Skip to content

Hermes Agent

CompanyNous Research
TypeTerminal TUI + messaging gateway + desktop
PricingFree — open source, bring your own model
LicenseMIT
ProtocolMCP, ACP, agent skills
Config FormatsAGENTS.md, SKILL.md
RepositoryNousResearch/hermes-agent
Websitehermes-agent.nousresearch.com

Hermes is a self-hosted agent built around a closed learning loop. Where most agents start each session from nothing but your repo, Hermes keeps what it learns: it curates its own memory, searches its past conversations, and builds a model of how you work across sessions.

It is not tied to a terminal. Hermes runs as a persistent daemon on a VPS, a GPU box, or serverless infrastructure, and you talk to it from Telegram, Discord, Slack, WhatsApp, Signal, or the CLI through a single gateway process.

This is the part worth paying attention to, and it is unusual.

  • Autonomous skill creation: after a complex task, Hermes writes the procedure down as a skill — the steps, the pitfalls it hit, and how to verify the result
  • Skills that improve during use: a skill is revised as the agent works with it, rather than frozen at the moment it was written
  • Memory with nudges: the agent is periodically prompted to persist what matters instead of losing it at the end of a session
  • Cross-session recall: full-text search over past conversations, with summarization for retrieval

The premise matches the case for skills generally. The first run of a new task is expensive because the agent has to work everything out; the second should not be. A skill is where that first run gets stored.

  • Model-agnostic: Nous Portal, OpenRouter, NVIDIA NIM, z.ai/GLM, Moonshot, MiniMax, Hugging Face, OpenAI, or your own endpoint — switch with hermes model
  • Lives where you do: 16+ messaging platforms from one gateway, with voice memo transcription and continuity across them
  • Runs anywhere: a $5 VPS is enough; it costs nearly nothing when idle on serverless
  • Real terminal interface: multiline editing, slash-command autocomplete, interrupt-and-redirect, streaming tool output
  • Scheduled work: cron tasks, so the agent does things without being asked
  • Self-hosting is the model, so you own the deployment, the keys, and the upgrades
  • Broader in scope than a pure coding agent — strong at long-running personal and operational work, less specialized than a harness built only for repositories
  • The learning loop needs time to be worth anything; on day one it knows nothing about you

Hermes writes its own skills, but it does not have to start from a blank file. The largest open catalogue is atskills.one, with 60,000+ skills — see @skills for how referencing and saving work.

  • Long-running assistants that should get better at your work over time
  • Teams that want the agent reachable from chat rather than only a terminal
  • Self-hosted and privacy-bound setups
  • Anyone who wants the agent to capture procedures instead of rediscovering them