Claw profile

Hermes Agent

Name: Hermes Agent
Author: Nous Research

An open-source agent that writes its own skills, keeps them in a searchable store, and reuses them the next time it runs. Here is the mechanism, and where it parts ways with OpenClaw.

Built by: Nous Research
Released: February 2026
License: MIT
Language: Python 3.11+
Runs on: macOS 12+ · Windows · Linux
Floor cost: ~$5/mo VPS

An agent with a learning loop, not a chat wrapper

Hermes Agent is a long-running process you own. It reaches you through messaging apps, keeps state between sessions, and — the part that distinguishes it — converts completed work into reusable procedures without being asked.

Nous Research describes it as “the only agent with a built-in learning loop — it creates skills from experience, improves them during use, nudges itself to persist knowledge, searches its own past conversations, and builds a deepening model of who you are across sessions.”

Install on Linux, macOS, WSL2, or Termux

$curl -fsSL https://hermes-agent.nousresearch.com/install.sh | bash

Install on Windows (PowerShell)

$iex (irm https://hermes-agent.nousresearch.com/install.ps1)

Run the configuration wizard

$hermes setup

Switch model provider — no code changes, no lock-in

$hermes model

Requirements are Python 3.11+, Node.js, ripgrep, and ffmpeg. Git is required and bundled as MinGit on Windows when absent.

The agent loop is the core, not the gateway

Hermes puts the AIAgent loop at the center as a synchronous orchestration engine. Everything else — the messaging gateway, the cron scheduler, the CLI — is an edge that feeds it. That single choice is what the OpenClaw comparison below turns on.

Every edge converges on one loop. The highlighted path is the write into the store — the edge that makes the learning loop below possible.

The repository mirrors this layout: /agent holds the loop, /gateway the platform router, /skills the procedural memory, /tools the integrations, /providers the model adapters, and /hermes_cli the terminal interface.

Where skills come from

This is the claim the product rests on, so it is worth being precise. Hermes does not merely store transcripts. When a task completes, the agent abstracts the procedure into a skill file, keeps it in a store it can search, loads it the next time a similar task appears, and edits it in place when it turns out to be wrong.

Dashed lines are reads and writes against the store; every stage touches it, which is why recall works across sessions rather than only within one. Skills follow the open agentskills.io format.

Memory is layered rather than monolithic, which matters because the layers carry very different context costs.

Only the top band is a standing context cost. The other two are fetched, which is what keeps a long-lived agent from growing an unbounded prompt.

Personality is separate again: a global SOUL.md carries identity, while SKILL.md files carry procedure. Hermes deliberately splits the two — a distinction that becomes the awkward part of migrating from OpenClaw, where both live in one file.

Hermes vs OpenClaw: what actually differs

Both are open source, both run locally, both talk to chat platforms. The structural difference is where control sits — and that one choice cascades into how skills, memory, and security work in each.

The highlighted box is the component that owns control in each system. It sits above the work in OpenClaw and below it in Hermes — that inversion is the whole difference.

Dimension	Hermes Agent	OpenClaw
Control point	AIAgent loop as a synchronous orchestration engine	Central gateway as the control plane for sessions and routing
Skills	Generated from experience, self-improved during use	Human-authored tool and workflow instructions, loaded per scope
Memory	Core notes + SQLite/FTS5 history + optional Honcho user model	Markdown files as canonical truth, hybrid retrieval over files and transcripts
Security posture	Five layers: authorization, dangerous-command approval, container isolation, MCP credential filtering, context-file scanning	Operator-driven — control and inspectability, hardening is the operator’s job
Positioned for	Developers running a long-lived worker that learns and schedules	A broader audience wanting a rich personal-assistant control plane

Migrating off OpenClaw

Hermes ships this as a first-party path. The CLI imports the persona file, memories, user-created skills, messaging settings, and API keys. Run the dry run first and read it — the known friction points are a large asset copy, a conflicting SOUL.md, and overflowed memory entries.

Preview exactly what would move

$hermes claw migrate --dry-run

Run the full interactive migration

$hermes claw migrate

Against the nine Claw signals

Scored against our inclusion standard, with the evidence rather than just the verdict. Hermes clears all nine, which is unusual — most catalogued projects clear four or five.

Persistent / long-runningDesigned to run 24/7; state survives restarts

Autonomous executionPlans and executes multi-step tasks unprompted

Tool use60+ built-in tools plus MCP servers

Memory / stateThree-layer memory with FTS5 cross-session recall

Scheduled tasks / heartbeatCron plus natural-language scheduling

Browser / shell / computer accessBrowser automation and six terminal backends

Proactive communicationSends scheduled reports and briefings unprompted

Multi-step goalsSubagent delegation for parallel workstreams

Skills / plugins / MCPagentskills.io format, MCP servers, Bot Mode

What it costs to keep one alive

Path	Setup	Monthly	Notes
Desktop app	Installer	$0 + tokens	macOS 12+, Windows 10/11, Linux. Not always-on.
VPS, native	One-line script	~$5–7	1 vCPU / 2 GB floor. 20–30 min for a competent Linux user.
VPS, Docker	Container	~$5–7	Recommended for production: isolation and cleaner profiles.
Serverless	Modal / Daytona	near $0 idle	Hibernates when idle; pays only while working.
Local GPU	RTX / DGX Spark	hardware only	Local inference, no per-token cost.

LLM spend runs roughly $2–15 a month on top, depending on model choice. Providers are swappable at runtime, so the deployment decision and the model decision stay independent.

Questions people ask

Is Hermes Agent free?

The agent itself is open source under the MIT license and free to self-host. You pay for model inference — roughly $2–15 a month depending on model choice — plus a server if you want it always on. Nous Research also sells paid tiers that bundle monthly credits and access to 300+ models.

What makes Hermes different from other agents?

The closed learning loop. When a task completes, Hermes abstracts the procedure into a skill file, stores it in a searchable SQLite database, retrieves it in later sessions, and edits it in place when it proves wrong. Most agents load human-authored skills; Hermes writes and revises its own.

Can I migrate an existing OpenClaw setup?

Yes. Hermes ships a first-party migration path — run hermes claw migrate --dry-run to preview, then hermes claw migrate to import your persona file, memories, user-created skills, messaging settings, and API keys. Because Hermes splits identity from procedure, a single OpenClaw SOUL.md has to be pulled apart into a personality file plus skill files.

What does it take to run Hermes Agent 24/7?

A 1 vCPU / 2 GB VPS is the practical floor, around $5–7 a month. Docker is recommended over a native install for production because it isolates the agent and keeps profiles separable. Serverless backends such as Modal and Daytona hibernate when idle and cost close to nothing between tasks.

Which models can Hermes Agent use?

Nous Portal, OpenRouter, OpenAI, or any custom endpoint. Providers are swapped at runtime with hermes model — no code changes — so the deployment decision and the model decision stay independent.

Resources

Sorted by how much weight each source can carry. Primary sources are publisher-controlled and safe to cite as fact; the rest are useful for deployment detail but vary in freshness.