Hermes Agent is Nous Research’s open-source agent, described in its README as “The self-improving AI agent built by Nous Research.” It is MIT licensed, and its defining characteristic is that it is not confined to a terminal on one machine.
Reach is the actual feature
Most agent projects give you a CLI. Hermes gives you a terminal UI (hermes) plus a messaging gateway spanning Telegram, Discord, Slack, WhatsApp, Signal and CLI — the README’s phrasing is that it is “not tied to your laptop.”
Two further pieces extend that. A built-in cron scheduler can deliver results to any supported platform, so a task can run overnight and report to Signal. And it can spawn isolated subagents for parallel work.
Execution is similarly spread out: seven terminal backends are listed, including local, Docker, SSH, Modal and Daytona. So the agent you message from your phone can be executing in a container, on a remote host, or on managed compute.
Models are deliberately unopinionated — any provider works, including Nous Portal, OpenRouter, OpenAI or a custom endpoint, switched with hermes model, which the README sells as “no code changes, no lock-in.”
The “self-improving” claim, unpacked
This is the phrase that needs examining, because it is doing a lot of work. The README describes a “closed learning loop” with four components:
- Agent-curated memory, with periodic nudges to persist knowledge.
- Autonomous skill creation after complex tasks, with skills that improve during use.
- FTS5 session search with LLM summarisation for cross-session recall.
- Honcho dialectic user modelling, building a model of the user across sessions.
It also states compatibility with the agentskills.io open standard, which matters for portability of anything it writes.
Read carefully, three of those four are retrieval and persistence rather than learning in any weight-updating sense: the agent writes notes, writes skills, searches its own history and models your preferences. That is a sensible architecture and it genuinely compounds over time — an agent that wrote itself a working skill last week starts ahead this week.
But the README says it “builds a deepening model of who you are across sessions” without explaining the mechanism, and offers no measurement of whether the loop improves outcomes. Take “self-improving” as a description of the design’s intent rather than a demonstrated result, and judge it on your own usage.
Installing
curl -fsSL https://hermes-agent.nousresearch.com/install.sh | bashNative Windows PowerShell is supported without WSL:
iex (irm https://hermes-agent.nousresearch.com/install.ps1)There is also a signed APT repository for aarch64 Android devices via Termux. Reload your shell, then run hermes; other commands include hermes model, hermes tools, hermes gateway, hermes setup, hermes doctor and hermes update. A Nous Portal subscription covers model, web search, image generation, TTS and a cloud browser in one, with hermes setup --portal; otherwise bring your own provider and per-tool keys are optional.
The installer handles Python 3.14, Node.js, npm, ripgrep, FFmpeg and Python dependencies. As always with a piped install script, read it first.
Rough edges
Native Windows support carries a “Heads up” note and a request to file bugs. Windows Defender or Bitdefender may quarantine uv.exe as a false positive — the documentation gives verification steps and folder-exclusion advice, which is the correct response rather than telling you to disable your scanner. The OpenClaw migration imports a specific set of items (persona, memories, skills, allowlists and a limited set of API keys), with flags like --overwrite needed for some options.
Notably, the README has no dedicated limitations section, which for a project this ambitious is itself worth knowing.
Who it is for
Someone who wants an agent reachable from wherever they are, running wherever makes sense, accumulating skills and context as it goes — and who is comfortable being an early adopter on Windows. The messaging gateway plus cron scheduler is the combination few other projects offer, and it changes what an agent is for: not a session you open, but something that works while you are elsewhere.
Start with the TUI and one messaging platform before wiring up remote execution backends. And treat the accumulated memory as data you should inspect — an agent modelling you across sessions is useful exactly in proportion to how right it is.
