresh sessions / `/compress`.

## Error Fixes — OpenClaw
- ["The AI provider rejected this request" / "provider rejected the request schema or tool payload"](https://www.betterclaw.io/blog/ai-provider-rejected-request-fix): the hub for every "provider rejected" string. Six causes of the OpenClaw 400 (failover replay, JSON Schema keywords, no tool support, message shape with compat.strictMessageKeys and compat.requiresStringContent, reasoning parameters, a WAF in front of the endpoint), plus the variants that are a different error: "check the selected model and proxy settings (400)" (model ID or base URL), "your AI provider rejected the API key" (401, not 400), "I kept the raw provider error out of chat" (OpenClaw's newer redacted wording, read `openclaw logs --follow`), "provider rejected a test request" (model access or quota), "could not auto resolve a provider" (a Bifrost error: use provider/model), and "please adjust your message or attachments" (content filter).
- [Model does not support tools](https://www.betterclaw.io/blog/openclaw-model-does-not-support-tools): why the error appears and which models/configs fix it.
- [Ollama fetch failed](https://www.betterclaw.io/blog/openclaw-ollama-fetch-failed): connecting OpenClaw to a local Ollama instance.
- [Local model not working](https://www.betterclaw.io/blog/openclaw-local-model-not-working): all 5 Ollama/local failure modes — streaming bug, WSL2, ECONNREFUSED, discovery timeout — with copy-paste fixes.
- [Config validation failed (ollama.models expected array)](https://www.betterclaw.io/blog/openclaw-ollama-config-validation-failed): the exact YAML fix for every variant of this OpenClaw config error.
- [API key not working](https://www.betterclaw.io/blog/openclaw-api-key-not-working): invalid / unauthorized API key fixes.
- [DeepSeek 503 errors](https://www.betterclaw.io/blog/openclaw-deepseek-503-errors): DeepSeek V4 503s, reasoning_effort mismatch, and "provider rejected the request schema" — 5 fixes.
- [Context window reached](https://www.betterclaw.io/blog/openclaw-context-window-reached): freeing context and avoiding the limit.
- [Memory fix](https://www.betterclaw.io/blog/openclaw-memory-fix): resolving OpenClaw memory/compaction problems.
- [OOM errors (agent crashes)](https://www.betterclaw.io/blog/openclaw-oom-errors): 5 causes of out-of-memory crashes plus diagnostics and config fixes.
- [TypeError: Cannot read properties of undefined (reading 'trim')](https://www.betterclaw.io/blog/openclaw-trim-error): 5 causes — stale sessions.json is #1.
- [Rate limit reached](https://www.betterclaw.io/blog/openclaw-rate-limit): handling provider 429 rate-limit errors.
- [DNS lookup failed](https://www.betterclaw.io/blog/openclaw-dns-lookup-failed): network/DNS resolution fixes.
- [Gateway won't start after updating](https://www.betterclaw.io/blog/openclaw-gateway-wont-start-after-update): ten post-update failures matched to their exact log lines — ERR_MODULE_NOT_FOUND from a pruned dist/, macOS updates run against a live LaunchAgent, config schema drift, unhandled-rejection loops, OOM during sidecar startup, EADDRINUSE from split-brain installs, protocol mismatch after a rollback, plugin load failures, a missing supervisor service, and 2026.9.4's "existing managed handoff lease is incompatible" config-write failure (issue #144742) — plus the safe rollback order.
- [Back up and restore an OpenClaw agent before updating](https://www.betterclaw.io/blog/openclaw-backup-restore-before-update): what actually needs backing up (openclaw.json, the two SQLite databases, credentials/, the workspace), why copying live .sqlite files corrupts backups, `openclaw backup create --verify`, the two-step stage-then-activate restore and its full-second-copy disk-space requirement, and rolling the version back with `openclaw update --tag`.
- [Agent stuck in a loop](https://www.betterclaw.io/blog/openclaw-agent-stuck-in-loop): breaking infinite tool-call loops.
- [Agent hallucinating (describes tasks instead of doing them)](https://www.betterclaw.io/blog/openclaw-agent-hallucination-fix): 5 causes — wrong model, Docker down, skill inactive, loop, or SOUL.md conflict.
- [OpenClaw slow](https://www.betterclaw.io/blog/openclaw-slow): 5 causes of slow responses, ranked — context bloat is #1.
- [Pairing required error](https://www.betterclaw.io/blog/openclaw-pairing-required-error): gateway "pairing required" — cleared browser data, new device, gateway restart, or proxy.
- [Cron not running](https://www.betterclaw.io/blog/openclaw-cron-not-running): scheduled tasks silent — check nextWakeAtMs first; 4 causes.
- [Dashboard / Control UI not working](https://www.betterclaw.io/blog/openclaw-dashboard-not-working): localhost:18789 not loading, auth token rejected, WebSocket failing — 9 fixes.
- [Telegram bot not responding](https://www.betterclaw.io/blog/openclaw-telegram-bot-not-working): 7 fixes — token mismatch is the cause 60% of the time.
- [Discord bot not replying](https://www.betterclaw.io/blog/openclaw-discord-bot-not-working): 6 fixes — Message Content Intent disabled 50% of the time.
- [Skill install failed](https://www.betterclaw.io/blog/openclaw-skill-install-failed): the 3 hidden dependencies — Chromium, Python version, Docker permissions.
- [Docker troubleshooting](https://www.betterclaw.io/blog/openclaw-docker-troubleshooting): 8 Docker errors — permission denied, OOMKilled, port conflicts, broken updates.
- [Claude subscription and OpenClaw: the ban, the reversal, and what to do](https://www.betterclaw.io/blog/openclaw-anthropic-subscription-ban): the April 4 subscription ban, the May 13 reversal, the June 15 pause of the Agent SDK credit, and the 5-minute API key setup that stays reliable.
- [10 common OpenClaw errors](https://www.betterclaw.io/blog/openclaw-common-errors): the 10 most common errors in 2026 with the fastest fix for each.
- [OpenClaw not working](https://www.betterclaw.io/blog/openclaw-not-working): full first-hour error checklist.

## Error Fixes — Claude
- [Rate limit reached on Claude](https://www.betterclaw.io/blog/claude-rate-limit-reached): the four separate limit systems (five-hour session, weekly cap, per-minute API 429s, shared Cowork pool), exactly when each resets, and 5 workarounds starting with prompt caching.
- [Claude Code and ChatGPT rate limits, September 2026](https://www.betterclaw.io/blog/claude-code-chatgpt-rate-limit-alternatives-2026): the two dated changes (OpenAI's five-hour Codex cap returning August 25, Claude Code's weekly limit dropping 17% from the summer boost on September 14), why the "+25%" headline and the "-17%" reality are both true, and the priced alternatives — OpenRouter routing, local Ollama, and free tiers with the catch on each.
- [Virtual Machine Platform not available](https://www.betterclaw.io/blog/claude-virtual-machine-platform-not-available-fix): the single Claude Cowork Virtual Machine Platform fix page. Covers "Virtual Machine Platform not available" and "Claude's workspace requires the Virtual Machine Platform": enabling the feature (PowerShell, Windows Features, Windows 11 25H2 Settings path), Hyper-V, BIOS VT-x/SVM, the yukonSilver detection bug, Windows Home (vmcompute/hns/vfpext check, bcdedit, web Cowork fallback), and admin rights.
- [Activar la Plataforma de máquina virtual (Español)](https://www.betterclaw.io/blog/plataforma-de-maquina-virtual-claude): Spanish-language version — how to turn on Virtual Machine Platform on Windows 11, 10 and Home (Windows Features, PowerShell, DISM), what to do when it does not appear, BIOS keys by brand, the yukonSilver detection bug, Windows Home limits and the September 2026 update bug.
- [Ativar a Plataforma de Máquina Virtual (Português)](https://www.betterclaw.io/blog/ativar-plataforma-de-maquina-virtual): Brazilian Portuguese version — same coverage: enabling the feature, when it does not appear, BIOS keys by brand, yukonSilver, Windows Home and the September 2026 update bug.
- [VM-Plattform aktivieren (Deutsch)](https://www.betterclaw.io/blog/vm-plattform-aktivieren-claude): German version — same coverage: enabling VM-Plattform, when it is missing, BIOS keys by brand, yukonSilver, Windows Home and the September 2026 update bug.
- [Cowork requires modern installer](https://www.betterclaw.io/blog/claude-cowork-modern-installer-error-fix): why the Squirrel-to-MSIX migration breaks Cowork, why the in-app Reinstall button and Check for Updates do nothing, and the clean-uninstall-plus-MSIX-reinstall fix.
- [n8n MCP tools not loading in Claude](https://www.betterclaw.io/blog/n8n-mcp-server-trigger-tools-rejected-by-claude-fix): the connector shows running but zero tools appear — Claude's MCP client now rejects any `outputSchema` declaring JSON Schema draft-07, n8n's `/mcp-server/http` emits exactly that, and the fix is deleting one `$schema` line or setting it to 2020-12.

## Error Fixes — Agents (General)
- [Failed to Fetch error: what it means and how to fix it](https://www.betterclaw.io/blog/failed-to-fetch-error): English explainer — "TypeError: Failed to fetch" means the browser got no response at all (not a 404/500). Per-browser variants (NetworkError on Firefox, Load failed on Safari), the 5 causes (CORS, wrong URL/port/DNS, mixed content, extensions/VPN/CSP, aborted requests), the curl test that separates network from browser problems, a DevTools Network checklist, Node.js "fetch failed" and err.cause codes, and Ollama/n8n/MCP/OpenClaw fixes.
- [Failed to fetch: qué significa](https://www.betterclaw.io/blog/failed-to-fetch-que-significa): Spanish-language explainer — "Failed to fetch" means the request never completed at the network layer, not that the server errored. CORS, wrong URL, port or DNS, mixed content, extension and VPN blocking, and browser-initiated aborts (a reload or navigation, not an AbortController, which throws AbortError).
- [Failed to Fetch: tradução, significado e solução](https://www.betterclaw.io/blog/failed-to-fetch-significado-e-solucao): Brazilian Portuguese explainer — "Failed to fetch" means the browser got no response at all (not a 404/500). Fixes for app users (hard reload, incognito, VPN, network switch, outage) and developers (curl vs CORS, why no-cors hides the bug, Node.js err.cause codes, FastAPI/Flask CORS, n8n in Docker needing host.docker.internal).
- [Failed to Fetch ne demek? Türkçesi, anlamı ve çözümü](https://www.betterclaw.io/blog/failed-to-fetch-ne-demek-cozumu): Turkish explainer — same coverage as the Portuguese post: what "Failed to fetch" means, per-browser message variants (Load failed on Safari, NetworkError on Firefox), a symptom-to-cause table, Node.js "fetch failed" vs browser "Failed to fetch", and Ollama/n8n/OpenClaw fixes.
- [Failed to Fetch: Bedeutung und Lösung](https://www.betterclaw.io/blog/failed-to-fetch-bedeutung-und-loesung): German explainer — same coverage as the Portuguese and Turkish posts, plus why clearing cookies is rarely needed (use a private window instead) and a warning that OLLAMA_HOST=0.0.0.0 exposes the unauthenticated Ollama API to the whole network.
- [Debug MCP tool calls](https://www.betterclaw.io/blog/debug-mcp-tool-calls): a 5-stage decision tree to find why an MCP tool is silently not firing.
- [AI agent slow / high latency](https://www.betterclaw.io/blog/ai-agent-slow-latency-fix): diagnose context bloat, tool lag, and step compounding with a 5-layer framework.
- ["Does not support tools" in Ollama](https://www.betterclaw.io/blog/ollama-models-tool-calling-support): the full tool-calling compatibility table for 38 Ollama models, why gemma3, llama3, and vision models throw the error, and how to check capability with `ollama show`.
- [Qwen3.8 27B tool calls hang in agents](https://www.betterclaw.io/blog/qwen-3-8-27b-tool-calling-fails-agents-fix): the model does support tools — agents hanging on Apple MLX (fixed by Ollama PR #17901, so update Ollama) and `/v1/chat/completions` never returning on AMD ROCm (issue closed without a fix), with the two-minute confirmation test and the `qwen3.6:27b` and `/api/chat` workarounds.
- [Ollama's default context window silently truncates your agent config](https://www.betterclaw.io/blog/ollama-context-window-truncates-agent-config): why agents ignore soul.md or drop skills after a few turns on local Ollama — the VRAM-derived default (4096 under 24 GiB), front-of-prompt truncation with HTTP 200, why `num_ctx` is ignored on the OpenAI-compatible `/v1` endpoint (issue #11409, which hits Hermes), and the `OLLAMA_CONTEXT_LENGTH` / Modelfile fixes verified with `ollama ps`.

## Workflows (ready-made AI agent workflows)
- [Workflow directory](https://www.betterclaw.io/workflows): 28 ready-made AI agent workflows for email, sales, marketing, finance and engineering. Each one is already built: connect your apps (one click each), fill in a few settings in plain English, and it runs on its own schedule. Every page lists each step, the apps it uses and what it sends.
- [Brand Mention Monitor](https://www.betterclaw.io/workflows/brand-mention-monitor): Every morning, finds new mentions of your brand on Reddit, Hacker News, YouTube, review sites and the web, sorted by sentiment, with suggested replies.
- [SEO & AI-Search Article Writer](https://www.betterclaw.io/workflows/seo-article-writer): Every Monday, finds a query your site almost ranks for in Search Console, studies the top results and drafts a better, answer-first article in Google Docs.
- [Daily Content Drafts](https://www.betterclaw.io/workflows/daily-content-drafts): Every weekday morning, finds what's new in your niche and drafts ready-to-edit LinkedIn or X posts in your voice, saved to a Google Doc with sources.
- [Competitor Monitor](https://www.betterclaw.io/workflows/competitor-monitor): Every Monday, compares competitors' pricing, feature and changelog pages with last week, scans for launches and funding, and explains what changed.
- [SEO Click-Through Monitor](https://www.betterclaw.io/workflows/search-console-ctr-monitor): Every Monday, finds Search Console pages with impressions but too few clicks for their position, explains why and suggests better titles and descriptions.
- [Review & Reputation Monitor](https://www.betterclaw.io/workflows/review-reputation-monitor): Every morning, checks Google and review sites like Trustpilot, Yelp and G2 for new reviews, flags the bad ones and drafts a reply to each. It never posts.
- [Content Repurposing](https://www.betterclaw.io/workflows/content-repurposing): Every week, turns your newest blog post, video or podcast into an X thread, LinkedIn post, newsletter blurb, short-video script and carousel outline.
- [Morning Briefing](https://www.betterclaw.io/workflows/morning-briefing): One short email every morning: overnight Gmail sorted by urgency, today's meetings with prep notes and your top three priorities first. Read-only.
- [Follow-Up Reminder](https://www.betterclaw.io/workflows/follow-up-reminder): Every weekday, finds Gmail threads still waiting on a reply, theirs or yours, and drafts a polite nudge in each. Nothing goes out until you press Send.
- [Email Triage](https://www.betterclaw.io/workflows/email-triage): Every hour on weekdays, sorts new Gmail into urgent, needs reply, FYI, newsletter and noise, alerts you to urgent mail and drafts replies in your style.
- [Meeting Notes](https://www.betterclaw.io/workflows/meeting-notes): At the end of each weekday, turns your meeting notes and transcripts into summaries with decisions and action items, and drafts follow-ups to attendees.
- [Meeting Prep](https://www.betterclaw.io/workflows/meeting-prep): Every weekday morning, a one-page brief for each outside meeting: who they are, what's new at their company, your last conversation and what to ask.
- [B2B Prospect Hunter](https://www.betterclaw.io/workflows/b2b-prospect-hunter): Every weekday, finds companies that fit your ideal customer with a reason to buy now, like hiring, funding or a launch, and adds each to a Google Sheet.
- [Proposal Composer](https://www.betterclaw.io/workflows/proposal-composer): Paste your notes after a sales call. It researches the client and drafts a complete proposal in Google Docs, priced only from your services.
- [Lead Qualification](https://www.betterclaw.io/workflows/lead-qualification): Every 2 hours on weekdays, finds new lead emails, researches and scores each Hot, Warm or Cold, logs them in a Sheet and drafts replies to hot leads.
- [Buying Signal Hunter](https://www.betterclaw.io/workflows/buying-signal-hunter): Every weekday, finds people on Reddit, X, Hacker News and forums asking for what you sell, and logs each with a helpful suggested reply. It never posts.
- [Software Spend Auditor](https://www.betterclaw.io/workflows/software-spend-auditor): On the 1st of every month, lists every subscription you pay for from your email receipts in a Google Sheet, and flags price rises, duplicates and renewals.
- [Invoice & Expense](https://www.betterclaw.io/workflows/invoice-expense): Every morning, finds receipts and invoices in Gmail, including PDFs and photos, reads the amounts, sorts them into categories and logs them in a Sheet.
- [Market Watchlist Brief](https://www.betterclaw.io/workflows/market-watchlist-brief): Every weekday before the US market opens, checks your watchlist, explains big moves with the news behind them and lists upcoming earnings. Never trades.
- [PostHog Errors to Linear](https://www.betterclaw.io/workflows/posthog-errors-to-linear): Every morning, pulls the last day of PostHog frontend errors, files each new one as a Linear ticket and posts a Slack summary. Known errors are skipped.
- [Standup Notes from Linear](https://www.betterclaw.io/workflows/linear-standup-notes): Every weekday morning, collects the Linear tickets you closed yesterday, picks up to six for today by priority and posts your standup update to Slack.
- [GCP Error Log Triage](https://www.betterclaw.io/workflows/gcp-error-log-triage): Every morning, pulls Google Cloud error logs, groups them by root cause, files new issues in Linear with personal data removed and posts a Slack summary.
- [Gmail Cleaner](https://www.betterclaw.io/workflows/gmail-cleaner): Every 2 hours, AI labels cold email, newsletters, old one-time codes and delivery updates in Gmail so you clear them in one go. Nothing is ever deleted.
- [Failed Payment Recovery](https://www.betterclaw.io/workflows/failed-payment-recovery): Every morning, finds failed Stripe payments, emails each customer a friendly reminder with a secure link to pay and follows up until the invoice is paid.
- [Job Application Tracker](https://www.betterclaw.io/workflows/job-application-tracker): Every morning, finds new jobs that fit you, drafts a tailored résumé and cover letter for the best ones and tracks your applications from your email.
- [Weekly Team Report](https://www.betterclaw.io/workflows/weekly-team-report): Every Friday afternoon, a short report from Linear on what shipped, what's in progress, what slipped and what's stuck, against the last four weeks.
- [Applicant Screener](https://www.betterclaw.io/workflows/applicant-screener): Every weekday, reads new applications in your email, scores each CV against the role's must-haves, logs everyone in a Google Sheet and sends a shortlist.
- [Industry Signal Digest](https://www.betterclaw.io/workflows/industry-signal-digest): Every weekday morning, one email with what happened in your market: launches, funding, hires, regulation and discussions, each with why it matters to you.

## Guides
- [Why AI agents fail on multi-step tasks](https://www.betterclaw.io/blog/why-ai-agents-fail): real failure rates (Agents' Last Exam: best system 24.0%, hardest tier under 1%), the compounding-error arithmetic (95% per step = 36% over 20 steps), Berkeley's MAST taxonomy of 14 failure modes (44.2% design, 32.3% misalignment, 23.5% verification), and the setup that works: fixed steps, the agent only at judgment points, a machine-checkable done-check and a separate checker. Includes workflow vs agent.
- [Running AI agents overnight](https://www.betterclaw.io/blog/run-ai-agents-overnight): which tasks Claude Code, Codex or Hermes Agent can safely run unattended (narrow, checkable in two minutes) and which never should (deploys, payments, auth, migrations). A dry-run-tested Bash script that runs Codex (`codex exec --sandbox workspace-write`) per task in its own git worktree and opens draft PRs, the `claude -p` swap, Hermes cron, and a 40-minute morning review.
- [Hermes vs Pi vs OpenCode](https://www.betterclaw.io/blog/hermes-vs-pi-vs-opencode): what each open-source agent is for (Pi or OpenCode for coding, Hermes for automation, memory, schedules and chat apps), a side-by-side table, measured fixed overhead per turn on fresh installs (Hermes v0.22.5 ~10,400 tokens, OpenCode 2.1.4 ~5,100, Pi 1.4.0 ~1,500), install and MCP config for each, using Hermes as an orchestrator over coding tools, and cost.
- [Is Hermes Agent still free?](https://www.betterclaw.io/blog/is-hermes-agent-free): after Nous Research's $90M Series B and Hermes for Businesses (7 Oct 2026), Hermes Agent is still MIT-licensed (v0.22.2, 15 Oct), including the Desktop app. What costs nothing (the software, your own computer, local models via Ollama) and what costs money (VPS from $6.49/mo, Hermes Cloud per day, API keys, Nous Portal plans $20 to $200/mo, Tool Gateway). Which subscriptions Hermes can sign in with (ChatGPT/Codex yes; Claude Max extra usage only; Claude Pro no; SuperGrok/X Premium+ yes; Gemini API key only) and the cheapest setups.
- [Use your ChatGPT subscription with OpenClaw](https://www.betterclaw.io/blog/openclaw-chatgpt-subscription): OpenAI lets Go, Plus and Pro plans power OpenClaw (OpenClaw was a Sign in with ChatGPT launch partner at DevDay, 29 Sep 2026). Codex login vs Sign in with ChatGPT (beta) vs API key, setup commands for OpenClaw 2026.10.5 (`openclaw models auth login --provider openai`, `--method device-code`, `--method siwc`), default model GPT-6 Astra, API-key failover via `auth.order.openai`, how shared usage limits and per-app caps work, subscription vs API cost, and connecting a ChatGPT plan on BetterClaw.
- [BYOK AI explained: bring your own key for AI agents](https://www.betterclaw.io/blog/byok-ai-bring-your-own-key): what BYOK means (your provider API key, billed at list price, no resale margin), BYOK vs managed credits vs subscription sign-in, a worked cost example on Claude Sonnet 5.5 (direct $120 vs OpenRouter card credits $126.60 vs a 20% reseller margin $144), key safety (one key per agent, provider spend limits, restricted keys, rotation), and how BYOK works on BetterClaw (zero inference markup, 28+ providers, custom OpenAI/Anthropic-compatible endpoints, Ollama).
- [Claude Max API credits: how to use them with an AI agent](https://www.betterclaw.io/blog/claude-max-api-credits-ai-agents): the monthly API credits Anthropic added to Max and Team plans on 7 Oct 2026 (Max 5x $100, Max 20x $200, Team $20 per Standard / $100 per Premium seat pooled up to $500; Free, Pro and Enterprise not eligible; no rollover). Covers how to claim them by linking one Console organization, what they cover (API, Managed Agents, Agent SDK, Playground) and what they don't (interactive Claude Code, extra usage, Bedrock/Vertex/Foundry), how to use them from OpenClaw, Hermes and BetterClaw with an API key rather than a Claude login, runs per $100/$200 on Sonnet 5.5 and Haiku 5.5, and when Max beats pay-as-you-go.
- [OpenClaw latest version](https://www.betterclaw.io/blog/openclaw-latest-version): the current stable and extended-stable (LTS) releases, what changed, how to check your version (`openclaw --version`, `openclaw update status --json`), how to update safely (`openclaw update`, then `openclaw doctor --fix` and `openclaw gateway restart`), which release channel to use, and where OpenClaw 2.0 (v2026.8.1) sits.
- [Complete OpenClaw setup guide](https://www.betterclaw.io/blog/openclaw-setup-guide-complete)
- [OpenClaw memory wiki](https://www.betterclaw.io/blog/openclaw-memory-wiki-guide): how OpenClaw's memory system works in 2.0 (conversation, daily notes, Dreaming, MEMORY.md/USER.md, memory-wiki), the working memory-wiki config, all 5 wiki tools, vaultMode isolated/bridge/unsafe-local, per-agent vaults, Active Memory, `openclaw memory forget`, and what OpenClaw 2.0 changed (QMD retired, Dreaming on by default).
- [Install custom skills](https://www.betterclaw.io/guide/install-custom-skills)
- [Integrate Telegram](https://www.betterclaw.io/guide/integrate-telegram-with-betterclaw)
- [Generate a Telegram bot token](https://www.betterclaw.io/guide/generate-telegram-bot-token)
- [Slack integration](https://www.betterclaw.io/guide/slack-integration)
- [Discord integration](https://www.betterclaw.io/guide/discord-integration)
- [Connect Gmail](https://www.betterclaw.io/guide/connect-gmail)
- [Claude Cowork on Windows: requirements and setup](https://www.betterclaw.io/blog/claude-cowork-windows-setup): Claude Cowork runs on Windows 10 (build 19041+) and Windows 11, x64 and Arm64. It requires Virtual Machine Platform enabled and admin rights for initial setup. The service runs as cowork-svc.exe (CoworkVMService). Cloud sessions work on web and mobile without the desktop app. Covers the requirements table, what cowork-svc.exe is, installing without admin rights (Add-AppxProvisionedPackage), and which fix page handles each setup error.
- [Claude Cowork en Windows (Español)](https://www.betterclaw.io/blog/claude-cowork-windows-instalar): Spanish-language version — what Cowork is, price by plan (no free plan; Pro 20 US$/month), Windows requirements, step-by-step install, Windows 11 Home, the "Additional setup needed" prompt, the September 2026 Windows update bug and fix KBs, cowork-svc.exe, installing without admin rights, desktop vs cloud sessions (6 Oct 2026 change) and Cowork vs Claude Code vs ChatGPT Work.
- [Claude Cowork no Windows (Português)](https://www.betterclaw.io/blog/como-instalar-claude-cowork-windows): Brazilian Portuguese version — same coverage, with the real cost in Brazil (BRL checkout vs USD plus 3.5% IOF), Windows requirements, install steps, Windows 11 Home, the September 2026 update fix KBs, cowork-svc.exe, no-admin installs, cloud vs desktop sessions and Cowork vs Claude Code vs ChatGPT Work.
- [Claude Cowork für Windows (Deutsch)](https://www.betterclaw.io/blog/claude-cowork-windows-installieren): German version — same coverage plus giving Cowork access to a folder, the cost in Germany (EU Pro €18 plus 19% VAT shown at checkout), Windows requirements, install steps, Windows 11 Home, the September 2026 update fix KBs, cowork-svc.exe, no-admin installs, cloud vs desktop sessions and Cowork vs Claude Code vs ChatGPT Work.
- [Hermes Agent Docker install: setup, Compose and troubleshooting](https://www.betterclaw.io/blog/hermes-agent-docker-install): Hermes Agent runs in Docker using the official nousresearch/hermes-agent image (`:latest` or a dated tag; there is no `:stable`). All data persists in one volume, host ~/.hermes mounted at /opt/data. Port 9119 is the web dashboard (`HERMES_DASHBOARD=1`), port 8642 the optional OpenAI-compatible API server (`API_SERVER_ENABLED=true` + `API_SERVER_KEY`). The container runs as UID 10000 (remap with `HERMES_UID`/`HERMES_GID` or `PUID`/`PGID`). Covers the official Compose file, updating (`docker compose pull && docker compose up -d`), the Docker terminal backend (`hermes config set terminal.backend docker`), and Portainer, Synology and Unraid setup.
- [Inbox Triage agent for Gmail and Outlook](https://www.betterclaw.io/use-cases/email-triage): classifies every message (urgent, needs reply, FYI, spam), drafts replies for approval, escalates urgent mail to Slack or Telegram, and archives noise on a heartbeat schedule.
- [Lead Qualification agent for forms, email and Instagram](https://www.betterclaw.io/use-cases/lead-qualification): scores inbound leads Hot/Warm/Cold on weighted criteria you define (budget, timeline, company size, decision-maker language), drafts first replies for approval, books calls from your calendar, and routes leads to HubSpot, Salesforce or any webhook CRM. Inbound only: no phone calls, voice or cold outreach.
- [Customer Support Triage agent for Help Scout, Intercom, Zendesk and shared inboxes](https://www.betterclaw.io/use-cases/customer-support-triage): classifies every ticket (billing, technical, shipping, account, feature request, spam), assigns urgent/normal/low priority, drafts first replies from your knowledge base, and escalates low-confidence, high-value, refund, legal and angry-customer tickets to a human. Starts at Intern trust level (every reply approved); auto-sends routine replies only once promoted. About $0.01–0.02 per ticket in model cost on Sonnet 5, vs $1.50–2.00 per automated resolution on Zendesk AI agents. Shopify order and tracking lookups for ecommerce.
- [GPT-6 Astra vs Fable 5.1 vs Sonnet 5 on real agent work](https://www.betterclaw.io/blog/gpt-6-astra-vs-fable-5-1-vs-sonnet-5-agents): five agent tests run September 7, 2026 — structured tool calling, failure honesty, instruction decay, context honesty, and measured cost per task ($1.10 Astra, $0.65 Fable 5.1, $0.19 Sonnet 5) — plus the cache-rate math and the routing rules that follow from it.
- [Opus 5.5 vs GPT-6 Sol vs Grok 4.7: cheapest model for agents](https://www.betterclaw.io/blog/opus-5-5-vs-gpt-6-sol-vs-grok-4-7-agents): prices checked September 28, 2026 for Opus 5.5 ($4/$20, $0.20 cache reads), GPT-6 Sol ($2/$10, $0.20), Grok 4.7 ($2/$6, $0.50 cache reads, doubling above 200K prompt tokens) and GPT-6 Luna ($0.10/$0.50, $0.01); worked cost per agent task (40K cached, 3K input, 2K output: $0.060, $0.034, $0.038, $0.0017); why Grok's cheap output loses on cache-heavy agent work (and wins via OpenRouter at $1.60/$4.80/$0.40), what Luna is and is not fit for, and a routing table across all four.
- [Claude Opus 5.5 for agents: real cost and the rerouting problem](https://www.betterclaw.io/blog/claude-opus-5-5-agents-cost-rerouting): Opus 5.5 (released September 22, 2026) pricing vs Opus 5 and Sonnet 5 — $4/$20, cache reads cut 60% to $0.20, cache writes $5, fast mode $8/$40; cost per agent task ($0.060 vs $0.085 vs $0.034); the safety-classifier fallback that silently answers flagged cybersecurity requests with Opus 4.8 and biology/frontier-LLM requests with Opus 5; how to detect it via the response `model` field and handle it (split agents, pin per step, alert on rate); Opus 5.5 vs Sonnet 5 routing; thinking cannot be disabled.
- [Claude Haiku 5.5 for AI agents: what it really costs](https://www.betterclaw.io/blog/claude-haiku-5-5-agents): Haiku 5.5 (released October 7, 2026) pricing — $0.10/$0.50 per million up to 100K-token prompts, the whole request at $0.50/$2.50 above 100K (cache tokens count toward the threshold); the ~25-30% tokenizer increase vs Haiku 4.5; cost per agent run vs Haiku 4.5, Sonnet 5.5 and Opus 5.5; benchmarks (OSWorld 72.4%, Terminal-Bench 4.0 39.2%); Haiku 5.5 vs GPT-6 Luna's 272K price step; effort levels and disabling thinking (high effort or below); setup in OpenClaw 2026.10.1+, Claude Code and BetterClaw.
- [Migrate an LLM app to a new provider with zero downtime](https://www.betterclaw.io/blog/migrate-llm-provider-zero-downtime): the 7-step procedure (abstraction layer, shadow traffic, eval set, tool-schema diff, rate-limit map, staged cutover with fallback, one-switch rollback) and the provider differences that break migrations: OpenAI vs Anthropic tool-call formats, GPT-6.1 Sol tool calling only on the Responses API, cache pricing, long-context price steps (Haiku 5.5 100K, GPT-6.x 272K), DeepSeek's weekday peak-hour pricing; OpenRouter fallbacks and fees vs a BYOK platform.
- [OpenClaw API costs: what you actually pay per model](https://www.betterclaw.io/blog/openclaw-api-costs): per-million rates as of Oct 2026 for Claude Opus 5.5 ($4/$20), Sonnet 5.5 ($2/$10), Haiku 5.5 ($0.10/$0.50 up to 100K-token prompts, $0.50/$2.50 above), Sonnet 4.6 ($3/$15) and deepseek-flash ($0.15/$0.60 off-peak), the 136K-token system-prompt overhead at real cache-hit rates, retry loops and context overhead, and the five ways an agent bleeds tokens, plus the 8-step cost-reduction stack (model routing, session resets, a contextTokens cap, memory hygiene, a cheap heartbeat model, fallbacks, loop detection, spending caps) and a 10-minute OpenClaw cron/automations audit.
- [OpenRouter vs direct API: is it cheaper?](https://www.betterclaw.io/blog/openrouter-vs-direct-api-agents): no per-token markup and no monthly subscription, a 5.5% fee when buying credits by card ($0.80 minimum; 5% with crypto; 8% on the Business plan), BYOK free up to $25,000/month of list-price inference then 5% ($200,000 on Enterprise), worked monthly cost math vs direct API for Opus 4.8, DeepSeek V4 Pro and GPT-5.5, the DeepSeek off-peak catch (most OpenRouter hosts charge DeepSeek's peak rate all day; pin `deepseek/deepseek-v4-pro-0813` to the DeepSeek provider to keep off-peak pricing), when OpenRouter is worth it, and OpenRouter vs a $20 Claude Pro or ChatGPT Plus subscription. Re-verified September 22, 2026.
- [OpenRouter fee calculator](https://www.betterclaw.io/tools/openrouter-fee-calculator): free tool comparing monthly cost of direct API, OpenRouter credits (card or crypto) and BYOK (standard or Enterprise allowance) for a chosen model and daily token volume.
- [OpenRouter alternatives for AI agents: where the fee actually hides](https://www.betterclaw.io/blog/openrouter-alternatives-ai-agents): ten gateways (OpenRouter, nexos.ai, Vercel AI Gateway, Cloudflare AI Gateway, Portkey, LiteLLM, Helicone, Requesty, Eden AI, LLM Gateway) compared on token markup, where the fee sits, self-hosting and catalogue, verified September 2026; OpenRouter BYOK is a $25,000/month list-price dollar cap, not 1M requests; the $0.80 top-up minimum; the four gateway behaviours that break agents (dropped provider-specific parameters, silent failover, per-minute rate limits, stacked privacy policies); and which gateway to pick by constraint.
- [OpenRouter 401 "User not found": six causes](https://www.betterclaw.io/blog/openrouter-401-user-not-found): OpenRouter reports an expired key as "User not found"; curl /api/v1/auth/key to split a dead key from a stack that never sent it; misspelled OPENROUTER_API_KEY, sub-agents reading an unset environment variable via LiteLLM, stale keys from other providers, provider/model-prefix/key mismatch, 402 and 403 lookalikes, the activity-log check, and why sk-or-v2 keys don't exist.
- [Which AI providers train on your agent's data?](https://www.betterclaw.io/blog/ai-provider-training-data-policy-2026): 21 providers and endpoints checked against their own terms on September 15, 2026 — training default, opt-out, retention and a customer-data verdict for OpenAI, ChatGPT, Anthropic (including Fable 5 / 5.1 frontier retention), Gemini free vs paid, Vertex, Azure OpenAI, DeepSeek, Groq, OpenRouter, Together, Mistral, Cohere, xAI, and local Ollama, LM Studio, llama.cpp and Jan — plus the structural / contractual / policy safety tiers and a ten-minute agent audit.

## Model Comparisons
- [GLM 5.2 vs Sonnet 4.6 vs MiniMax M3](https://www.betterclaw.io/blog/glm-5-2-vs-sonnet-4-6-vs-minimax-m3): benchmarks, licences and first-party prices for the three agent-model families, plus GLM 5.3 (custom licence, not MIT), GLM-5.3-Flash (MIT), Sonnet 5 and Opus 5.5 — and why MiniMax M3's $0.30/$1.20 is the standard rate under 512K input, not a promo. Also covers per-task and daily agent cost at 100, 500 and 2,000 tasks, the routing math, and which model to pick for an OpenClaw agent.
- [GLM 5.2 vs Sonnet 4.6](https://www.betterclaw.io/blog/glm-5-2-vs-sonnet-4-6): seven agent tasks tested head-to-head, updated for GLM 5.3 and Sonnet 5, with the GLM 5.1 → 5.2 → 5.3 lineage and GLM 5.3's non-MIT licence.
- [MiMo-V2.6-Pro for AI agents: setup and cost vs GLM and MiniMax](https://www.betterclaw.io/blog/mimo-v2-6-pro-agents-setup-cost): Xiaomi's MIT-licensed open-weight flagship (released 21 Sept 2026; 1.02T total / 42B active, 1M context) against GLM-5.3, GLM-5.3 Flash and MiniMax M3 on Artificial Analysis score (46 vs 45), list price ($0.435/$0.87 per million, $0.0036 cached) and cost per finished task, monthly agent cost at 300 tasks, setup via Xiaomi's API, OpenRouter or vLLM (the `--tool-call-parser mimo` gotcha), Ollama status, speed and verbosity caveats, and Anthropic's distillation allegations. Prices checked 5 October 2026.
- [MiniMax M3 vs Claude Sonnet 4.6](https://www.betterclaw.io/blog/minimax-m3-vs-claude-sonnet-4-6): when the Anthropic premium is worth paying — agent tasks tested, tool-chain failure economics, and the M3/Opus routing blend.
- [GLM vs LLM: what the GLM model is, and why it is an LLM](https://www.betterclaw.io/blog/glm-vs-llm-difference-explained): GLM (General Language Model) is a family of large language models built by Z.ai (formerly Zhipu AI). The current generation is GLM-5.3 (August 2026). GLM models are LLMs, not a different architecture, and GLM-5.3-Flash is available with open weights under the MIT licence. Covers the three meanings of GLM (including the statistics "generalized linear model"), a GLM vs Claude/GPT/MiniMax price table, the full version history, and why GLM-5.5 is not out.

## Hardware and Local AI
- [Zanus AI Review: private AI servers vs NVIDIA DGX](https://www.betterclaw.io/blog/zanus-ai-vs-nvidia-dgx): a Zanus AI review (on-premise servers with 15+ preinstalled modules and quote-only pricing, plus separate cloud Front Office/Back Office software; not an AI agent builder) and turnkey appliance vs component compute — what each vendor publishes, why the third-party Zanus price tables are unreliable, and the staffing question that decides the purchase.
- [DGX Station price (Oct 2026) and 6 alternatives](https://www.betterclaw.io/blog/dgx-station-alternative): no official NVIDIA price; US listings from $92,887.47 (Supermicro) to $159,283 (Dell Pro Precision with GB300) for the 748 GB GB300 workstation (checked 13 Oct 2026; every listing is the Linux version, DGX Station for Windows is "Coming in Q4" with no price), the monthly cost (about $2,570 over 36 months), plus six cheaper alternatives compared on memory, bandwidth and price.
- [NVIDIA DGX alternatives: every DGX compared](https://www.betterclaw.io/blog/nvidia-dgx-alternative): the full DGX line — Spark, Station, SuperPOD, BasePOD, DGX Cloud — what each tier is for, why NVIDIA publishes no list prices, and the cheaper alternative at every budget band.
- [DGX Spark alternatives](https://www.betterclaw.io/blog/dgx-spark-alternative): cheaper paths to local inference than a $6,950 Spark (raised from $4,699 on 2 October 2026; a 64 GB model is $4,999 from 23 October), including when cloud APIs remove the hardware question entirely.
- [RTX Spark vs DGX Spark for local AI agents](https://www.betterclaw.io/blog/rtx-spark-vs-dgx-spark): NVIDIA's Windows-on-Arm RTX Spark laptops (laptops from 16 Oct 2026, up to 128GB unified memory, no clustering) against the Linux DGX Spark (273 GB/s, ConnectX-7 pairing, 128GB $6,950 / 64GB $4,999) on specs, power, what models fit in 64GB vs 128GB, Windows on Arm vs DGX OS software readiness (CUDA 13.4, Ollama, MXC containers), price per GB, Mac Studio / M5 Max / Strix Halo / RTX 5090 comparisons, and running BetterClaw Desktop against a Spark over the local network. Prices checked 13 October 2026.
- [DGX Spark Alternativen 2026 (Deutsch)](https://www.betterclaw.io/blog/dgx-spark-alternativen): German-language version for Germany, Austria and Switzerland — DGX Spark at $6,950 (64 GB $4,999 from 23 Oct), retailer prices checked 6 Oct 2026 (Germany from 5.799 €, Austria about 6.499 €, Switzerland CHF 5.785 at Alternate.ch), the six cheaper alternatives, a DGX Spark vs RTX 5090 / Mac Studio / Strix Halo / GX10 / ZGX Nano / RTX Pro 6000 table, and buying-used advice.
- [Alternativas a DGX Spark 2026 (Español)](https://www.betterclaw.io/blog/alternativas-dgx-spark): Spanish-language version for Spain and Latin America — DGX Spark at $6,950 (64 GB $4,999 from 23 Oct), local prices checked 6 Oct 2026 (Spain 6.499 € with VAT at IETRES, Mexico 199.999 MXN at Core Gaming), what to expect elsewhere in Latin America, the six cheaper alternatives and a DGX Spark vs RTX 5090 / Mac Studio / Strix Halo comparison table.
- [Alternativas ao DGX Spark 2026 (Português)](https://www.betterclaw.io/blog/alternativas-dgx-spark-preco): Brazilian Portuguese version — DGX Spark at $6,950 (64 GB $4,999 from 23 Oct), Brazil prices checked 6 Oct 2026 (KaBuM partner listings from R$ 53.649, shown without warranty), why buying in Paraguay rarely pays (US$ 500 land exemption, 50% tax on the excess), the six cheaper alternatives and a DGX Spark vs RTX 5090 / Mac Studio / Strix Halo comparison table.
- [OpenClaw gratis: los mejores modelos gratis (Español)](https://www.betterclaw.io/blog/openclaw-modelos-gratis): Spanish-language version of the OpenClaw free models list — OpenClaw is free (open source); the model and server may cost. Gemini 3.8 Flash free tier, Qwen3.8-27B on Groq's free plan, local Ollama (qwen3.5:9b on 16 GB, qwen3.8:27b on 24 GB+), DeepSeek deepseek-flash as the near-free step up, OpenRouter :free limits, changing the model in OpenClaw, Gemini setup (GEMINI_API_KEY or GOOGLE_API_KEY), offline use, and privacy per provider (Gemini free tier uses prompts in Latin America but not in Spain/EEA).
- [OpenClaw grátis: os melhores modelos gratuitos (Português)](https://www.betterclaw.io/blog/openclaw-modelos-gratuitos): Brazilian Portuguese version of the OpenClaw free models list — OpenClaw itself is free (open source); the model and server may cost. Gemini 3.8 Flash free tier first, Qwen3.8-27B on Groq's free plan second (30 RPM, 1,000 RPD, 8K TPM, 200K TPD), local Ollama (qwen3.5:9b on 16 GB, qwen3.8:27b on 24 GB+), DeepSeek deepseek-flash as the near-free step up (peak hours in Brasília time), OpenRouter :free limits, how to change the model (agents.defaults.model.primary/fallbacks, openclaw models set), and what each free provider does with data (Gemini free tier trains on prompts outside the EEA/CH/UK; LGPD note).
- [OpenClaw kostenlos: die besten Gratis-Modelle (Deutsch)](https://www.betterclaw.io/blog/openclaw-kostenlose-modelle): German version — same model ranking and limits, DeepSeek peak hours in German time, GDPR/DSGVO section (Gemini API paid-tier data terms apply to the free tier in the EEA, Switzerland and UK; DeepSeek has no EU adequacy decision), using several models at once (fallbacks, subagents, agents.entries, /model), and free hosting (Oracle Always Free A1 halved to 2 OCPUs / 12 GB on 15 June 2026).
- [Ornith-1.5-35B-A3B for AI agents](https://www.betterclaw.io/blog/ornith-1-5-35b-agent-setup-guide): the 35B-total, 3B-active MoE coding-agent model (79.0 SWE-bench Verified per its card) — GGUF file sizes and hardware by quant (Q4_K_M 21.7 GB on a 24GB card), the Ollama pull and tool-call template check, exact OpenClaw and Hermes configs, the Ornith 9B vs Gemma 4 12B vs Qwen 3.5 9B benchmark table with its vendor-run caveat, and where a local coding model fits in a routing stack.
- [Qwen 3.8 on Ollama (Qwen 3.7 is API-only)](https://www.betterclaw.io/blog/qwen-3-7-ollama-honest-review): no qwen3.7 exists on Ollama or Hugging Face; Qwen3.8-27B setup — official tag sizes (18GB Q4 to 56GB BF16), Ollama's VRAM-based default context and num_ctx, model-card sampling settings, turning thinking off, known Ollama agent issues, and Model Studio pricing for qwen3.7-max and qwen3.7-plus.
- [LM Studio vs Ollama (and Jan)](https://www.betterclaw.io/blog/ollama-vs-lm-studio): the 2026 comparison, re-checked 13 Oct (LM Studio 0.4.26, Ollama 0.40.3; Jan 0.8.4 as of 28 Sep) — llmster headless daemon (0.4.0, 28 Jan 2026; systemd, continuous batching, MCP over REST), Ollama's GGML engine and MLX on Apple Silicon (the default for supported models since Ollama 0.40.0, 6 Oct 2026), an agent-backend section (OLLAMA_NUM_PARALLEL default 1 and RAM scaling, keep_alive 5 min vs JIT TTL 60 min, VRAM-based Ollama context default, MCP client vs orchestrator, CPU-only x86 llmster-preview Docker image), LM Studio Bionic as a separate app, the "Friends Don't Let Friends Use Ollama" critique, and Jan's AGPLv3/opt-in telemetry.
- [Gemma 4 vs Qwen 3.5, 3.6 and 3.8](https://www.betterclaw.io/blog/gemma-4-vs-qwen-3-5): every size compared (Gemma 4 ships E2B, E4B, 12B, 26B A4B and 31B — there is no Gemma 4 27B), which pairings are fair, VRAM by tier, and where Qwen 3.6 27B and Qwen 3.8 27B fit.
- [Gemma 4 31B vs Qwen 3.8 27B](https://www.betterclaw.io/blog/gemma-4-27b-vs-qwen-3-6-27b): the dense 24GB matchup for local agents. Tool calling, coding, decode speed and VRAM headroom at 64K, plus why there is no Gemma 4 27B and no Qwen 4 yet.
- [Gemma 4 vs Qwen 3.8](https://www.betterclaw.io/blog/gemma-4-vs-qwen-3-8): the family-level pick for local agents. Shared benchmarks (Qwen leads all three), tool calling and preserved thinking, Ollama download sizes, the Flash Next licence limits, and which model to run for which agent job.

## Competitors and Positioning
- [Hermes Cloud vs BetterClaw](https://www.betterclaw.io/blog/hermes-cloud-vs-betterclaw): Hermes Cloud pricing (a per-day instance fee, metered per second, with stopped instances paying storage only at $0.03/day), plus Nous Portal plans for inference and tools (Plus $20, Super $100, Ultra $200; 10% credit bonus; capped rollover). Works out what a real month costs and compares it with BetterClaw's flat plans with BYOK and zero markup and with a self-hosted VPS. Also covers Hermes for Businesses.
- vs KiloClaw ($9/mo): BetterClaw has a free plan, skill security audit, and multi-agent
- vs Ampere (free beta): BetterClaw is production-grade with security auditing
- vs Hostinger ($6.99/mo): BetterClaw is fully managed, Hostinger is a VPS with 1-click install
- vs xCloud ($24/mo): BetterClaw has a free plan and skill security audit
- vs self-hosting: BetterClaw removes 3 to 6 hours per month of maintenance
- [Free AI agent platforms compared](https://www.betterclaw.io/blog/free-ai-agent-platforms-2026): the three meanings of "free" (free software, free trial, free plan) across BetterClaw, OpenClaw, Hermes, CrewAI, n8n, Gumloop, MindStudio, and Lindy, with the real all-in monthly cost of each.
- [Hermes Agent alternatives](https://www.betterclaw.io/blog/hermes-agent-alternative): Option 0, Hermes Cloud (Nous-hosted Hermes, from $0.51/day running, $0.03/day stopped, inference extra), then six options ranked by setup time, with hosting and cost — BetterClaw, managed OpenClaw hosting, OpenClaw, n8n, CrewAI, LangGraph — plus the cases where staying on Hermes is the right call.
- [Muse vs Instinct (2026), plus BetterClaw](https://www.betterclaw.io/blog/betterclaw-vs-muse-vs-instinct-2026): a Muse vs Instinct head-to-head (Meta's Muse, launched September 8, 2026, US and Canada only, $0/$20/$100, vs invite-only Instinct reached by text and phone), then both consumer personal agents against BetterClaw (multiple work agents), compared on pricing re-checked October 5, 2026 including the model bill, Instinct's $10B Series C, who owns the environment, Amazon's September 20 block on Muse shopping (why Amazon did it and what it means for browsing agents), locked vs BYOK models, and what happens when an update breaks — including where BetterClaw loses. Also covers Klaus (managed OpenClaw hosting), which shut down September 1, 2026, and what its former customers can do.
- [MCP vs direct API calls: does your agent need MCP?](https://www.betterclaw.io/blog/mcp-vs-direct-api-tools-agents): where MCP's cost comes from, using Anthropic's measurements (GitHub MCP server ~26K tokens, five servers / 58 tools ~55K tokens sent on every model call unless the client defers loading), the case against MCP (models now call APIs and run code directly), why MCP isn't dead (Tool Search Tool cuts ~77K to ~8.7K tokens, code mode 150K to 2K, cross-client auth, Agentic AI Foundation governance), when to use MCP vs a direct API call, and how BetterClaw scopes connectors per agent.
- [AWS Bedrock AgentCore pricing, every line item](https://www.betterclaw.io/blog/aws-bedrock-agentcore-pricing-alternatives): all twelve AgentCore meters re-verified against AWS's pricing page on October 13, 2026 — Runtime v1 at $0.0895/vCPU-hour and $0.00945/GB-hour, Runtime v2 at $0.1276 and $0.0169 (committed baseline $0.0997/$0.0132 launching by October 2026), instances at EC2 On-Demand plus a 12% fee, short-term memory billed per GB since October 6, 2026 ($1.00/GB ingested, $0.20/GB retrieved, $0.10/GB-month), Web Search $7.00/1K, Gateway $0.005/1K API vs $0.025/1K Search, long-term Memory's $0.75 vs $0.25 per 1K records, Policy, Evaluations, Agent Registry, Payments and the unmetered Optimization feature — plus AWS's own worked examples, the token amplification that turns a $45 model bill into $128-$257, what one 10,000-query agent costs a month (about $195-$325), an AgentCore vs BetterClaw table (setup, GovCloud, agent payments, 10,000-conversation cost), and five alternatives.
- [Vertex AI pricing 2026](https://www.betterclaw.io/blog/google-vertex-ai-agent-builder): what a Vertex AI (Gemini Enterprise Agent Platform) agent actually costs, Gemini, Model Garden and agent runtime rates re-verified October 13, 2026 — what the platform is (Agent Studio, ADK, Agent Runtime, Model Garden; renamed from Vertex AI Agent Builder on April 23, 2026), Gemini 3.8 Flash at $0.75/$3.75 per million tokens through December 31, 2026 (then $1.50/$7.50, paid as 50% credits back), Flex/Batch, Priority and regional multipliers, Claude Opus 5.5 at $4/$20, Sonnet 5.5 at $2/$10 and Haiku 5.5 at $0.10/$0.50 in Model Garden, Agent Compute $0.085/vCPU-hour, Agent Memory $0.009/GiB-hour, Agent Storage $0.30/GiB-month with per-operation charges (Sessions and Memory Bank billing began September 1, 2026), 1- and 3-year savings-plan rates for all three, Agent Search at $1.50/$4.00 per 1,000 plus $4.00 for Advanced Generative Answers (10,000 free queries a month), Model Garden per-token prices (Claude, Llama, DeepSeek, gpt-oss), three worked monthly bills ($91 to $13,870), the standing monthly free allowance (50 vCPU-hours, 100 GiB-hours, 1 GiB-month), a quick-reference table for training, prediction, Vector Search, RAG Engine and CX Agent Studio ($0.50/session), and Copilot Studio and AWS AgentCore comparisons.
- [Vertex AI Agent Builder alternatives](https://www.betterclaw.io/blog/vertex-ai-agent-builder-alternatives): eight cheaper or simpler options with a 10,000-query cost comparison, a BetterClaw vs Vertex AI Agent Builder side-by-side table (setup time, code, hosting, free plan, pricing model, model providers, integrations, lock-in), and why the old "$0.25 per 1,000 session events" SKU never existed.
- [OpenAI Dots vs Meta Muse vs Grok Bot (2026)](https://www.betterclaw.io/blog/openai-dots-vs-meta-muse-vs-grok-bot): the three always-on agents launched in seven weeks (Grok Bot 11 Aug, Muse 8 Sept, Dots 29 Sept 2026) compared on price and access (Dots from ChatGPT Pro $100/mo, excluded on Pro in the EEA, UK and Switzerland; Muse free plus Power $20 and Max $100, US and Canada only; Grok Bot bundled with paid Cursor plans from $20 or SuperGrok from $30), model (GPT-6 Astra, Muse Spark, unnamed with no picker or BYOK), where each runs, privacy terms, and Amazon's September block on Muse shopping, plus short head-to-head answers for Dots vs Muse, Grok Bot vs Muse and Dots vs Grok Bot. Facts re-checked October 5, 2026.
- [BetterClaw vs Grok Bot](https://www.betterclaw.io/compare/grok-bot): xAI's Grok Bot (bundled with SuperGrok $30/mo, Plus $100, Heavy $300, and every paid individual Cursor plan; model chosen for you with no model picker or BYOK; unpublished weekly limits; works apps on its own cloud computer) vs BetterClaw (Basic $19, Pro $49, Business $149; 28+ providers via BYOK at zero markup), with model-cost tables for light/medium/heavy use at 29 September 2026 prices and where each one is the better choice.
- [BetterClaw vs Meta Muse](https://www.betterclaw.io/compare/muse): Meta's personal AI agent (launched 8 Sept 2026; US since launch, Canada since 18 Sept; Free, Power $20/mo, Max $100/mo, card required even on free; Muse Spark model only, no BYOK; iOS, Android, muse.ai, WhatsApp and a Mac app) vs BetterClaw (Basic $19, Pro $49, Business $149; 28+ providers via BYOK at zero markup; Slack, Telegram, Discord, Gmail), covering what Muse can and cannot do (Amazon's September block, CNN's hands-on), Meta's privacy terms (Secure VM, Sentinel, training opt-out), and Muse for Small Business.
- [BetterClaw vs OpenAI Dots](https://www.betterclaw.io/compare/openai-dots): OpenAI's always-on agents in ChatGPT (launched 29 Sept 2026; one Dot included with ChatGPT Pro $100/$200/$500 or Business Premium $100/user/mo annual, $125 monthly; not on Plus; Pro excluded in the UK, EEA and Switzerland; GPT-6 Astra only, even after ChatGPT moved paid plans to GPT-6 Sol on 7 Oct; no API; since 9 Oct a Dot can be created in the iOS and Android apps and can start Codex tasks; texting is a limited beta for US Pro users) vs BetterClaw (from $19/mo, 28+ providers including OpenAI via BYOK or ChatGPT sign-in, any remote MCP server, local models through BetterClaw Desktop, works in the UK and EU), with who can use Dots, safety controls on both sides, limits and when each one fits. Facts re-checked October 15, 2026.
- [OpenClaw hosting: 10 best providers from $2/mo compared](https://www.betterclaw.io/openclaw-hosting): xCloud, Hostinger, ClawHosters, Elestio, ClawTrust, DigitalOcean, AWS Lightsail, IONOS, HostGator and Railway ranked on intro and renewal price, what is actually managed, server location (EU/GDPR options) and gateway security, with prices checked on 9 October 2026, a head-to-head table of managed OpenClaw (xCloud, ClawHosters, Hostinger Managed) vs an OpenClaw VPS vs a managed agent platform, Hostinger Managed OpenClaw vs Hostinger VPS, OpenClaw VPS requirements, model cost per task, and BetterClaw listed as an alternative that is not hosted OpenClaw.

## Contact
- Website: https://www.betterclaw.io
- App: https://app.betterclaw.io
- Community: https://reddit.com/r/better_claw