Hermes Agent vs OpenClaw: Same Licence, Different Custodian

Put Hermes Agent and OpenClaw side by side on capabilities and the comparison comes out close to level, which is the first thing worth saying. Both are MIT licensed. Both let you point them at a model of your choosing and swap it later. Both were last pushed on 23 October 2026, and both describe themselves as running on hardware you control. Neither asks for a subscription. The interesting difference is not a feature one has and the other lacks. It sits one level up, in who is answerable for the project and what each one does with the record it accumulates about you.

The public numbers, on GitHub, read 23 October 2026: NousResearch/hermes-agent carries 256,360 stars and 56,015 forks, created 22 July 2025. openclaw/openclaw carries 392,829 stars and 82,566 forks, created 24 November 2025. Hermes is the older project by roughly four months and holds about two-thirds of the stars. Those counts measure attention, not work done, and the section on measurement comes back to that. Everything between here and there runs both projects through the same six dimensions — where it runs, how you reach it, how the model is chosen, what it remembers, who governs it, and what is measured — and marks each question neither project answers.

The two opening sentences of the two READMEs explain most of what follows. Nous Research calls Hermes "the self-improving AI agent built by Nous Research" and leads with a loop: it creates skills from experience, improves them during use, nudges itself to persist knowledge, and searches its own past conversations. OpenClaw leads with location instead of learning — "an open-source AI assistant that runs on your own computer and meets you in the channels you already use." One project's headline is memory. The other's is address. Hold those two sentences in mind and the six dimensions stop looking like a checklist.

Where it runs

Hermes treats the machine as a deployment target you pick. Its README offers a $5 VPS, a GPU cluster, or serverless infrastructure that costs nearly nothing when idle, and it is explicit that the agent need not live on your laptop: you can talk to it from Telegram while it works on a cloud VM. Six terminal backends ship built in — local, Docker, SSH, Modal, Daytona and Vercel Sandbox — with plugin backends such as Singularity and Apptainer on top. Two of those, Daytona and Modal, give serverless persistence: the environment hibernates when idle and wakes on demand.

OpenClaw has one deployment unit and varies it by configuration. "One Gateway runs it as a personal assistant on a laptop or as a shared team deployment; configuration is the only difference," its README says, and the Gateway is described as the local control plane for sessions, tools, events and channel connections. The README is equally clear about what does not exist: "no paid tier, hosted service, or token." Whether Nous Research runs a hosted version of Hermes: not published in the material read for this comparison. The contrast is between a project that spreads execution across optional remote backends and one that keeps a single control plane and changes only its settings.

How you reach it

Hermes has a first-class operator console. Its README describes a "full terminal interface" with multiline editing, slash-command autocomplete, conversation history, interrupt-and-redirect, and streaming tool output, and the same gateway process also serves Telegram, Discord, Slack, WhatsApp, Signal and the command line. The messaging channels are a second front door bolted onto a terminal product, which is why the terminal still reads as the native surface.

OpenClaw inverts that emphasis. Channels are the product: Discord, iMessage, Slack, Teams, Telegram, WhatsApp and, in the README's words, "20+ more", alongside native apps for macOS, iOS, Android, Windows and Linux. A Control UI, the CLI and a TUI connect to the same Gateway, but they arrive second in the description. The practical difference for a team is small and real: address one in the terminal and chat from the phone; address the other from whichever window your colleagues already have open. Round-trip latency for either, and the number of channels each actually supports in practice: not published.

How the model is chosen

Both projects make the same promise in different words. Hermes says use any model you want and switch with the command "hermes model", with no code changes and no lock-in; in its desktop app, subagents can each be given their own provider, model and limits. OpenClaw says models and agent harnesses — it names Claude and Codex, and local models — are plugins you can swap without changing anything else, and its onboarding verifies model access before it finishes. Neither ships a mandatory first-party model, and neither publishes a recommended model or a default the project endorses.

Model-agnostic is a claim about architecture, not about where the tokens come from. You still need an endpoint, and for either agent that endpoint can be any provider you configure, including ours: one OpenAI-compatible key reaches more than 200 models with automatic failover and 0% markup, the provider's list price passed straight through. Claude Opus 5.5 sits on it at $4.00 per million input tokens and $20.00 per million output tokens, with a 1M-token context window and cache reads at $0.20 per million. The scope is worth stating precisely: we host models, not Anthropic's apps or anyone's assistant, and whichever of these two agents you run remains yours to install and configure.

What it remembers

Hermes is the more explicit of the two, because memory is where it starts. The README lists agent-curated memory with periodic nudges, autonomous skill creation after complex tasks, skills that improve during use, FTS5 session search with LLM summarisation for cross-session recall, and Honcho dialectic user modelling, and adds compatibility with the agentskills.io open standard. Nous Research's own phrasing is that it "builds a deepening model of who you are across sessions."

OpenClaw's published statement is about custody rather than curation: "State, memory, and credentials live on your hardware." Whether the Gateway summarises old sessions, curates that state, or writes its own skills: not published in the README read here. What is published is the boundary. Your prompts go to the model provider and the chat platforms you configure, plus any diagnostics export you enable yourself; by default the software checks for a new version once a day, anonymous feature statistics are opt-in, and "update.checkOnStart: false" disables both. Hermes' telemetry defaults: not published in the source read here.

The asymmetry has a consequence worth thinking through before you install either. A system that deepens its model of you gets more useful the longer it runs, and it also holds a richer record of the person using it. A system that keeps state on your hardware makes a narrower promise and a more checkable one. Those are different bets, and they are not about features at all.

Who governs it

Hermes is built by Nous Research and says so in the first line of its own README — a company's project carrying a company's name. Beyond that the published record is thin: MIT licence, no foundation, and no governance document among the material read here. Whether there is a commercial tier, a hosted service or a roadmap commitment from Nous Research: not published in that material.

OpenClaw answers the same question in the first screen of its README. It is stewarded by the OpenClaw Foundation, which the README describes as "an independent 501(c)(3)", and it states that there is no paid tier, hosted service, or token. That is a governance claim rather than a licensing one, and it matters because the licence is identical on both sides. MIT gives you the same rights whichever project you clone. What differs is who can change direction and what incentive sits behind the change.

The honest limit of that distinction: a foundation structure is not a guarantee of anything, and neither project publishes a funding disclosure, a succession plan or a roadmap commitment in the text read for this piece. What you can say is that one project has written down who is accountable and what it will not sell, and the other has not published that. When the question is "what happens to this in three years," one of them has a paragraph you can quote and the other has a silence.

What is measured

Neither project publishes a benchmark, a task-success rate, a latency figure or an audited head-to-head against the other. For a comparison piece that is the most important sentence in it, because it means there is no number on which to rank these two, and any ranking would have to be invented. The quantitative record that does exist is GitHub's: 256,360 stars and 56,015 forks against 392,829 and 82,566, read 23 October 2026. Stars count attention; forks count people who thought they might change something. Neither counts completed work.

What both publish instead is scope. Hermes lists six terminal backends, a cron scheduler for scheduled automations, subagents for parallel work, batch trajectory generation for training the next generation of tool-calling models, and its compatibility line with the agentskills.io standard. OpenClaw lists channels, native apps on five platforms, a Gateway that becomes a team deployment by configuration, and a security guide that opens with "treat inbound messages as untrusted input." Those are claims about surface area, which is the easiest thing for a README to describe and the hardest thing for a reader to compare. Token spend per completed task, time to first useful action, and how often either agent needs a human to unstick it: not published for either.

How to choose without a scoreboard

Since no published measurement separates them, the decision has to come from the shape of your work rather than from a score.

• Long-running and cloud-shaped. A nightly audit, a research batch, something you want to check from your phone while it runs on a VM — Hermes is built around that shape, with serverless backends that hibernate and a scheduler that fires in natural language.

• Conversational and multi-party. A shared gateway that answers in the chat app your team already uses, on a laptop or across a group — OpenClaw is built around that shape, and it pairs unknown senders by default rather than trusting them.

• Accountability first. If the deciding factor is who is answerable when the project changes direction, the comparison is a 501(c)(3) with no paid tier against a company's product, and it should be read as exactly that rather than as a feature gap.

• Memory first. Ask what happens to the record after a year of use. Hermes publishes an answer about curation; OpenClaw publishes a location. Which of those you want is a policy decision, not a technical one.

Both projects are MIT, both are model-agnostic, and both are actively pushed, so the licence will not settle this for you. What separates them is custody in the ordinary sense — who holds the code, who holds the record, and who has written down what they will do with both. Pick the custodian whose answer you can live with, then let the six dimensions confirm it.