[{"data":1,"prerenderedAt":1246},["ShallowReactive",2],{"blog-post-hermes-vs-pi":3,"related-posts-hermes-vs-pi":1219,"decoder-links-hermes-vs-pi":1245},{"id":4,"title":5,"author":6,"body":11,"category":1193,"date":1194,"description":1195,"extension":1196,"featured":1197,"hideToc":1197,"image":1198,"imageAlt":1199,"imageHeight":1200,"imageWidth":1201,"itemList":1202,"lang":1202,"meta":1203,"navigation":1204,"noindex":1197,"path":1205,"readingTime":1206,"redirected":1197,"relatedSlugs":1202,"seo":1207,"seoTitle":1208,"stem":1209,"subtitle":1202,"tags":1210,"updatedDate":1194,"__hash__":1218},"blog/blog/hermes-vs-pi.md","Hermes vs Pi: A Durable Agent or a Minimal Coding Agent (and Where Pi Durable Fits)",{"name":7,"role":8,"avatar":9,"bio":10},"Shabnam Katoch","Growth Head","/img/avatars/shabnam-profile.jpeg","Shabnam leads growth at BetterClaw and has spent the past three years building organic distribution for AI-native products. She writes about no-code agent builders, content strategy, and what it actually takes to grow without an ads budget.",{"type":12,"value":13,"toc":1164},"minimark",[14,27,79,85,88,96,101,104,109,112,115,122,245,249,256,265,271,275,278,281,287,290,309,312,318,321,325,335,342,349,357,363,381,384,389,393,396,399,459,470,482,485,547,558,572,575,579,582,594,597,600,603,607,610,617,628,631,689,693,696,699,702,709,713,716,719,722,804,808,814,817,824,828,831,881,891,938,945,953,957,960,974,977,983,988,991,995,1002,1009,1016,1023,1026,1030,1035,1038,1042,1050,1054,1057,1061,1068,1074,1078,1081,1084,1087,1091,1094,1101,1105,1109,1112,1116,1119,1123,1126,1130,1136,1140,1143,1147,1150,1154,1160],[15,16,18],"callout",{"type":17},"quick-fix",[19,20,21,22,26],"p",{},"In the ",[23,24,25],"strong",{},"Hermes vs Pi"," choice, Hermes Agent is a durable personal agent: it remembers you, runs cron jobs and answers on more than 20 chat platforms. Pi is a minimal terminal coding agent with four tools that you extend in TypeScript. Pick Hermes for an always-on assistant, and Pi for coding work you watch.",[28,29,30,43],"table",{},[31,32,33],"thead",{},[34,35,36,40],"tr",{},[37,38,39],"th",{},"If you want...",[37,41,42],{},"Pick",[44,45,46,55,63,71],"tbody",{},[34,47,48,52],{},[49,50,51],"td",{},"An assistant that remembers you and works on a schedule",[49,53,54],{},"Hermes Agent",[34,56,57,60],{},[49,58,59],{},"A lean coding agent in your terminal, with nothing hidden",[49,61,62],{},"Pi",[34,64,65,68],{},[49,66,67],{},"To build your own long-running agent in TypeScript",[49,69,70],{},"Pi Durable (experimental)",[34,72,73,76],{},[49,74,75],{},"Both jobs covered",[49,77,78],{},"Hermes for your day, Pi for your repo",[19,80,81],{},[82,83,84],"em",{},"Checked 16 November 2026.",[19,86,87],{},"Picture this: it's Friday night, Pi is halfway through a refactor, and your laptop lid closes.",[19,89,90,91,95],{},"On Monday you run ",[92,93,94],"code",{},"pi --continue",", read where it stopped and type \"carry on\". The session was saved. The run wasn't.",[19,97,98],{},[82,99,100],{},"So is Pi broken?",[19,102,103],{},"No. Earendil, the company that now builds Pi, describes exactly that as what the coding agent is for. The real question is whether your job needs an agent that keeps going while you're away.",[105,106,108],"h2",{"id":107},"hermes-vs-pi-at-a-glance","Hermes vs Pi at a glance",[19,110,111],{},"We checked the release tags, changelogs and docs of both projects on 16 November 2026.",[19,113,114],{},"Hermes Agent v0.23.1 is the latest stable release, dated 9 November. It's a patch that rolls up about 618 merged pull requests since v0.23.0. Pi's latest release is 1.5.2 from 10 November, and the experimental Pi Durable package ships at the same version number.",[19,116,117,118,121],{},"One naming trap first. The Pi we compare is the coding agent Mario Zechner started; he joined Earendil in April 2026 and moved the repository from badlogic/pi-mono to earendil-works/pi. Both addresses serve the same code today, and the npm package is now ",[92,119,120],{},"@earendil-works/pi-coding-agent",".",[28,123,124,134],{},[31,125,126],{},[34,127,128,130,132],{},[37,129],{},[37,131,54],{},[37,133,62],{},[44,135,136,147,158,168,179,190,201,212,223,234],{},[34,137,138,141,144],{},[49,139,140],{},"What it is",[49,142,143],{},"A personal agent with memory, schedules and chat apps",[49,145,146],{},"A minimal terminal coding agent you extend",[34,148,149,152,155],{},[49,150,151],{},"Made by",[49,153,154],{},"Nous Research",[49,156,157],{},"Mario Zechner, now developed by Earendil",[34,159,160,163,166],{},[49,161,162],{},"Licence",[49,164,165],{},"MIT",[49,167,165],{},[34,169,170,173,176],{},[49,171,172],{},"Latest stable (16 Nov 2026)",[49,174,175],{},"v0.23.1, released 9 Nov",[49,177,178],{},"1.5.2, released 10 Nov",[34,180,181,184,187],{},[49,182,183],{},"Tools on a fresh install",[49,185,186],{},"20 in our test",[49,188,189],{},"4: read, bash, edit, write",[34,191,192,195,198],{},[49,193,194],{},"Memory of you",[49,196,197],{},"MEMORY.md and USER.md, plus search over past sessions",[49,199,200],{},"None built in; AGENTS.md files and saved sessions",[34,202,203,206,209],{},[49,204,205],{},"Scheduling",[49,207,208],{},"Built-in cron, run by the gateway",[49,210,211],{},"None; use your system's cron",[34,213,214,217,220],{},[49,215,216],{},"Chat apps",[49,218,219],{},"More than 20 platforms from one gateway",[49,221,222],{},"None built in; a separate extension covers Discord and Telegram",[34,224,225,228,231],{},[49,226,227],{},"Safety default",[49,229,230],{},"Command approval in smart mode; chat access denied until you allow it",[49,232,233],{},"No approval prompts; run it in a container or VM",[34,235,236,239,242],{},[49,237,238],{},"Best for",[49,240,241],{},"An assistant that runs your day",[49,243,244],{},"Coding work you watch, and building your own agent",[19,246,247],{},[82,248,84],{},[19,250,251],{},[252,253],"img",{"alt":254,"src":255},"Hermes vs Pi at a glance: type, licence, latest version, tools, memory, scheduling, chat apps, safety default and best use","/img/blog/hermes-vs-pi-at-a-glance.jpg",[19,257,258,259,264],{},"If OpenCode is also on your shortlist, our ",[260,261,263],"a",{"href":262},"/blog/hermes-vs-pi-vs-opencode","three-way comparison of Hermes, Pi and OpenCode"," puts all three side by side. This page sticks to the two-way question: an agent that runs your day, or one that runs in your terminal.",[266,267,268],"blockquote",{},[19,269,270],{},"Hermes ships the whole assistant and asks you to trim it. Pi ships almost nothing and asks you to build on it.",[105,272,274],{"id":273},"pi-durable-vs-hermes-three-meanings-of-durable","Pi Durable vs Hermes: three meanings of \"durable\"",[19,276,277],{},"Here's what changed. On 1 October 2026, Earendil shipped Pi 1.0 together with an experimental package called Pi Durable, so \"Pi\" and \"durable\" now belong in the same sentence.",[19,279,280],{},"Earendil's launch post is clear about the split. Pi the coding agent runs in a terminal, driven by one person, and if the process dies you look at what happened and tell it to continue.",[19,282,283,284],{},"Pi Durable \"does not replace the Pi coding agent\". It's a framework for building long-running agents of your own, so the honest Pi Durable vs Hermes answer is this: ",[23,285,286],{},"Hermes is a durable agent you install. Pi Durable is a durable agent you write.",[19,288,289],{},"What Pi Durable gives a developer, according to its README and launch post:",[291,292,293,297,300,303,306],"ul",{},[294,295,296],"li",{},"Every step of a run is a task that stores a checkpoint, so a new process picks up an interrupted run where it stopped.",[294,298,299],{},"Tasks can sleep until a set time, and the sleep survives restarts. Earendil's own example is a reminder that fires tomorrow.",[294,301,302],{},"One process runs many conversations at once, and any conversation can fork another.",[294,304,305],{},"In-memory, SQLite and JSONL storage, plus an adapter for Cloudflare Durable Objects.",[294,307,308],{},"About 15,000 lines of source without tests, small enough for an agent to read.",[19,310,311],{},"What it leaves out matters just as much. There's no chat gateway, no ready-made memory of you, no scheduler command and no built-in subagents, and its README opens with a warning that the API changes without notice between releases.",[19,313,314],{},[252,315],{"alt":316,"src":317},"What survives a restart in Hermes Agent, the Pi coding agent and Pi Durable","/img/blog/hermes-vs-pi-durable.jpg",[19,319,320],{},"Hermes takes the opposite route: durability comes in the box. A chat on a messaging platform is one continuous session that survives restarts and reboots, and past sessions sit in a SQLite database the agent can search. If the gateway comes back inside a job's catch-up window (up to two hours by default), a missed cron run fires late instead of vanishing.",[105,322,324],{"id":323},"hermes-vs-pi-on-memory-what-each-one-remembers","Hermes vs Pi on memory: what each one remembers",[19,326,327,330,331,334],{},[23,328,329],{},"Hermes keeps a small, curated memory of you."," Two files in ",[92,332,333],{},"~/.hermes/memories/"," hold it: MEMORY.md for what the agent has learned (capped at 2,200 characters) and USER.md for your preferences (capped at 1,375). Hermes's docs put the pair at about 1,300 tokens when full.",[19,336,337,338,341],{},"Both load into the system prompt as a frozen snapshot at the start of each session. The agent edits them with its memory tool, entries are scanned for prompt injection before they're accepted, and ",[92,339,340],{},"memory.write_approval: true"," makes every save wait for your yes.",[19,343,344,345,348],{},"Older conversations are covered too. Every CLI and chat session is stored in ",[92,346,347],{},"~/.hermes/state.db"," with full-text search, so the agent can find something you discussed weeks ago.",[19,350,351,352,356],{},"Two external memory providers, RetainDB and ByteRover, ship with Hermes, and Honcho, Mem0, Supermemory and others install from its plugin catalog. Our guide to ",[260,353,355],{"href":354},"/blog/hermes-agent-memory-providers","Hermes Agent memory providers"," compares them.",[19,358,359,360,121],{},"One trap catches most people. On Telegram or Slack a chat is one long session, so new memories stay invisible to it until you start a fresh one with ",[92,361,362],{},"/new",[19,364,365,368,369,372,373,376,377,380],{},[23,366,367],{},"Pi doesn't remember you, by design."," It reads AGENTS.md (and CLAUDE.md) from ",[92,370,371],{},"~/.pi/agent/"," and your project folders, and it saves every session as a JSONL tree in ",[92,374,375],{},"~/.pi/agent/sessions/"," that you can resume, branch with ",[92,378,379],{},"/tree"," or fork. When the context fills, it compacts older messages into a summary and keeps the originals on disk.",[19,382,383],{},"Anything more is yours to add. Pi's site says extensions can inject context before each turn, implement RAG or build long-term memory, and community packages already do.",[266,385,386],{},[19,387,388],{},"Hermes remembers you across weeks. Pi remembers the session, and only when you ask it to pick that session back up.",[105,390,392],{"id":391},"scheduling-a-built-in-cron-or-your-own-crontab","Scheduling: a built-in cron or your own crontab",[19,394,395],{},"Hermes has a real scheduler. You describe the job in plain language, in chat or from the shell, and the gateway checks for due jobs every 60 seconds and delivers the result to Telegram, Slack, email or wherever you point it.",[19,397,398],{},"We created this job on a fresh v0.23.1 install on 16 November:",[400,401,406],"pre",{"className":402,"code":403,"language":404,"meta":405,"style":405},"language-bash shiki shiki-themes github-light","hermes cron create \"weekdays at 8am\" \"Summarise my unread email in five bullets\" \\\n  --deliver telegram --name inbox-brief\nhermes cron list\n","bash","",[92,407,408,434,449],{"__ignoreMap":405},[409,410,413,417,421,424,427,430],"span",{"class":411,"line":412},"line",1,[409,414,416],{"class":415},"s7eDp","hermes",[409,418,420],{"class":419},"sYBdl"," cron",[409,422,423],{"class":419}," create",[409,425,426],{"class":419}," \"weekdays at 8am\"",[409,428,429],{"class":419}," \"Summarise my unread email in five bullets\"",[409,431,433],{"class":432},"sYu0t"," \\\n",[409,435,437,440,443,446],{"class":411,"line":436},2,[409,438,439],{"class":432},"  --deliver",[409,441,442],{"class":419}," telegram",[409,444,445],{"class":432}," --name",[409,447,448],{"class":419}," inbox-brief\n",[409,450,452,454,456],{"class":411,"line":451},3,[409,453,416],{"class":415},[409,455,420],{"class":419},[409,457,458],{"class":419}," list\n",[19,460,461,462,465,466,469],{},"It saved the job and showed the next run straight away. Then it warned that no gateway was serving the profile, which is the catch: ",[23,463,464],{},"nothing fires until the gateway is running",", so run ",[92,467,468],{},"hermes gateway install"," on any machine you expect to keep working.",[19,471,472,473,476,477,481],{},"Two safety defaults matter here. Dangerous commands are denied inside cron runs unless you change ",[92,474,475],{},"approvals.cron_mode",", and a cron run can't create more cron jobs. If your jobs still don't fire, our guide to ",[260,478,480],{"href":479},"/blog/hermes-cron-not-working","Hermes cron jobs that don't run"," covers the usual causes.",[19,483,484],{},"Pi has no scheduler, and its site's answer to background work is \"use tmux\". What it does have is print mode, which runs one prompt and exits, so your operating system's cron can drive it. We tested this read-only version on Pi 1.5.2:",[400,486,488],{"className":402,"code":487,"language":404,"meta":405,"style":405},"# crontab -e   (weekdays at 8:00; use the full path from `which pi`)\n0 8 * * 1-5 cd /home/you/project && /path/to/pi --tools read,grep,find,ls --no-session -p \"List every TODO in this repo with its file\" >> /home/you/todo-report.txt\n",[92,489,490,496],{"__ignoreMap":405},[409,491,492],{"class":411,"line":412},[409,493,495],{"class":494},"sAwPA","# crontab -e   (weekdays at 8:00; use the full path from `which pi`)\n",[409,497,498,501,504,507,509,512,515,518,522,525,528,531,534,537,540,544],{"class":411,"line":436},[409,499,500],{"class":415},"0",[409,502,503],{"class":432}," 8",[409,505,506],{"class":432}," *",[409,508,506],{"class":432},[409,510,511],{"class":419}," 1-5",[409,513,514],{"class":419}," cd",[409,516,517],{"class":419}," /home/you/project",[409,519,521],{"class":520},"sgsFI"," && ",[409,523,524],{"class":415},"/path/to/pi",[409,526,527],{"class":432}," --tools",[409,529,530],{"class":419}," read,grep,find,ls",[409,532,533],{"class":432}," --no-session",[409,535,536],{"class":432}," -p",[409,538,539],{"class":419}," \"List every TODO in this repo with its file\"",[409,541,543],{"class":542},"sD7c4"," >>",[409,545,546],{"class":419}," /home/you/todo-report.txt\n",[19,548,549,550,553,554,557],{},"With ",[92,551,552],{},"--tools read,grep,find,ls",", Pi offered the model only those four tools, so the run can read your code but can't edit it or run shell commands. MCP tools stay available unless you filter them too, and ",[92,555,556],{},"--no-session"," keeps scheduled runs out of your session list.",[19,559,560,561,564,565,568,569,121],{},"One more detail for unattended runs. Print mode can't show Pi's project-trust prompt, so a folder's own ",[92,562,563],{},".pi"," settings and extensions load only with a saved trust decision, ",[92,566,567],{},"--approve"," or ",[92,570,571],{},"defaultProjectTrust: \"always\"",[19,573,574],{},"The result lands in a file, not a chat app. That's the real difference.",[105,576,578],{"id":577},"chat-apps-one-gateway-or-a-bridge-you-run","Chat apps: one gateway or a bridge you run",[19,580,581],{},"Hermes runs a single background process, the gateway, that connects to every platform you configure. Nous lists more than 20, including Telegram, Discord, Slack, WhatsApp, Signal, Microsoft Teams, Matrix, LINE, SMS, email and iMessage through BlueBubbles.",[19,583,584,585,588,589,593],{},"Access is closed by default. The gateway answers only users you allowlist, or people who DM it, receive an 8-character pairing code and get approved by you with ",[92,586,587],{},"hermes pairing approve",". That default matters, because a bot anyone can message is a bot anyone can steer, and our ",[260,590,592],{"href":591},"/blog/ai-agent-security-guide","AI agent security guide"," covers what else to lock down.",[19,595,596],{},"Pi has no gateway. Its README points chat automation to pi-chat, a separate Earendil extension that bridges Discord and Telegram to a sandboxed Pi session, with one Gondolin micro-VM per channel plus its own memory files and skills.",[19,598,599],{},"It's a promising design, but read the label before you rely on it. pi-chat needs QEMU (plus tmux to run several channels), sits at version 0.1.0, was last updated on 5 June 2026, and its README lists Discord and Telegram only.",[19,601,602],{},"Pi Durable's launch post uses a Slack channel and its threads to show many conversations running at once. The Slack connection itself is code you'd write.",[105,604,606],{"id":605},"models-providers-and-cost","Models, providers and cost",[19,608,609],{},"Both projects are free, MIT-licensed software. You pay the model provider you connect, and each maker sells an optional shortcut.",[19,611,612,613,616],{},"Hermes switches models with ",[92,614,615],{},"hermes model",". Its docs list more than 40 provider options, from Nous Portal, OpenRouter, Anthropic and OpenAI to Bedrock, Vertex and any OpenAI-compatible endpoint, and it can download and manage llama.cpp to run local models itself.",[19,618,619,620,623,624,627],{},"Pi connects with ",[92,621,622],{},"/login"," or an environment variable. Its docs list more than 30 providers, with subscription sign-ins such as ChatGPT and GitHub Copilot, and local models run through its llama.cpp router support or any compatible endpoint in ",[92,625,626],{},"models.json",", Ollama and LM Studio included.",[19,629,630],{},"If you plan to use a Claude subscription, both say the same thing. Hermes needs Claude Max plus purchased extra usage credits (Pro can't sign in at all), and Pi warns that a subscription sign-in may draw on paid extra usage.",[28,632,633,644],{},[31,634,635],{},[34,636,637,640,642],{},[37,638,639],{},"Cost item",[37,641,54],{},[37,643,62],{},[44,645,646,656,667,678],{},[34,647,648,651,654],{},[49,649,650],{},"Software",[49,652,653],{},"Free, MIT",[49,655,653],{},[34,657,658,661,664],{},[49,659,660],{},"Model usage",[49,662,663],{},"Your provider, or a Nous Portal plan",[49,665,666],{},"Your provider, or Earendil's Radius gateway (early alpha)",[34,668,669,672,675],{},[49,670,671],{},"Hosted option from the maker",[49,673,674],{},"Hermes Cloud: one agent included with paid Nous Portal plans, from Plus at $20 a month",[49,676,677],{},"None listed; Radius is a model gateway, not an agent host",[34,679,680,683,686],{},[49,681,682],{},"Where it usually runs",[49,684,685],{},"An always-on machine: a VPS, a home server or Hermes Cloud",[49,687,688],{},"Your terminal, while you work",[19,690,691],{},[82,692,84],{},[19,694,695],{},"Hermes Cloud also deploys without a plan from a $2 minimum credit balance, billed daily for the compute and storage the agent uses, with model usage on top. Pi costs nothing to keep running, because it only runs while you do.",[19,697,698],{},"If what you want from Hermes is the always-on half (a schedule, a chat app, a memory) and not a machine to look after, that's the job BetterClaw does. It's a managed platform, not a Hermes or Pi host: your agents run on our infrastructure, recurring tasks follow a cron schedule, and Telegram and Slack come with Pro and above.",[19,700,701],{},"You bring your own key across 28+ providers at zero markup, with 500+ connectors plus any remote MCP server. The trade-offs are real: it isn't open source, it won't work inside your repo the way Pi does, and Starter runs schedules at most once a day and keeps agent memory for 7 days.",[19,703,704,705],{},"Plans start at $19 a month. ",[260,706,708],{"href":707},"/pricing","See BetterClaw plans",[105,710,712],{"id":711},"what-each-one-sends-before-you-type-we-measured","What each one sends before you type (we measured)",[19,714,715],{},"Every turn, an agent sends the model its system prompt and the definitions of every tool it can use, on top of your message. That fixed overhead decides how much of the context window is left for your work.",[19,717,718],{},"So we measured it on 16 November 2026. We installed both fresh with default settings (Hermes v0.23.1 from a source checkout of its release tag, Pi 1.5.2 from npm), pointed each at a local stand-in endpoint that records requests, opened an empty git repository and asked \"Say OK\".",[19,720,721],{},"We counted the first request with OpenAI's o200k tokenizer, after swapping our long sandbox folder paths for typical home-folder paths.",[28,723,724,736],{},[31,725,726],{},[34,727,728,730,733],{},[37,729],{},[37,731,732],{},"Hermes Agent v0.23.1",[37,734,735],{},"Pi 1.5.2",[44,737,738,749,760,771,782,793],{},[34,739,740,743,746],{},[49,741,742],{},"System prompt",[49,744,745],{},"1,862 tokens",[49,747,748],{},"638 tokens",[34,750,751,754,757],{},[49,752,753],{},"Tools offered",[49,755,756],{},"20",[49,758,759],{},"4",[34,761,762,765,768],{},[49,763,764],{},"Tool definitions",[49,766,767],{},"8,931 tokens",[49,769,770],{},"714 tokens",[34,772,773,776,779],{},[49,774,775],{},"Total before you type",[49,777,778],{},"about 10,800 tokens",[49,780,781],{},"about 1,350 tokens",[34,783,784,787,790],{},[49,785,786],{},"Share of a 32K context window (calculated)",[49,788,789],{},"about 33%",[49,791,792],{},"about 4%",[34,794,795,798,801],{},[49,796,797],{},"Share of a 64K context window (calculated)",[49,799,800],{},"about 16%",[49,802,803],{},"about 2%",[19,805,806],{},[82,807,84],{},[19,809,810],{},[252,811],{"alt":812,"src":813},"Hermes vs Pi measured: tokens of system prompt and tool definitions sent before the first message on fresh installs","/img/blog/hermes-vs-pi-tokens.jpg",[19,815,816],{},"Hermes's 20 tools cover the terminal, files, web search, a browser, memory, skills, delegation, code execution, connections, speech, vision and three tools for finding and calling more tools. Its system prompt also carries a one-line index of its 44 bundled skills, and full memory files add up to about 1,300 tokens more.",[19,818,819,820,823],{},"On a big cloud model with prompt caching, that gap is mostly a cost question. On a local model with a 32K window, ",[23,821,822],{},"Hermes spends a third of the context before you've said anything",", and Pi spends about 4%.",[105,825,827],{"id":826},"install-footprint-one-npm-package-or-a-full-toolchain","Install footprint: one npm package or a full toolchain",[19,829,830],{},"Pi installs as one npm package on Node.js 22.19 or newer, or through its installer, which pins every dependency. In our 16 November test, the npm install added 123 packages, took 17 seconds and used about 160 MB of disk.",[400,832,834],{"className":402,"code":833,"language":404,"meta":405,"style":405},"# Pi (macOS or Linux)\ncurl -fsSL https://pi.dev/install.sh | sh\n# or, on Node.js 22.19+\nnpm install -g --ignore-scripts @earendil-works/pi-coding-agent\n",[92,835,836,841,858,863],{"__ignoreMap":405},[409,837,838],{"class":411,"line":412},[409,839,840],{"class":494},"# Pi (macOS or Linux)\n",[409,842,843,846,849,852,855],{"class":411,"line":436},[409,844,845],{"class":415},"curl",[409,847,848],{"class":432}," -fsSL",[409,850,851],{"class":419}," https://pi.dev/install.sh",[409,853,854],{"class":542}," |",[409,856,857],{"class":415}," sh\n",[409,859,860],{"class":411,"line":451},[409,861,862],{"class":494},"# or, on Node.js 22.19+\n",[409,864,866,869,872,875,878],{"class":411,"line":865},4,[409,867,868],{"class":415},"npm",[409,870,871],{"class":419}," install",[409,873,874],{"class":432}," -g",[409,876,877],{"class":432}," --ignore-scripts",[409,879,880],{"class":419}," @earendil-works/pi-coding-agent\n",[19,882,883,884,887,888,121],{},"Hermes's installer clones the source, then pulls a pinned Python, Node.js, npm, pnpm, ripgrep and FFmpeg, plus a pinned Chromium for its browser tools and a computer-use driver. You can leave out the last two with ",[92,885,886],{},"--skip-browser"," and ",[92,889,890],{},"--skip-computer-use",[400,892,894],{"className":402,"code":893,"language":404,"meta":405,"style":405},"# Hermes Agent (Linux, macOS, WSL2)\ncurl -fsSL https://hermes-agent.nousresearch.com/install.sh | bash\nhermes model            # pick a provider and model\nhermes gateway setup    # connect Telegram, Slack, Discord and others\n",[92,895,896,901,915,925],{"__ignoreMap":405},[409,897,898],{"class":411,"line":412},[409,899,900],{"class":494},"# Hermes Agent (Linux, macOS, WSL2)\n",[409,902,903,905,907,910,912],{"class":411,"line":436},[409,904,845],{"class":415},[409,906,848],{"class":432},[409,908,909],{"class":419}," https://hermes-agent.nousresearch.com/install.sh",[409,911,854],{"class":542},[409,913,914],{"class":415}," bash\n",[409,916,917,919,922],{"class":411,"line":451},[409,918,416],{"class":415},[409,920,921],{"class":419}," model",[409,923,924],{"class":494},"            # pick a provider and model\n",[409,926,927,929,932,935],{"class":411,"line":865},[409,928,416],{"class":415},[409,930,931],{"class":419}," gateway",[409,933,934],{"class":419}," setup",[409,936,937],{"class":494},"    # connect Telegram, Slack, Discord and others\n",[19,939,940,941,944],{},"The official Docker image, ",[92,942,943],{},"nousresearch/hermes-agent",", is about 1.1 GB compressed for x86 on Docker Hub. Nous recommends 2 to 4 GB of RAM for the container, and at least 2 GB once browser tools are on.",[19,946,947,948,952],{},"That machine has to stay up, because Hermes earns its keep while you're away. Our guide to ",[260,949,951],{"href":950},"/blog/hermes-agent-hosting","Hermes Agent hosting"," compares Hermes Cloud, a VPS and a home server on real prices.",[105,954,956],{"id":955},"security-guardrails-in-the-agent-or-the-agent-in-guardrails","Security: guardrails in the agent, or the agent in guardrails",[19,958,959],{},"Hermes's security docs describe eight layers. The two you'll notice first are command approval and chat access.",[291,961,962,965,968,971],{},[294,963,964],{},"Hermes checks commands against a curated danger list, and in the default smart mode a helper model rates each flagged one: harmless ones run, dangerous ones are blocked, unclear ones ask you.",[294,966,967],{},"A hard blocklist (wiping the root folder, fork bombs, writing over a disk) stays on even in YOLO mode.",[294,969,970],{},"Chat access is denied until you allowlist a user or approve a pairing code.",[294,972,973],{},"Docker, Modal and Daytona backends skip the command check, because the container is the boundary.",[19,975,976],{},"Pi's docs take the other view. Pi doesn't ask for approval before every tool call and has no built-in permission system, so its tools and extensions run with your user's permissions.",[19,978,979,980,982],{},"Project trust decides whether a folder's ",[92,981,563],{}," settings and extensions load, but it doesn't sandbox anything, and AGENTS.md files load either way. Pi's security page calls running the whole process inside a container, VM or sandbox usually the strongest practical option, and its docs walk through four ways to do it: plain Docker, Docker Sandboxes, OpenShell and a Gondolin micro-VM extension.",[266,984,985],{},[19,986,987],{},"Hermes builds the guardrails into the agent. Pi expects you to build the box around the agent.",[19,989,990],{},"Neither default is safe on its own. Hermes still needs a tight allowlist and a sensible approval mode, and Pi's box only protects what you leave outside it.",[105,992,994],{"id":993},"skills-extensions-and-mcp","Skills, extensions and MCP",[19,996,997,998,1001],{},"Hermes skills live in ",[92,999,1000],{},"~/.hermes/skills/",". Version 0.23.0 trimmed the bundled set from 58 to 44, the agent writes new skills after complex tasks and improves them as it uses them, and a background curator archives the ones that go stale.",[19,1003,1004,1005,1008],{},"Its plugin catalog grew past 1,050 entries in v0.23.0. If you're still on that version, update: the v0.23.1 notes say the published catalog outgrew a 2 MiB client limit on 9 November, so ",[92,1006,1007],{},"hermes plugins install"," quietly fell back to the catalog your release shipped with.",[19,1010,1011,1012,1015],{},"For coding, Hermes ships skills that hand work to Claude Code and Codex, plus an optional OpenCode skill. Our ",[260,1013,1014],{"href":262},"Hermes vs OpenCode comparison"," walks through that hand-off; there's no equivalent skill for Pi.",[19,1017,1018,1019,1022],{},"Pi's extensions are TypeScript modules loaded into the Pi process. They can add tools, commands, keyboard shortcuts, providers and screen elements, and ",[92,1020,1021],{},"/reload"," picks up changes without a restart; sub-agents and plan mode are left to you, to Pi itself if you ask it to build them, or to a package.",[19,1024,1025],{},"The package gallery on pi.dev listed 5,366 packages when we checked on 16 November. Both projects follow the Agent Skills format for SKILL.md folders, and both speak MCP; Pi made it built in with version 0.99.0 on 29 September 2026.",[105,1027,1029],{"id":1028},"who-each-one-is-for","Who each one is for",[1031,1032,1034],"h3",{"id":1033},"you-want-an-assistant-that-runs-your-day","You want an assistant that runs your day",[19,1036,1037],{},"Pick Hermes. Morning briefings, a weekly report in Slack and reminders on Telegram need memory, cron and a gateway, and Hermes ships all three. Give it one job first, and widen its scope after a clean week.",[1031,1039,1041],{"id":1040},"you-write-code-and-want-to-see-every-token","You write code and want to see every token",[19,1043,1044,1045,1049],{},"Pick Pi. Four tools and a small prompt leave the context for your code, and every behaviour you add is code you can read. Run it in a container, or add a guard extension, before it touches anything important; our ",[260,1046,1048],{"href":1047},"/blog/pi-coding-agent-setup","Pi coding agent setup guide"," builds one step by step.",[1031,1051,1053],{"id":1052},"youre-building-an-agent-into-your-own-product","You're building an agent into your own product",[19,1055,1056],{},"Look at Pi Durable, or the Pi SDK if you don't need crash recovery. You get checkpoints, durable timers and many conversations as TypeScript parts, and you write the rest. Treat it as experimental, because its README does.",[1031,1058,1060],{"id":1059},"you-want-both","You want both",[19,1062,1063,1064,1067],{},"Run both. Hermes is the always-on front door and Pi is the worker in your terminal, and Hermes can call ",[92,1065,1066],{},"pi -p"," through its terminal tool like any other command. Both read AGENTS.md, so project instructions carry over.",[19,1069,1070],{},[252,1071],{"alt":1072,"src":1073},"Who should pick Hermes, Pi, Pi Durable or both, by job","/img/blog/hermes-vs-pi-who-for.jpg",[105,1075,1077],{"id":1076},"the-honest-answer-to-hermes-vs-pi","The honest answer to Hermes vs Pi",[19,1079,1080],{},"We went in expecting a heavyweight and a lightweight. What we found are two different promises.",[19,1082,1083],{},"Hermes promises to be there when you aren't: it keeps the memory, holds the schedule and answers the chat at 8 AM. Pi promises never to do more than you asked, and to show you everything it did.",[19,1085,1086],{},"Pick the promise your job needs. If you can't choose, you probably need both, in different places.",[105,1088,1090],{"id":1089},"want-the-always-on-part-handled-for-you","Want the always-on part handled for you?",[19,1092,1093],{},"If the Hermes side of this page sounds right but the server, the updates and the gateway don't, start with a conversation instead of a weekend of setup.",[19,1095,1096,1097],{},"Our AI automation audit starts with a 30-minute call about how your team works, and within 5 business days you get a written report with a proposed agent setup. ",[260,1098,1100],{"href":1099},"/ai-automation-audit","Book an AI automation audit",[105,1102,1104],{"id":1103},"frequently-asked-questions","Frequently Asked Questions",[1031,1106,1108],{"id":1107},"what-is-the-difference-between-hermes-agent-and-pi","What is the difference between Hermes Agent and Pi?",[19,1110,1111],{},"Hermes Agent is a durable personal agent from Nous Research: it keeps memory files about you, runs cron jobs from its gateway and answers on more than 20 chat platforms. Pi is a minimal terminal coding agent, started by Mario Zechner and now developed by Earendil, with four default tools and no approval prompts. Hermes runs your day; Pi works in your repo while you watch.",[1031,1113,1115],{"id":1114},"pi-durable-vs-hermes-is-pi-a-durable-agent-now","Pi Durable vs Hermes: is Pi a durable agent now?",[19,1117,1118],{},"The Pi coding agent saves your sessions but stops when its process stops, so you resume it yourself. Pi Durable is a separate experimental TypeScript package, shipped with Pi 1.0 on 1 October 2026, for building agents that survive crashes, sleep until a set time and run many conversations at once. It's a framework you code against, while Hermes is a durable agent you install.",[1031,1120,1122],{"id":1121},"is-hermes-or-pi-better-for-coding","Is Hermes or Pi better for coding?",[19,1124,1125],{},"Pi, for most coding work. It starts with four tools and, in our 16 November test, sent about 1,350 tokens before the first message against about 10,800 for Hermes, which leaves more context for your code. Hermes can code through its terminal tool, but it's built for automation and chat.",[1031,1127,1129],{"id":1128},"can-pi-run-on-a-schedule-or-reply-on-telegram-like-hermes","Can Pi run on a schedule or reply on Telegram like Hermes?",[19,1131,1132,1133,1135],{},"Not out of the box. Pi has no scheduler, so you call ",[92,1134,1066],{}," from your system's cron, and chat needs the separate pi-chat extension, which bridges Discord and Telegram to a sandboxed Pi session. Hermes does both through its gateway.",[1031,1137,1139],{"id":1138},"which-uses-fewer-tokens-hermes-or-pi","Which uses fewer tokens, Hermes or Pi?",[19,1141,1142],{},"Pi, by about eight times on the first request in our test: about 1,350 tokens of system prompt and tool definitions for Pi 1.5.2 against about 10,800 for Hermes v0.23.1, both fresh installs with default settings. Hermes's memory files can add up to about 1,300 more once they fill.",[1031,1144,1146],{"id":1145},"are-hermes-agent-and-pi-free","Are Hermes Agent and Pi free?",[19,1148,1149],{},"Yes. Both are free, MIT-licensed software, and you pay your model provider. Nous also sells Nous Portal plans from $20 a month that include one hosted Hermes Cloud agent, and Earendil runs Radius, a model gateway for Pi that is in early alpha.",[1031,1151,1153],{"id":1152},"can-i-use-hermes-and-pi-together","Can I use Hermes and Pi together?",[19,1155,1156,1157,1159],{},"Yes. Hermes can run ",[92,1158,1066],{}," through its terminal tool, though it ships no official Pi skill, and both read AGENTS.md files and Agent Skills folders. Most people who run both use Hermes as the always-on front door and Pi in the terminal for code.",[1161,1162,1163],"style",{},"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 pre.shiki code .sYu0t, html code.shiki .sYu0t{--shiki-default:#005CC5}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);}html pre.shiki code .sAwPA, html code.shiki .sAwPA{--shiki-default:#6A737D}html pre.shiki code .sgsFI, html code.shiki .sgsFI{--shiki-default:#24292E}html pre.shiki code .sD7c4, html code.shiki .sD7c4{--shiki-default:#D73A49}",{"title":405,"searchDepth":436,"depth":436,"links":1165},[1166,1167,1168,1169,1170,1171,1172,1173,1174,1175,1176,1182,1183,1184],{"id":107,"depth":436,"text":108},{"id":273,"depth":436,"text":274},{"id":323,"depth":436,"text":324},{"id":391,"depth":436,"text":392},{"id":577,"depth":436,"text":578},{"id":605,"depth":436,"text":606},{"id":711,"depth":436,"text":712},{"id":826,"depth":436,"text":827},{"id":955,"depth":436,"text":956},{"id":993,"depth":436,"text":994},{"id":1028,"depth":436,"text":1029,"children":1177},[1178,1179,1180,1181],{"id":1033,"depth":451,"text":1034},{"id":1040,"depth":451,"text":1041},{"id":1052,"depth":451,"text":1053},{"id":1059,"depth":451,"text":1060},{"id":1076,"depth":436,"text":1077},{"id":1089,"depth":436,"text":1090},{"id":1103,"depth":436,"text":1104,"children":1185},[1186,1187,1188,1189,1190,1191,1192],{"id":1107,"depth":451,"text":1108},{"id":1114,"depth":451,"text":1115},{"id":1121,"depth":451,"text":1122},{"id":1128,"depth":451,"text":1129},{"id":1138,"depth":451,"text":1139},{"id":1145,"depth":451,"text":1146},{"id":1152,"depth":451,"text":1153},"Comparison","2026-11-16","Hermes vs Pi compared on memory, cron, chat apps, models, cost and safety. In our test Hermes sent about 10,800 tokens before you type, Pi about 1,350.","md",false,"/img/blog/hermes-vs-pi.jpg","Hermes vs Pi cover: a Hermes Agent chat card with memory and a cron job next to a Pi terminal card with four tools",512,1024,null,{},true,"/blog/hermes-vs-pi","13 min read",{"title":5,"description":1195},"Hermes vs Pi: Durable Agent or Minimal Coding Agent?","blog/hermes-vs-pi",[1211,1212,1213,1214,1215,1216,1217],"hermes vs pi","pi vs hermes","hermes agent vs pi","pi agent vs hermes","pi durable vs hermes","hermes vs pi vs opencode","pi coding agent setup","9QelTLCLWLzNQmMjnqZ9W5hKZQ86mUanSOXs_TD1m38",[1220,1229,1237],{"stem":1221,"title":1222,"description":1223,"date":1224,"image":1225,"category":1226,"redirected":1197,"author":1227,"readingTime":1228},"blog/hermes-agent-hosting","Hermes Agent Hosting: Hermes Cloud, VPS, Home Server or Managed","Hermes Agent hosting: Hermes Cloud, a VPS like Hostinger or Hetzner, a home server or a managed alternative. Real costs and trade-offs.","2026-11-14","/img/blog/hermes-agent-hosting.jpg","Guides",{"name":7,"role":8,"avatar":9,"bio":10},"10 min read",{"stem":1230,"title":1231,"description":1232,"date":1233,"image":1234,"category":1226,"redirected":1197,"author":1235,"readingTime":1236},"blog/hermes-agent-latest-version","Hermes Agent Latest Version: What Changed in v0.23 and How to Update Safely","Hermes Agent latest version is v0.23.1. What changed in v0.23 (Webapp, one gateway, /rewind, plugins), what moved, how to update and what breaks.","2026-11-10","/img/blog/hermes-agent-latest-version.jpg",{"name":7,"role":8,"avatar":9,"bio":10},"9 min read",{"stem":1238,"title":1239,"description":1240,"date":1241,"image":1242,"category":1226,"redirected":1197,"author":1243,"readingTime":1244},"blog/hermes-agent-desktop-app","Hermes Agent Desktop App (v0.22.7): Install, Local Models, Plugins and Fixes","Install the Hermes Agent desktop app on Windows, Mac or Linux, run a local model with no API key, add plugins safely and fix common v0.22.7 errors.","2026-11-05","/img/blog/hermes-agent-desktop-app.jpg",{"name":7,"role":8,"avatar":9,"bio":10},"11 min read",[],1795019549967]