# Hermes Agent review: experiences, strengths and weaknesses | llmnet Apps

[← Apps overview](/en/)

# Hermes Agent: review

By Ivo Donker · October 28, 2026 — created with AI assistance (writer: DeepSeek V4.1 Flash; critic: Gemini 3.8 Flash + gpt-oss-120b via Groq)

Hermes Agent: [News](/en/hermes-nieuws)

Current as of September 2026, compiled from our own notes and news archive; text machine-generated and editorially checked. Check versions and commands in the official documentation.

Hermes Agent from Nous Research is not a coding agent in the classic sense, but a personal, self-hosted agent that learns from what it does. It remembers your preferences, creates and improves its own skills, carries out scheduled tasks and is reachable via Telegram, Discord or Slack. This review looks at what that delivers in practice (as of September 2026).

## What it is and who it is for

Nous Research describes Hermes as "the self-learning AI agent": an agent with a built-in learning loop that creates skills from experience, that improves during use, records knowledge and builds an increasingly better picture of you across sessions. It is open source under the MIT licence.

Hermes suits anyone who wants an agent that runs for them day in, day out: research, monitoring, summaries and automation, alongside programming work. It is less focused on interactively working through a single codebase than Claude Code, Codex or Cursor.

## How it works

Hermes runs on your own machine or server. The configuration lives in ~/.hermes/: config.yaml for settings, .env for secrets, SOUL.md for the agent's identity, plus folders for memory and skills.

- Memory: maintained by the agent itself and searchable, with memories from earlier sessions.

- Skills: folders with a SKILL.md following the open standard from agentskills.io. The agent creates, modifies and improves them itself, and you can install them from various registries.

- Terminal backends: commands run locally, in Docker, via SSH or in cloud sandboxes such as Modal, Daytona and Vercel Sandbox, or on HPC clusters via Singularity.

- Messaging gateway: more than twenty chat platforms, with an allowlist or pairing codes as access control.

- Planner: scheduled tasks in plain language or via hermes cron, with delivery to any linked channel.

- Also: subagents for parallel work and MCP for external tools.

## Strengths

- It really does get better with use. Because the agent records recurring ways of working as skills, you don't have to explain things over and over again. The step-by-step loading strategy keeps context costs low, even with many skills.

- Always reachable. Via the gateway you can send a task from your phone, and scheduled tasks deliver their result wherever you want.

- Flexible in where it runs. From your laptop to a Docker container or an HPC cluster: the backend is a single setting.

- Security taken seriously. No chat user is allowed anything by default. Skills from outside are scanned, and a "dangerous" verdict cannot be overruled. Risky commands can require approval, and updates first make a snapshot with automatic rollback.

- Open and model-agnostic. MIT licence, multiple providers, and a cheaper helper model for side tasks such as summarising context.

## Weaknesses and risks

- An agent that adapts itself needs watching. Automatically updated skills and memory are powerful, but a mistake that ends up in a skill repeats itself. Turn on write_approval for skills until you trust the behaviour.

- The default backend is your own machine. Without Docker or another sandbox, the agent executes commands directly on your system. For long-running, autonomous work, a sandboxed backend is effectively mandatory.

- A gateway is an attack surface. A chatbot with access to a terminal is dangerous if access control is misconfigured; GATEWAY_ALLOW_ALL_USERS=true is therefore not advisable.

- Limited number of platforms. Macs with an Intel processor are not supported.

- Less suitable as a pure code editor assistant. For intensive work in a single repository, with diffs, reviews and IDE integration, Claude Code, Cursor or Codex are better equipped.

- Fast-changing project. Frequent releases bring new features, but also migrations. hermes config check after every update is no unnecessary luxury.

## Price and licence

Hermes Agent itself is free and open source (MIT). You only pay for the models you use and possibly for the infrastructure it runs on (a server or cloud sandboxes). Anyone who routes routine work to a cheap helper model or a local model keeps costs low.

## Alternatives

- OpenCode: open-source coding agent with free choice of model, stronger at interactive work on code.

- Claude Code: strong at code work, hooks and CI, with a built-in sandbox, but Claude models only and not open source.

- Codex: OpenAI's open-source CLI with a good sandbox model.

- Cursor: for anyone who wants a graphical editor.

## Conclusion

Hermes Agent is the best choice in this overview if you are looking for an agent that keeps working for you: scheduled tasks, reachability via chat, a growing memory and self-learned skills, fully self-managed. Do set it up deliberately, though: a sandboxed terminal backend, approval for changes to skills, and a tightly locked-down gateway. For intensive programming work in one codebase, you are best off combining it with a coding-agent such as Claude Code, Codex or OpenCode.