Hermes Agent vs OpenClaw: Which Self-Hosted AI Agent to Run in 2026

As of 28 October 2026, OpenClaw has 393.2k stars on GitHub and Hermes Agent has 257.8k, and both are free, MIT-licensed agents that you host yourself. We read each count from the project's own repository page that day, where OpenClaw also showed 82.7k forks and Hermes 56.5k.

Stars do not tell you which one fits. The short version is this: pick OpenClaw if the agent will serve several people or sit behind many chat apps. Pick Hermes Agent if one person wants an assistant that keeps what it learns, with a paid hosted route for the day self-hosting gets tedious. Neither is an editor copilot, and the section on coding agents says what to use instead.

The short verdict: OpenClaw for groups, Hermes for one person

Both projects do the same basic job. You run a gateway on a machine you control, connect it to Telegram, Slack, Discord, WhatsApp or Signal, point it at a model, and message it like a colleague. The difference is what each project treats as the hard problem.

OpenClaw treats trust as the hard problem. Its docs describe a trusted gateway with untrusted execution, and say policy is enforced by code rather than requested in the prompt. Hermes treats accumulated experience as the hard problem: its homepage says the agent learns your projects, writes its own skills and keeps them between sessions.

That gives three plain recommendations:

• Several users, or a shared channel: OpenClaw, because its docs describe person-level roles and a written trust model.
• One person, one long-running assistant: Hermes Agent, because memory and skill-writing are its headline features.
• A team that wants someone else to host it: Hermes Business, but only after Nous publishes a price (more on that below).

If your situation is not on that list, try Smart Match. You describe the job, and it returns a shortlist.

What does each one cost to run?

The software costs nothing in both cases. Your real bill is the model provider, plus a server if you do not run it on a laptop. Our guide to what a personal AI agent costs each month breaks down that second part in detail.

The difference is that Hermes sells an optional subscription and OpenClaw sells nothing. According to the OpenClaw repository, the OpenClaw Foundation offers no paid tier, hosted service or token. Nous Research, which makes Hermes, sells credits through Nous Portal, and its pricing page lists these plans as of 28 October 2026.

Every paid tier carries a 10% credit bonus, access to 200+ models and hosted tool usage, per the same page. You never have to buy any of it. The Hermes FAQ says you can bring your own provider, and the model bill then goes straight to that provider.

Hermes Business is the pricing gap. Nous describes one central balance for the team, per-member spend caps and an isolated tenant on its own infrastructure, but the Business page we read shows no price. Enterprise, which runs on infrastructure you control with SSO and SLAs, is also unpriced. If a quote matters to your decision, ask Nous Research directly.

OpenClaw has no equivalent, so a team that wants a vendor to carry the server has to look elsewhere. Our roundup of AI agent builders covers the hosted options.

How do setup and platforms compare?

Both are a terminal install away. Here are the two commands as the projects print them. The first needs Node 24.22.1 or 26.12.1, per the OpenClaw repository. The second runs on macOS, Linux and Windows, per the Hermes homepage.

npm install -g openclaw@latest --allow-scripts=openclaw
curl -fsSL https://hermes-agent.nousresearch.com/install.sh | bash

Hermes' source installer pulls in Python 3.14 along with Node.js and ripgrep, according to its README.

Hermes also ships a desktop app as its front door. Its homepage lists a download for macOS 12 or newer and for Windows 10 and 11, with Linux through the terminal. For someone who has never opened a terminal, that matters more than any feature table.

OpenClaw's strength is breadth of where it runs. Its docs carry install guides for Docker, Podman, Kubernetes, a Raspberry Pi, DigitalOcean, Fly.io and Cloudflare Containers, among others. A cheap virtual server from Hetzner is the usual always-on home, and the docs have a Hetzner page too. Hermes lists five terminal sandbox backends on its homepage: local, Docker, SSH, Singularity and Modal.

Run either on a spare machine and you also get an always-on agent that answers while your laptop is shut. NanoClaw is a smaller option in the same family, with source short enough to read, and its record says it wants an always-on machine with 4 GB of memory. If you would rather not self-host at all, AutoClaw is a desktop app that skips the gateway. It comes from Z.ai, which also makes the GLM models it runs on.

How do memory and learning compare?

This is the real fork between them. Hermes says it stores conversations, memories and skills so you can return to your work in a later session. Its homepage adds that the agent auto-generates skills as it solves problems, and that it can spawn isolated subagents with their own terminals.

OpenClaw handles memory through plugins and a skill registry called ClawHub, and its docs make a different promise. They describe tracked session origins and an attributable purge, so you can see where a memory came from and remove it. Hermes, by OpenClaw's reading of the source at commit 6defe7eb6c on 27 August 2026, stores built-in memory as plain entries without source-session lineage or deletion tombstones.

Read that last claim carefully. It comes from the OpenClaw project's own comparison page, which is a competitor's account, and it describes a snapshot two months old. We have not run either memory system ourselves. Treat it as a reason to test deletion on Hermes before you rely on it, not as a verdict.

In practice the difference shows up over months. A Hermes agent that has written skills for your weekly reports should get faster at them. An OpenClaw agent gives you more control over what it keeps. If you want a deeper look at the decision to self-host at all, our self-hosted agent playbook walks through it.

Where does Hermes Agent win?

Hermes wins for a single person who wants a desktop app and an agent that improves with use. Three things stand out.

First, the on-ramp. A downloadable app for Mac and Windows, plus a hosted option on Nous Portal, is a gentler start than a Node install and a config file. Second, scheduling: its homepage describes natural-language schedules for reports, backups and briefings that run unattended and deliver to any connected platform.

Third, money. The optional Portal ladder starts free and gives a paid tier a 10% credit bonus, so a person who does not want to juggle API keys can buy credits in one place. People who prefer to pay a provider directly can do that too, including a model router. To pay no model bill at all, run a local model through Ollama or LM Studio, both of which OpenClaw's docs list as providers.

Hermes also gives you a clean exit. It can import settings, memories and skills from OpenClaw, Claude Code, Codex CLI and Cursor, which lowers the cost of trying it. The cost of that openness is a fast release cadence. Our own record notes breaking changes in recent releases, including platform minimums for the Windows and macOS apps, so pin a version if you want stability.

Where does OpenClaw win?

OpenClaw wins when more than one person depends on the agent. Its gateway owns tool permissions, so a rule such as "this channel may not run shell commands" is enforced by code, not by asking the model nicely. Its docs also describe role ceilings per person and secret handles that keep credentials out of model context.

It also wins on reach. Its own comparison page says its channel catalog includes Nextcloud Talk, Twitch, Urbit and three Zalo paths, while Hermes adds SimpleX, ntfy and conversational email. Counting channels is a weak test, because the two catalogs mix transports and plugin types. Check the exact adapter you need.

Then there is governance. OpenClaw is run by the OpenClaw Foundation, which its repository describes as an independent 501(c)(3). OpenAI is listed as a donor, not an owner. Hermes is made by a company, Nous Research, that also sells a hosted product. Neither arrangement is wrong. They do predict different futures for a free download.

Finally, OpenClaw can sit on top of other agents. Its docs list Codex, the Copilot SDK and Claude Code as runtimes that can run underneath a chat session. If your team already pays for one of those, OpenClaw lets it answer in Slack. Our guide to approval gates for personal agents explains why that wrapper matters.

What happens to your data and your credentials?

Both keep state on your own hardware. The OpenClaw repository says the project itself phones home for nothing except a daily version check, with anonymous statistics opt-in. Your prompts go to whichever model provider you configure, which is true of Hermes too.

The default settings are where they differ, and OpenClaw is blunt about its own. Its repository says tools run on the host for the main session unless you configure sandboxing. Unknown senders on direct-message channels must be paired before they can talk to the agent, which is a sensible default. Our record of OpenClaw marks it as not sandboxed out of the box.

Hermes offers its five sandbox backends, but the same caution applies: a backend you did not turn on protects nothing. Its security policy, as summarised by OpenClaw's docs, treats authorized callers within one adapter as equally trusted, and those callers' tools can include shell access. We could not verify that wording ourselves beyond what OpenClaw's page quotes.

Whichever you pick, the safe setup is the same checklist:

1. Run the agent on a machine that holds nothing you cannot lose.
2. Turn on sandboxing and approvals before you add a second user.
3. Give it its own API keys with a spend cap at the provider.
4. Pin a release and read the changelog before each update.

Who should pick which?

The verdict above is a rule of thumb. Here it is applied to six real situations, each with a named winner.

Two rows deserve a caveat. The agency row assumes you can maintain a server, because OpenClaw offers no hosted version to fall back on. The company row assumes Nous will quote you, since no price is published.

If you are still unsure, run the cheaper experiment first: one weekend, one spare machine, one agent. You lose almost nothing if it fails, and both projects are free to try.

How do they compare with Claude Code, OpenCode and Codex?

They are different tools for a different job. The two agents above live in your chat apps and handle general work. The three below live in a repository and write code. Plenty of people run one of each.

Claude Code's docs say most surfaces need a Claude subscription or a Console account from Anthropic, while the terminal, VS Code and JetBrains also accept third-party providers. The Codex repository, from OpenAI, tells you to sign in with a ChatGPT Plus, Pro, Business, Edu or Enterprise plan, and offers an editor extension for VS Code, Cursor and Windsurf. OpenCode's repository calls itself an open-source AI coding agent, with a read-only plan agent and a full-access build agent, and it showed 222.4k stars and 21,521 commits on 28 October.

So which wins for code? For most developers, the answer is whichever of the three matches the model account you already pay for. Cursor and GitHub Copilot compete for the same seat. Our comparison of Cursor against Claude Code has the details.

Hermes does not replace any of them. HokAI's record lists no editor integration for it. The overlap is the other way round: both Hermes and OpenClaw can import from or wrap these agents, so you can keep your coding tool and add a chat-facing assistant beside it. For lighter self-hosted options in that second role, look at nanobot or Agent Zero.

What is the strongest case against this verdict?

The best objection is that you are leaning on a comparison written by one of the two projects. A fair reader should ask why the OpenClaw docs have a page titled "OpenClaw and Hermes Agent" at all. Its own caveat says the page reflects source at a given commit, not a live adversarial test.

That objection is right about the evidence and wrong about the recommendation. Our verdict leans on three facts that do not depend on the OpenClaw page: OpenClaw has no paid tier, Hermes has a desktop app and a Portal ladder, and the two sets of sandbox defaults come from each project's own text. The memory-lineage claim is the only point that rests on the rival's reading, and we flagged it.

There is a second objection. Fast-moving projects change, and a verdict from October may be stale by December. Both ship quickly. OpenClaw's docs showed release v2026.10.8 on the day we read them, and the Hermes repository listed 60,040 commits.

So test with your own workload. Install one on a spare machine, give it a week of real tasks, and try deleting a memory. That test costs an afternoon. It will tell you more than any table, including ours.

What does switching cost?

Less than you would fear, because both projects wrote an exit for the other. Hermes documents hermes import-agent openclaw, with a --dry-run flag to preview and a --migrate-secrets flag to bring allowlisted API keys across. It reads from ~/.openclaw, and it also imports from Claude Code, Codex and Cursor.

OpenClaw's docs list a "Migrating from Hermes" page under their install section, and a separate one for Claude. We did not walk through that guide, so we cannot say what it does and does not carry over.

What you lose in either direction is the thing each project is proud of. Skills that Hermes wrote about your habits arrive as files, but the learning loop starts again. OpenClaw's role settings and approvals may not carry over to Hermes's slash-command controls, so expect to rebuild them by hand.

Plan a day, not an hour. Run the import in preview mode, keep the old agent running for a week, and move one channel at a time. The migration is easy. The re-teaching is slow.

What would change our pick?

Three shifts would move us.

First, a Hermes Business price. Today a team cannot budget for it, which is the main reason we steer groups to OpenClaw. A published number would open that route to teams that do not want to run a server.

Second, sandbox defaults. If OpenClaw turned isolation on out of the box, or Hermes did, the security section would read differently. Third, a head-to-head test by someone with no stake in either project, covering memory deletion and permission enforcement under real use.

We will recheck both projects when the next major release lands from either side. If you want the wider field before you commit, start with our list of the best personal AI agents and the model leaderboard, because the model you plug in will move your results more than the agent wrapped around it.

Frequently asked questions

Is Hermes Agent or OpenClaw free?

Both programs are free under the MIT license, so you pay only your model provider and any server. Nous sells optional Portal plans from $0 to $200 a month for Hermes. OpenClaw offers no paid tier, hosted service or token, according to its repository.

Which is better for a team?

OpenClaw fits a shared setup better today, because its docs describe person-level roles and gateway-enforced tool policy. Hermes Business offers a pooled balance with per-member spend caps, but Nous publishes no price for it. Ask Nous for a quote before you plan around it.

Can I move from OpenClaw to Hermes Agent?

Yes. Hermes documents an import-agent command for OpenClaw, with a dry-run preview and an option to carry allowlisted API keys. It also imports from Claude Code, Codex CLI and Cursor. Expect to re-teach the agent, because the learning loop starts again.

Are Hermes Agent and OpenClaw safe to run?

Only if you configure them. OpenClaw says tools run on the host for the main session unless you set up sandboxing. Hermes offers five sandbox backends, but a backend you did not enable protects nothing. Use a spare machine, spend-capped keys and a pinned release.

Do they replace Claude Code or Codex?

No. Hermes and OpenClaw are general assistants that live in chat apps, while Claude Code, OpenCode and Codex work inside a code repository. Many people run one of each. OpenClaw can even run Claude Code or Codex underneath a chat session.

Covered in this guide

• Hermes Agent: Hermes Agent is Nous Research's MIT-licensed self-improving AI agent, past 250,000 GitHub stars by October 2026. Self-hosted, model-agnostic, with optional paid Nous Portal plans.
• OpenClaw: OpenClaw is a free, MIT-licensed agent gateway: one self-hosted process connects your chat apps to the AI model you choose, with the OpenClaw Foundation as steward.
• Agent Zero: Agent Zero is a free, MIT-licensed agent framework that runs in Docker with a Linux desktop, browser and projects. 19.5k GitHub stars.
• Anthropic: Anthropic is a Public Benefit Corporation started by 7 ex-OpenAI researchers. It builds the Claude model family and counts Amazon and Google among its largest backers.
• AutoClaw: Z.ai's desktop agent for Windows, macOS and Linux runs 50+ built-in skills on GLM models, with a free tier of daily credits.
• Cursor: Cursor is an AI code editor built on VS Code, widely adopted across large enterprise engineering teams, with Agent Mode, Tab completion, and Cloud Agents.
• GitHub Copilot: GitHub's AI coding assistant with agent mode and MCP support across 6 IDEs, plus a free plan with no credit card required.
• Hetzner: Hetzner sells cloud VPS, dedicated and GPU servers and S3-compatible object storage from Germany, Finland, the US and Singapore, with a large traffic allowance on EU cloud plans.
• LM Studio: Free desktop app that runs open LLMs locally on llama.cpp and MLX, with an OpenAI-compatible local server and an optional $20 cloud plan.
• nanobot: nanobot is a free, MIT-licensed personal AI agent you host yourself, with a browser WebUI, a terminal client and chat apps. 49k GitHub stars.
• NanoClaw: NanoClaw is a free, MIT-licensed personal AI agent for your own Mac or Linux machine, running one Docker container per session behind your chat apps.
• Nous Research: Nous Research, founded 2023, builds the open-source Hermes Agent and is launching Hermes for Businesses for enterprise teams.
• Ollama: MIT-licensed runtime that runs open-weight LLMs on macOS, Windows and Linux with a local API, plus an optional paid cloud for larger models.
• OpenAI: OpenAI builds ChatGPT, the GPT-6 family (Astra, Sol, Luna), Codex and the OpenAI API. It reports 1.2B weekly ChatGPT users and an $852B valuation after its March 2026 round.
• OpenAI Codex: Clones a GitHub repo into a sandbox, writes and tests code, then opens a pull request; OpenAI reported 20 million active users in August 2026.
• OpenCode: MIT-licensed coding agent for terminal, desktop and IDE that works with any model provider, from 208K GitHub stars.
• Z.ai: Z.ai (formerly Zhipu AI) is a Beijing lab behind the GLM model family, founded 2019 and listed on the Hong Kong Stock Exchange (2513) since January 2026.

Sources