Skip to content

Harness profile

Hermes Agent: an open-source agent from Nous Research that learns its own skills

Hermes Agent is an open-source AI agent from Nous Research that runs on a server or your own machine, talks to you through the terminal or chat apps, and writes reusable skills from the work it does. It suits people who want a long-lived agent that remembers, runs scheduled jobs and improves with use, on any model provider.

StudiedLast reviewed 4 min read

Explained from public sources. Mai has not built this yet.

On this page

In plain English

Hermes Agent is an AI assistant from Nous Research that you install and keep running, on a laptop, a cheap cloud server or a hosted sandbox. You talk to it in a terminal or through apps such as Telegram, Slack, WhatsApp, Teams or email. Its distinguishing idea is that it learns: when it works out how to do something, it can write that down as a skill and use it next time. It also keeps notes about you and its environment, and can run jobs on a schedule while you are away.

How it works

Hermes is open source under the MIT licence. It runs on Python (current installs use Python 3.14), and the official installer brings its own pinned Python and Node.js rather than using what is on your machine. It installs with one script on Linux, macOS and WSL2, natively on Windows through PowerShell, and on Android through Termux. Intel Macs are not supported. After install, hermes opens an interactive terminal session and hermes gateway starts the messaging side.

Model choice is open. The docs list Nous Portal, OpenRouter, OpenAI and custom endpoints, switched with hermes model and no code changes. Nous Portal also offers a single subscription for extra tools such as web search, image generation and a cloud browser, though you can bring your own keys for each.

Reach it: Terminal, Telegram, Slack, Email. Then Gateway and agent loop. Then Work: Built-in tools, MCP servers, Sub-agents. Then Runs on: Local, Docker, SSH, Cloud sandbox. Then Learns: Skills, Memory files, Session search.

One gateway, one agent loop, and a learning loop that writes skills and memory back to disk

The main parts:

  • Gateway. One process serves the terminal and a long list of platforms, including Telegram, Discord, Slack, WhatsApp, Signal, Matrix, Microsoft Teams, Google Chat, email and SMS. Unknown users get a pairing code that the owner approves before the agent will talk to them.
  • Terminal backends. Commands can run locally, in Docker, over SSH, or in hosted sandboxes such as Modal, Daytona and Vercel Sandbox. Some of the hosted options hibernate when idle and wake on demand, so a long-lived agent does not need a server running all the time.
  • Skills. A skill is a SKILL.md file with a procedure, pitfalls and checks. The agent sees a short list first and loads the full text only when needed. It creates, patches and deletes its own skills through a dedicated tool, and the docs say the system prompt asks it to save repeatable workflows. Skills follow the agentskills.io format and can be installed from a hub, with a built-in scanner checking each install for exfiltration, prompt injection and destructive commands.
  • Memory. Two capped files, MEMORY.md for the agent’s own notes and USER.md for your preferences, are loaded into each session as a fixed snapshot. Every past session is stored in SQLite and searchable by full text. External providers such as Honcho or Mem0 can run alongside for semantic recall.
  • Scheduling, sub-agents and MCP. A built-in cron scheduler runs jobs written in plain language and delivers results to any connected platform. The agent can spawn isolated sub-agents for parallel work. MCP servers connect over stdio or HTTP, with per-server tool filters.

What it does best

Hermes is built for an agent that stays with you and improves over time. The learning loop is its clearest strength: procedures become skills, skills get patched when they go wrong, and memory is deliberately small so it stays useful rather than turning into a junk drawer. The range of messaging platforms and execution backends is wide, and being free to pick any model provider avoids lock-in. For researchers, the docs also mention trajectory export and reinforcement learning training through Nous Research’s Atropos.

Where it stops

It is a complete agent application rather than a library for building your own. If you want a different control flow, you work within its structure or extend it with plugins. Memory is intentionally bounded: the two curated files hold roughly 800 and 500 tokens and do not compact themselves, so the agent must tidy them when they fill up, and mid-session memory changes only appear in the prompt from the next session. Self-written skills save freely by default; staging them for review is an option you have to turn on.

Where it goes wrong

The docs are candid about the limits of their own safeguards. Checks on context files for prompt injection are described as heuristics rather than an understanding of intent. File-write denylists are called defence in depth, not a hard boundary, because the terminal tool runs as the same user. Inside Docker and the hosted sandboxes the dangerous-command checks are skipped, on the basis that the container is the boundary, so anything forwarded into the container is exposed to code running there. A “yolo” mode switches approvals off for a session, with only a short hard blocklist left. Skills learned from a bad run can encode a bad habit until someone reviews them.

How Mai AI uses it

I would suggest Hermes to someone who wants a personal, long-running agent on their own server with a free choice of model. For a client system I would not hand a self-editing agent the keys to business data without the review gate switched on.

The idea I take from it is the learning loop itself: turning a solved problem into a short written procedure, and keeping memory small and curated rather than storing everything. That shapes how I think about self-improving skills and agent memory. Its tiered approval modes are a useful reference for human approval gates. For how these ideas fit a system built around one business, see how Mai AI designs custom AI operations systems.

Questions and answers

Is Hermes Agent free?

Yes. Hermes Agent is open source under the MIT licence. You pay for the model provider you connect, whether that is Nous Portal, OpenRouter, OpenAI or another endpoint.

How does Hermes Agent learn new skills?

After a non-trivial task, a dead end that turned into a working path, or a correction from you, the agent can save the procedure as a skill using its skill management tool. It can patch that skill later as it learns more. You can require approval before any skill change is applied.

How does Hermes Agent remember things?

It keeps two small curated files, one with its own notes and one about you, which are loaded at the start of each session. It can also search every past conversation, which is stored in SQLite with full-text search, and you can add an external memory provider.

Can Hermes Agent use MCP servers?

Yes. It connects to MCP servers over stdio or HTTP, with per-server allowlists and blocklists for their tools. An official plugin also lets Hermes act as an MCP server for other clients.

Does Hermes Agent ask before running dangerous commands?

By default it uses a smart mode, where a second model rates flagged commands, approves low-risk ones, denies dangerous ones and asks you about the rest. You can switch to asking every time, or turn checks off, which the docs advise only for trusted environments.

Sources

  1. hermes-agent on GitHub (README)github.com
  2. Hermes Agent documentation homehermes-agent.nousresearch.com
  3. Installationhermes-agent.nousresearch.com
  4. Securityhermes-agent.nousresearch.com
  5. Skillshermes-agent.nousresearch.com
  6. Memoryhermes-agent.nousresearch.com
  7. MCP integrationhermes-agent.nousresearch.com