Hermes Agent

A terminal agent you can also message on Telegram

Hermes Agent is Nous Research's open-source AI agent, released under the MIT license. You start it in a terminal with hermes, open it in the desktop app, or run a gateway so it answers on Telegram, Discord, Slack, WhatsApp, Signal or email. It runs shell commands and scheduled jobs, and it carries memory and self-written skills from one session to the next. The README stresses that the agent isn't tied to your laptop: it can work on a cloud VM while you talk to it from Telegram.

Skills written on the job, with swappable models and sandboxes

• Skills from experience: after a complex task, the agent turns what it did into a skill and keeps improving it as it uses it. Skills follow the open agentskills.io format.
• Cross-session memory: the agent curates its own notes and recalls old conversations through FTS5 full-text search, with the LLM summarizing the hits. Since 0.22.0, external memory providers such as Hindsight install as catalog plugins.
• Any model: Nous Portal, OpenRouter, OpenAI, your own endpoint or, since 0.22.0, a local model on llama.cpp, which Hermes installs and keeps updated. hermes model switches provider with no code change.
• Six execution backends: local, Docker, SSH, Modal, Daytona and Vercel Sandbox. On Modal and Daytona the environment hibernates between sessions, and the README says it costs nearly nothing while idle.
• Cron and subagents: the built-in scheduler runs jobs written in plain language and delivers the output to whichever platform you pick. Isolated subagents split parallel work between them.

Unattended reports and an assistant on a rented server

• An assistant on a VPS or cloud VM that you message from Telegram while it works.
• Daily reports, nightly backups or weekly audits described in plain language and run by cron with nobody watching.
• Batch trajectory generation for training tool-calling models, the research use Nous Research puts forward.

Host execution by default, a stable channel since 0.22

The default backend, local, isolates nothing. Commands run straight on the machine, as your own user. Those that match a list of dangerous patterns need approval: in the default smart mode, a second LLM lets low-risk ones through, blocks the clearly dangerous ones and only asks a human about the uncertain cases. The security page of the docs warns that the write guard won't stop a hostile or compromised agent, and its production checklist calls for the Docker backend and explicit user allowlists. With no allowlist at all, the gateway turns everyone away.

Only the desktop app comes with a trial model. The CLI, TUI and gateway start with no account and no free model, so they need a provider key, a Nous Portal subscription (300+ models plus tools such as web search) or a local model.

The README calls Hermes "the only agent" with a built-in learning loop, and it courts people coming from other agents: hermes setup looks for ~/.openclaw, ~/.claude, ~/.codex and ~/.cursor, then offers to import their settings, memories and skills.

The latest release, 0.22.5, came out on October 23, 2026, and the project is still at 0.x. Eight releases shipped between October 8 and 23. Since 0.22.0, every SemVer release goes through install and upgrade runs on each OS before publication, and source installs track those releases by default instead of every commit on main. The macOS desktop app is built for Apple Silicon only by default, and pip or Homebrew installs are not supported.