Short answer
Hermes Agent and OpenClaw are both open-source, MIT-licensed AI agents that you run on your own machine or server and talk to from a terminal or a chat app. Hermes Agent, from Nous Research, is built around a learning loop: it writes and improves its own skills, searches its past sessions, and can run its commands in seven different terminal backends. OpenClaw, from the OpenClaw Foundation, is built around one Gateway that routes your chat accounts to one or more isolated agents, with a control UI and companion apps.
Choose by how you want the agent to run and who will maintain it, not by feature count. Both projects change quickly, so check the current docs before you decide.
Sources for this comparison
This guide draws on four public sources, checked in October and November 2026:
- The Hermes Agent documentation, written by Nous Research.
- The OpenClaw documentation, including its own OpenClaw and Hermes Agent comparison, reviewed by the OpenClaw team on August 27, 2026. It is written by one of the two projects, so read it with that in mind.
- Dexio's Hermes Agent vs OpenClaw guide, updated October 5, 2026.
- Our own reading of both projects' docs. This is not a hands-on benchmark of either tool.
Hermes Agent vs OpenClaw at a glance
| Hermes Agent | OpenClaw | |
|---|---|---|
| Made by | Nous Research, venture-backed | OpenClaw Foundation, a 501(c)(3) nonprofit |
| License | MIT | MIT |
| Language | Python | TypeScript on Node.js |
| Installs on | macOS, Windows, Linux, WSL2, Docker | macOS, Linux, WSL2, Windows |
| Where tools run | Seven backends: local, Docker, SSH, Daytona, Singularity, Modal, Vercel Sandbox | The host, or a sandbox you configure. Sandboxing is off by default |
| Chat apps | One gateway process for 20+ platforms | One Gateway for 20+ channels, plus a control UI and companion apps |
| Several agents | Profiles, each a separate Hermes home | Isolated agents in one Gateway, routed by bindings |
| Memory | Capped MEMORY.md and USER.md, plus full-text session search |
Workspace Markdown, daily notes, semantic memory search |
| Skills | agentskills.io standard, written by the agent, Skills Hub | SKILL.md folders, ClawHub registry, Skill Workshop review |
| Scheduling | Built-in cron, managed by the agent | Automations |
| Paid offering | Optional Nous Portal plans | None |
Where commands run
This is the biggest practical difference. Hermes separates the agent from where its commands execute. You can run the agent on one machine and send its terminal work to a Docker container, a remote server over SSH, or a serverless sandbox such as Modal or Daytona. The Hermes security docs describe container isolation as one of their security layers.
OpenClaw runs tools on the host for the main session unless you configure a sandbox. OpenClaw's own comparison lists "sandboxing off by default" in its trust model. That is a reasonable default for a personal assistant on your own machine, and a setting to change before a team uses it.
Memory
Both projects keep memory as plain files and load a MEMORY.md and USER.md at the start of a session.
- Hermes caps those files at 2,200 and 1,375 characters, according to the memory docs. Everything older lives in a SQLite session store that the agent searches with full-text search. The small files keep the prompt short, and the search finds older detail when needed.
- OpenClaw keeps memory in the agent's workspace, adds daily notes, finds notes with a semantic search, and runs a background "dreaming" sweep that promotes items into
MEMORY.md, according to Dexio's comparison.
OpenClaw's comparison adds one point about Hermes: deleted memory entries leave no record of the session they came from, so a later session can add the same fact again. It labels this a limitation inferred from the design, not a reproduced failure. Hermes lets you require approval for memory writes if you want to review them.
Skills
Hermes skills follow the agentskills.io standard. The agent creates a skill after a complex task, improves it as it uses it, and you can install more from the Skills Hub, which scans each skill before installing. The skills docs cover the commands.
OpenClaw skills are folders with a SKILL.md file. ClawHub is its public registry, and its Skill Workshop lets you review and approve skills the agent drafts before they take effect. If you want a review step before an agent-written procedure goes live, OpenClaw has a dedicated tool for it.
Teams and several agents
Hermes runs separate agents as profiles. Each profile has its own configuration, memory, skills, sessions, and gateway, and the docs say never to point two agents at one profile. Within a profile, OpenClaw's comparison notes that Hermes treats every authorized user on a platform as equally trusted.
OpenClaw runs several isolated agents inside one Gateway and uses bindings to route a Slack workspace or WhatsApp number to the right agent. Its comparison lists person-level role limits as a feature. For a company with different permission levels, this difference matters, and you should test it with your own setup.
Security record
Both projects publish security policies, and neither is safe by default for company data.
OpenClaw's comparison says OpenClaw has a public record of repository security advisories, while Hermes had none in its repository on the review date. It also notes third-party CVE records for Hermes, such as CVE-2026-14625. It states that advisory counts are not a safety score: a project with more advisories may simply report more of them. We agree. Read the current advisories for both before you connect either to company systems.
On the Hermes side, the security docs describe command approval with a hardline blocklist, user allowlists, and a default that denies risky commands in scheduled jobs.
Cost and governance
Both agents are free software, and you pay for models and hosting either way.
- OpenClaw is run by an independent nonprofit funded by donations. Its docs say it has no paid tier, hosted service, or token.
- Hermes Agent is run by Nous Research, which OpenClaw's comparison describes as venture-backed with a Paradigm-led Series A. Nous sells an optional Nous Portal subscription for models and hosted tools. A September 2026 pricing review by AI//COST lists a free tier plus paid plans at $20, $100, and $200 a month. Hermes also works with your own API keys or a local model.
Migrating from OpenClaw to Hermes
Hermes includes an importer. According to the migration guide, you install the importer plugin and run it:
hermes plugins install agent-importers
hermes import-agent openclaw --dry-run
hermes import-agent openclaw
The dry run shows what will move. The import brings over your persona file, MEMORY.md, USER.md, and skills, which land in a separate openclaw-imports folder, and it creates a backup first. API keys move only if you add --migrate-secrets. Check imported skills before you rely on them, because the two projects load skills differently.
Which should you choose
- Choose Hermes Agent if you want commands to run in Docker, over SSH, or in a serverless sandbox, want an agent that writes its own skills, or want each agent fully separated in its own profile.
- Choose OpenClaw if you want one Gateway that routes several chat accounts to several agents, want the control UI and companion apps, want a review step for agent-written skills, or prefer a project run by a nonprofit with no paid tier.
For a company team, the deciding questions are the same for both: where the agent runs, which channels and tools it may use, which actions need approval, and who maintains it. If you are leaning toward Hermes, start with what Hermes Agent is and the installation guide. If you want help scoping a team agent in Slack, see our Slack AI agents service.
