Nous Research · Hermes Agent
Hermes Agent: free software, on someone else's meter
Nous Research's terminal coding agent — MIT-licensed, reachable from Telegram and Slack as well as a shell, and sold to you never: every way in buys a model instead.
Hermes Agent is a terminal coding agent with persistent memory, skills it writes for itself, and a gateway that answers on the messaging platforms you already use. It is MIT-licensed and the vendor sells no edition of it. What costs money is model access, bought ten ways — including one, a model on your own machine, that needs no account at all. The vendor's own gateway meters in dollars rather than a private unit.
What Hermes Agent actually is
Start with the name, because three different things answer to it. There is a French luxury house whose brand crowds the search results. There is Hermes 4, a family of chat models from the same company as this agent — and the company's own documentation tells you not to use those models inside this agent, because they are tuned for conversation rather than the rapid-fire tool-calling loop an agent runs on. And there is Hermes Agent, this one: a program you install and run in a project directory, from Nous Research, under the MIT licence.
What it does is broader than a coding agent's job. It is a terminal coding agent in the ordinary sense — it reads your repository, runs commands, edits files, delegates to subagents in their own git worktrees, and speaks MCP. It also keeps memory across sessions, writes and files its own reusable skills, and runs a gateway that puts the same agent, with the same memory, on Telegram, Discord, Slack, WhatsApp, Signal and email. There is a native desktop app for all three platforms, a cron scheduler you drive in plain language, and a one-command import of an existing Claude Code or Codex CLI setup — allowlists, MCP servers, skills and memories included.
Commercially it is the cleanest shape in this section. The client is free software with no paid edition above it, so nothing here is a trial. Ten routes reach a model, and they are genuinely different purchases: a local runtime that never signs in, your own API key, four tiers of the vendor's own gateway, two consumer subscriptions you may already hold, a developer seat administered by somebody else, and an enterprise deal. The gateway's credits are denominated in dollars — the vendor's own client documentation insists the interface print them that way — which spares this page the arithmetic that Codex, Cursor and Copilot all require.
We have not run it. The routes that would show how the meter behaves need an account with the vendor, and creating accounts is not something this desk does. Everything below was read on the vendor's own documentation, pricing surfaces, model catalogue, terms and privacy policy on a single day, and the places where two of those disagree are shown rather than quietly resolved.
Ten routes, and the agent is not for sale on any of them. The table below is a list of ways to buy a model, one of which reaches a true zero — and the vendor's own published discounts turn out to be checkable against our rate table, with one instructive exception.
| Route | Bought as | Price |
|---|---|---|
| Local model | Own hardware | Your hardware |
| Your own API key | API key | At model rates |
| Nous Portal Free | Subscription | $0/mo |
| Nous Portal Plus | Subscription | $20/mo |
| Nous Portal Super | Subscription | $100/mo |
| Nous Portal Ultra | Subscription | $200/mo |
| Claude Max 5x | Subscription | $100/mo |
| ChatGPT Plus | Subscription | $20/mo |
| GitHub Copilot Pro | Per seat | $10/seat/mo |
| Hermes Enterprise | Per seat | Contact sales |
When the honest answer is use something else
local, so commands the agent decides to run execute on your machine. Isolation is real and well documented — Docker, Modal, Daytona, Singularity and a serverless sandbox, with capabilities dropped, privilege escalation blocked, a process limit and a no-exec temp filesystem — but every one of those is a configuration you apply. What guards the default instead is an approval layer, and its default mode asks a second model to judge each risky command rather than matching a pattern list. The reasons to leave this tool are unusually specific — a training default, a sandbox that is off, a project that never stops moving, a bill you cannot see from inside — and each points at a different replacement rather than at a general top ten.
Frequently asked.
What people search next.