Hermes Agent vs OpenClaw: I Read Both Codebases Twice

In June I had to pick a runtime for a fleet of agents on a VPS. Hermes Agent or OpenClaw. Most comparisons I found read the README and counted stars, so I cloned both repos and read the source instead. The answer was clean: OpenClaw optimized for reach, Hermes for depth. I picked Hermes and built on it.

On October 1 I read both codebases again. Three of my June conclusions are now wrong. The gap I based my decision on mostly closed, and what is left is a different, more useful comparison.

The short answer

Pick OpenClaw if you want a personal assistant that lives on your devices: native apps for macOS, iOS, Android and Linux, a wake word, on-device voice, and a heartbeat that wakes it on its own. Pick Hermes if you want an agent runtime that lives on a server: seven execution backends including serverless sandboxes, a Mixture-of-Agents mode, and a model lab behind it.

Everything in between, they now both have. Learning loops, file-based memory, a way for the model to script its own tool calls, recursive subagents, MCP in both directions. Five months ago that sentence was false.

Two kinds of powerful

“Which one is more powerful” is the question everyone asks and the one that cannot be answered as asked. There are two senses of powerful in play. Reach: how many places the agent lives, how many integrations it has, how many people build on it. Depth: how much control you have over where it executes, how it manages context, how it learns.

In June, OpenClaw won reach without a contest and Hermes won depth without a contest. In October, OpenClaw still wins reach. Depth is now mostly a tie, with a few places where Hermes is still ahead.

What changed between June and October

My June conclusions, checked again in October

What I concluded in June	What the source says in October
OpenClaw does not create skills on its own	Wrong now. Self-learning is on by default: after substantial work, a detached background review proposes a reusable skill, and the Skill Workshop checks it before applying it.
Only Hermes lets the model script its tools	Wrong now. OpenClaw's Code Mode hides the tool schemas and lets the model write JavaScript against the catalog, in a Node vm or a hardened QuickJS guest.
OpenClaw subagents stop at depth one	Wrong now. Recursive delegation to depth five, five children per agent, eight concurrent.
OpenClaw has far more messaging channels	Parity. About 26 in-repo channels each. Hermes even ships WeChat and Yuanbao natively, which OpenClaw leaves to external plugins.
Hermes has six execution backends and 28 model providers	Seven backends, with Vercel Sandbox added, and 38 provider plugins.

The pattern is plain. OpenClaw spent the summer closing the depth gap. Its 2.0 release in August rebuilt the control UI, moved sessions to SQLite, and shipped the Skill Workshop. Hermes spent the same months widening its surface: Bot Mode for named agents in group chats, cron jobs that keep memory between runs, live steering of running subagents, an MCP command center. Each copied the other’s strengths. Hermes’ own MCP server says in its docstring that it matches OpenClaw’s channel bridge surface.

Hermes Agent vs OpenClaw, side by side

October 2026

Both are model-agnostic, both speak MCP as client and server, both run local models.

	Hermes Agent	OpenClaw
Design center	Agent runtime on a server	Personal assistant on your devices
Behind it	Nous Research, a model lab. MIT.	OpenClaw Foundation, a 501(c)(3). MIT.
GitHub	~250k stars, ~48k open issues and PRs	~391k stars, ~9k open issues and PRs
Surfaces	CLI, terminal UI, desktop app, web dashboard, API	Native macOS, iOS, Android and Linux apps, wake word, on-device voice
Wakes on its own	Cron jobs	Heartbeat every 30 minutes by default, plus cron
Where commands run	Local, Docker, SSH, Singularity, Modal, Daytona, Vercel Sandbox	Host by default; opt-in sandbox in Docker, Podman, SSH, OpenShell or Crabbox
Model scripts its tools	Python over RPC, 50 KB of stdout returned	Code Mode, JavaScript against a hidden catalog
Learning	Background review every 10 turns, weekly curator	Self-learning after substantial work, Skill Workshop
Memory	MEMORY.md, USER.md, full-text session search, 7 external providers	MEMORY.md, USER.md, daily notes, dreaming, session search, Honcho
Mixture of agents	Yes, /moa	No
Skills	SKILL.md, agentskills.io format	SKILL.md, ClawHub registry, 164 extensions

Where OpenClaw still leads

Presence. OpenClaw has native apps for four platforms, a wake-word daemon that runs on-device with zero network traffic, on-device text to speech on the Mac, and a canvas panel. Hermes has a desktop app, a terminal UI and a web dashboard, and no native mobile app. If the agent should live on your phone, this is not close. Windows is the gap on OpenClaw’s side: the gateway runs there, the companion app does not exist yet.

Ambient autonomy. OpenClaw’s heartbeat wakes the agent every 30 minutes by default, one hour when you use Anthropic OAuth, and lets it decide whether something needs doing. Hermes only wakes on cron or when something talks to it. For a personal assistant, the heartbeat is the feature. For a fleet driven by an orchestrator, it is a cost you do not need.

Governance and process. OpenClaw is run by an independent foundation, with OpenAI, NVIDIA, GitHub, Vercel and Red Hat among its sponsors. The README is explicit that OpenAI is a donor, not an owner. Its 2.0 release credited 987 contributors. It carries about a fifth of Hermes’ open issue and pull request count with more stars. That ratio tells you something about how each project absorbs its own growth.

Ecosystem. 164 extensions, 49 bundled skills, and ClawHub for installing and publishing more.

Where Hermes still leads

Where your commands run. Hermes can execute in seven backends, three of them cloud sandboxes: Modal, Daytona and Vercel Sandbox. OpenClaw runs tools on the host unless you turn sandboxing on, and its sandbox options are Docker, Podman, SSH, OpenShell and Crabbox. If you want the agent on one machine and its commands somewhere disposable, Hermes gives you more places to put them.

Mixture of agents. Hermes can send one question to several models and aggregate the answers. I found no equivalent in OpenClaw.

A model lab behind it. Hermes ships a batch runner and a trajectory compressor that write tool-calling trajectories in Nous’ own format. The runtime doubles as a way to train the lab’s models. That gives the project a reason to exist beyond its users, and it shows in how deep the loop primitives go.

Running subagents you can steer. Both delegate recursively. Hermes runs up to 10 children at once by default and, since v0.21, lets you list, steer and stop them while they run.

What each one costs to run

Neither is cheap by default. Both can be made cheap. In both cases the cost is a setting, not a property of the project.

OpenClaw’s bill comes from the heartbeat. Each beat is a full agent turn. The docs put it plainly: with isolatedSession on, a heartbeat drops from about 100k tokens to about 2k to 5k per run, because it stops resending the conversation history. Turn on lightContext too and it skips the workspace bootstrap files.

Hermes’ bill comes from the learning loop. With the same model, the background review replays the whole session history and leans on the prompt cache to keep that affordable. Route the review to a different model and it collapses the history to a digest. An issue arguing the full-history review is too expensive has been open since June with no reply.

The only number that matters is yours. Run the same task through both, on the same model, and read the token counts. I learned that the hard way when a “light” mode in my own setup cost six times more than the full one. The token cost guide has the rest of the method.

Both are still young software

Read the open issues before you put either one on a box you depend on. On OpenClaw: a gateway whose memory grows from 350 MB to 15.5 GB over days until it is OOM-killed (#91588), a SQLite write-ahead log that grows to gigabytes and blocks startup (#143524), an MCP server init timeout that crashes the gateway (#144911), and a September release that turned one user’s stable setup into an eight-hour recovery (#153257).

On Hermes: v0.21.0 shipped session-database regressions that took six pull requests and 44 closed issues to settle, a gateway executor with no timeout freezes the gateway for 120 seconds (#101033), and the dashboard leaks memory until clients get OOM-killed (#80527).

Same advice for both. Pin a version. Upgrade one box first. Keep the logs somewhere you can read them without asking the agent.

The claims the code killed

The method matters more than the verdict, because the verdict expires. In June three marketing claims fell apart once I opened the source. “3,200 MCP tools” turned out to be the size of the MCP ecosystem, not anything in either repo. The star gap described popularity, not what each runtime can do. And “OpenClaw burns more tokens” turned out to be a heartbeat setting. I wrote that up in the Lab as the marketing numbers were the story someone wanted me to tell.

In October the code killed three of my own conclusions. That is the same lesson pointed back at me. A comparison read from the source is only true for the commit you read.

Why I still run Hermes

My case is specific: a VPS, a cost-sensitive budget, and a fleet of work agents that an orchestrator wakes on its own schedule. I do not need a heartbeat, because the orchestrator is the heartbeat. I want commands to execute somewhere disposable. And my adapter, my agents’ instruction bundles and my MCP pooler are all built around Hermes. For that case, Hermes is still the right call.

If I were starting from zero today, the decision would be closer than it was in June. And it does not have to be one or the other. They plug into the same orchestrators, so you can route by role: Hermes for the work that runs on a server, OpenClaw for the assistant that lives on your phone.

Looking for an OpenClaw alternative?

“OpenClaw alternative” alone gets about 1,900 searches a month in the US. Before you switch, check that switching fixes your actual reason.

Before you replace OpenClaw

• ✓Required:
  The heartbeat bill is too high.Turn on isolatedSession and lightContext first. The docs put a tuned heartbeat at 2k to 5k tokens instead of about 100k.
• ✓Required:
  You do not want an agent running commands on your host.Turn sandboxing on, or move to Hermes and execute in Docker, SSH or a serverless sandbox. Either way, give it its own machine and keys you can rotate.
• ✓Required:
  The gateway keeps running out of memory.It is a known open bug. Pin a version that behaves. Hermes has its own memory leak in the dashboard, so moving does not make the class of problem go away.
• ✓Required:
  You want it on a server, driven by an orchestrator or cron.This is the honest reason to pick Hermes. Its design center is the agent that runs when nobody is talking to it.
• ✓Required:
  What you actually want is a coding agent.Neither is the best tool for that. Use Claude Code, OpenCode or pi inside the repo.

How to choose

Pick OpenClaw if

1. 01The agent should live on your phone and laptop
2. 02You want voice, a wake word and native apps
3. 03It should wake on its own and decide what to do
4. 04You value a foundation and a large ecosystem
5. 05You are not on Windows, or only need the gateway there

Pick Hermes if

1. 01The agent should live on a server
2. 02Commands should run in disposable sandboxes
3. 03An orchestrator or cron decides when it wakes
4. 04You want Mixture-of-Agents or live subagent steering
5. 05You will read Python and logs when it breaks

Hermes Agent vs OpenClaw, quick answers

Is Hermes Agent better than OpenClaw?+

Not in general. They have different design centers. OpenClaw is better as a personal assistant on your devices. Hermes is better as a runtime on a server, especially when commands should run in disposable sandboxes.

Since August most core capabilities overlap: both learn skills, keep file-based memory, let the model script its tools and delegate recursively.

Does OpenClaw learn skills like Hermes?+

Yes, since its 2.0 release in August 2026. Self-learning is on by default: after substantial work, a background review proposes a reusable skill, and the Skill Workshop checks proposals before applying them. Hermes runs its review every 10 turns or tool iterations and archives unused skills weekly.

Which one uses fewer tokens?+

It depends on configuration, not on the project. OpenClaw's heartbeat is the main cost and drops from about 100k to about 2k to 5k tokens per run with isolatedSession. Hermes' main extra cost is the background review replaying session history.

Measure both on the same task and the same model before you decide.

Can I run Hermes and OpenClaw together?+

Yes. They are not mutually exclusive, and orchestrators like Paperclip have adapters for both. A sensible split is Hermes for server-side work agents and OpenClaw for the personal assistant on your devices.

Is OpenClaw owned by OpenAI?+

No. OpenClaw is MIT-licensed and run by the OpenClaw Foundation, an independent 501(c)(3). OpenAI is one of its sponsors, and the README states it is a donor, not an owner.