What is Hermes? Nous Research's self-improving open-source agent
Hermes Agent is an MIT-licensed agent that writes its own skills, asks before risky commands and works with any model, local ones included. What it is, what it costs, how it installs.
Hermes Agent is Nous Research’s free, open-source personal AI agent: MIT-licensed, written mostly in Python, and released in February 2026. Nous’s own line for it is “The Agent That Grows With You”: it keeps a small memory of you, searches its past conversations, and writes, improves and archives its own skills as it works. It runs on a computer you choose, works with almost any model provider or a local model, and you talk to it from the terminal, its desktop app for macOS, Windows and Linux, a browser dashboard, or any of the 28 messaging platforms its docs list. The current release is v0.21.6, from October 8, 2026. One disambiguation before anything else: “Hermes” is also the name of Nous Research’s LLM model series. The agent and the models are separate products that share a brand.
Agent vs models: the naming trap
Nous Research has shipped Hermes-branded model weights since 2023: Hermes 3, Hermes 4 (405B/70B/14B, August 2025), Hermes 4.3, all open models with a strong local-inference story. Hermes Agent arrived later: a harness, like OpenClaw or Claude Code, that uses models rather than being one. Nothing requires pairing them, and Nous now says outright that its Hermes 4 models “are not recommended for use inside Hermes Agent,” because they are tuned for chat and reasoning rather than the rapid tool-calling an agent does. When someone says “Hermes” in an agent-hosting conversation, they usually mean the agent; when a benchmark thread says it, they mean the models.
What makes it distinctive
Every serious harness has tools, channels and memory. The README leads with the part Hermes claims as its own: “a built-in learning loop — it creates skills from experience, improves them during use, nudges itself to persist knowledge, searches its own past conversations, and builds a deepening model of who you are across sessions.” In practice that is three pieces:
- Skills it writes itself. The agent saves a skill when it has worked out a multi-step workflow, found a way past an error, or been corrected, and edits or deletes its skills as it goes; its docs call this “the agent’s procedural memory.” By default it writes freely; one setting stages every change for your approval. A curator runs every seven days and moves the unused skills its background review wrote to an archive, never deleting them. Skills are
SKILL.mdfiles in the open Agent Skills format, and Hermes also installs other people’s from skills.sh, LobeHub and ClawHub, with a security scan on each install. - A small memory, kept on purpose. Two plain Markdown files in
~/.hermes/memories/:MEMORY.md, the agent’s notes, capped at 2,200 characters, andUSER.md, its picture of you, at 1,375. Both go into the prompt at the start of each session. Everything else lives in a searchable database of past conversations that the agent queries when it needs it. - One agent everywhere. One background process carries the same agent, memory and skills into the terminal, Hermes Desktop, a browser dashboard and the 28 messaging platforms its docs list, among them Telegram, WhatsApp, Discord, Slack, Signal, iMessage, Microsoft Teams, LINE and email.
Around that sit about 100 built-in tools (a browser, terminal, vision, image generation, speech, MCP), scheduled jobs that remember what earlier runs did, subagents you can steer while they work, and background computer use on macOS, Windows and Linux with any model that can call tools. Four newer pieces have pages of their own: Hermes Desktop, the native app; Bot Mode, built into it since v0.21.0 in August, which turns profiles into a roster of named Bots; Hermes kanban, a task board those profiles work from; and Agent Client Protocol support, which seats Hermes inside editors.
How it stays safe, by default and in Nous’s words
Hermes asks before risky work. Its default approval mode has a second model judge each command: low-risk ones run, dangerous ones are refused, and the rest come to you, with an unanswered prompt refused after five minutes; scheduled and unattended runs refuse anything that needs approval, and a short list of catastrophic commands is refused whatever you set. Its security policy is clear about what that is not: “Hermes Agent is a single-tenant personal agent,” and “The only security boundary against an adversarial LLM is the operating system.” For an agent that reads the open web or inbound email, Nous’s supported answer is to run the whole thing in a container or on a separate machine. Its docs also record why the dashboard now demands a sign-in on any non-local address: in June 2026, scanners found dashboards left open to the internet and used them to make the agent plant an SSH backdoor.
Models: bring anything, including local
Hermes works with about 45 providers and routes, from OpenAI, Anthropic, Google and OpenRouter to DeepSeek, Kimi, xAI and any OpenAI-compatible endpoint. Some subscriptions sign in directly: ChatGPT through Codex, GitHub Copilot, SuperGrok, and Claude only on a Max plan with extra usage purchased, which is what that route draws on. Nous’s own provider page now calls its Nous Portal “the recommended way to run Hermes Agent”: one sign-in for hundreds of models plus hosted tools such as web search and image generation. For local models, Hermes can now download and run llama.cpp itself, choosing a build for your hardware, or point at Ollama, vLLM, SGLang or a llama.cpp server, stretching its timeouts to 30 minutes when it detects one, since local prefill on big contexts is slow. Whatever the model, it needs a context window of at least 64,000 tokens, or Hermes refuses it at startup. How well small local models hold up through long agentic runs is something practitioners debate, not something the docs promise.
The software is free. The paid parts are the model you choose and, if you want them, Nous’s own services: a free Portal tier limited to free models, paid Portal plans at $20, $100 and $200 a month that come with the same amount in credit plus a bonus, and Hermes Cloud, a hosted instance billed by the day while it runs.
Installing and running it
The official paths, from Nous’s installation page: a script for Linux, macOS and WSL2 (install.sh piped to bash, the distribution model the whole agent generation shares, with the same verify-what-you-run caveat), a PowerShell installer for native Windows, a Docker image that keeps everything in one mounted folder, and Hermes Desktop downloads for macOS and Windows. Android goes through a signed Termux package rather than the script, and on Linux the desktop app is built from a terminal install with hermes desktop. The installer brings its own Python, Node.js and tools, and by default a browser and the computer-use driver too; Homebrew, pip and the AUR are listed as unsupported.
Like OpenClaw, it is an always-on gateway process at heart: hermes gateway install writes the background service (systemd on Linux, launchd on macOS) and hermes gateway start starts it, so the where-does-it-live question is the same one, a machine that stays on. The README’s list starts at “a $5 VPS.” Nous publishes no hardware minimums; what it does publish is that the background process idles at about 300 MB of memory, and that its own hosted instances come with 2 GB and 4 vCPUs or 4 GB and 8 vCPUs. Budget it like a Python service with a browser beside it, unless you are serving local models on the same box, which changes the math entirely.
The OpenClaw relationship
Hermes and OpenClaw are the two most-watched open-source personal agents (252,201 and 391,534 GitHub stars on October 9, 2026), and Hermes courts switchers directly: its setup notices ~/.openclaw and offers hermes claw migrate, which brings across the persona, memory, skills, model settings, MCP servers and chat-app tokens after showing you a preview. API keys come only if you add --migrate-secrets, scheduled jobs and plugins are archived for you to set up again, and WhatsApp needs pairing again. One habit doesn’t carry: where OpenClaw binds multiple independent agents to specific chats, Hermes runs one agent across your surfaces and uses profiles for separate instances, each of which needs its own WhatsApp number. OpenClaw imports in the other direction too, and the two have grown alike: OpenClaw now writes its own skills as well. OpenClaw vs Hermes sets them side by side. Both assume what this whole category assumes: a real computer of the agent’s own, running around the clock.