Hermes Agent vs OpenClaw: which self-hosted agent should you run in 2026?

Run OpenClaw unless you have a specific reason not to. It is the safer default in 2026: it is the larger project by a wide margin, its install page states no architecture restriction for macOS where Hermes Agent's rules Intel out explicitly, and it installs natively on Windows. Pick Hermes Agent when you want its learning loop, which auto-generates skills from work it has already done, or its five execution backends for sandboxing, and when your Mac is Apple Silicon. Both are MIT licensed, both run on your own hardware, and both read the same SKILL.md format, so this is a choice about operations rather than a choice about lock-in.

#What are Hermes Agent and OpenClaw?

Both are self-hosted personal agents. You install one on a machine you control, connect it to the chat apps you already use, point it at a model provider, and it does work for you between conversations.

Hermes Agent is built by Nous Research under the MIT license. Its homepage calls it "The Agent That Grows With You" and leads on persistent memory: it "learns your projects, auto-generates skills, and never forgets how it solved a problem". It reaches you through Telegram, Discord, Slack, WhatsApp, Signal and email, delegates work to "Isolated subagents with their own conversations, terminals, and Python RPC scripts", and names five backends for running those subagents: local, Docker, SSH, Singularity and Modal, with container hardening and namespace isolation.

OpenClaw is maintained by the OpenClaw Foundation, also under MIT. Its repository describes it as "an open-source AI assistant that runs on your own computer and meets you in the channels you already use: Discord, iMessage, Slack, Teams, Telegram, WhatsApp, and 20+ more, plus native apps for macOS, iOS, Android, Windows, and Linux". If you are new to the category, start with what OpenClaw is and come back.

#How do they compare?

Every figure below was read from the vendor's own page on 4 October 2026. Star counts and version numbers in this category move weekly, so treat the date as part of the number.

Two cells deserve a note rather than a tick. "Not covered on the pages cited here" means exactly that: I did not find it on the pages this article cites, which is not the same as the project lacking it. OpenClaw's skills and model-provider pages do not discuss MCP, and the Hermes pages I read do not list local inference runtimes.

#Which hardware rules one of them out?

This is the question that decides the choice most often, and it has a clean answer.

Hermes Agent's installation page supports Linux, macOS on Apple Silicon, Windows, WSL2 and Android through Termux on aarch64. On Macs it is explicit: "The macOS installer is Apple Silicon only. macOS on x86 (Intel) processors is not a supported platform." If your Mac is a 2020 Intel machine, that is the end of the decision.

Worth flagging, because it will send people down the wrong path: the Hermes Agent homepage lists "macOS 12+" among its platforms and says nothing about the CPU. The installation page is the one to trust, both because it is the page describing the installer and because it is specific where the homepage is general. The homepage's wording is not wrong about the OS version, it is just silent on the architecture that matters.

OpenClaw's install page asks for "macOS, Linux, or Windows" and names no architecture restriction. Its runtime requirement is Node rather than Python: "Node 24.16+ or 26.1+", with Node 26 recommended, and the installer provisions Node for you if it is missing, Node 26 on macOS and Node 24 LTS on Linux. Hermes Agent runs on Python instead, and its docs say current first-party installations run on Python 3.14, with a wider range of 3.11 up to but not including 3.15 kept so older installs can run the updater.

Both install with one command:

# Hermes Agent, on Linux, Apple Silicon macOS or WSL2
curl -fsSL https://hermes-agent.nousresearch.com/install.sh | bash
# OpenClaw, on macOS, Linux or WSL2
curl -fsSL https://openclaw.ai/install.sh | bash

On Windows both ship a PowerShell one-liner, and OpenClaw additionally offers npm, pnpm, bun, Docker, Kubernetes and a source build. Hermes Agent has the one target OpenClaw's install page does not mention: a signed APT repository for Termux, so pkg install hermes-agent puts it on an Android phone.

#How big is the gap in project size?

OpenClaw's repository showed 391k stars and 104,998 commits on its main branch when I read it on 4 October 2026. Hermes Agent's showed 251.1k, displayed as "251k" in the badge. That is a real gap and it is worth something concrete: more community skills, more blog posts and forum threads to search at 2am, and more people who have already hit your error message.

Release cadence tells a different story from stars, and it is the more useful signal. OpenClaw's releases page lists v2026.9.8 on 3 October 2026, v2026.8.35 and v2026.8.34 both on 2 October 2026, and v2026.9.7 on 30 September 2026. The 2026.8 line is described as extended-stable, the project's equivalent of LTS, which means you can choose between a fast line and a conservative one. Hermes Agent's most recent tag is v2026.9.24, carrying Hermes Agent v0.21.5, released 24 September 2026, and described as a patch release rolling up the pull requests merged since v0.21.4 for downstream consumers.

Neither cadence is a problem. OpenClaw shipping several times a week with a separate stable line and Hermes Agent cutting a rollup tag every week or two are both normal shapes for a project this size. What they tell you is what to expect from an upgrade: with OpenClaw you pick a line, and with Hermes Agent you take the tag.

#Do skills and MCP servers transfer?

Mostly, and this is the strongest practical argument that the choice is reversible.

Both follow the same open standard. Hermes Agent's skills documentation says skills "follow a progressive disclosure pattern to minimize token usage and are compatible with the agentskills.io open standard". OpenClaw's says "OpenClaw follows the AgentSkills spec" and that "Every skill needs at minimum a name and description in the frontmatter". A folder with a SKILL.md carrying those two fields is a reasonable bet in either. Our guide to skills covers the format itself.

Where they diverge is the paths and the optional blocks:

• Hermes Agent keeps "All skills live in ~/.hermes/skills/" as the source of truth, with ~/.hermes/skill-bundles/ for bundles and project-local skills in .hermes/skills/ and .agents/skills/ inside a git repository. Its optional frontmatter block is metadata.hermes.
• OpenClaw loads from seven precedence-ordered roots, starting with <workspace>/skills, then <workspace>/.agents/skills, then ~/.agents/skills, down to bundled skills and extra directories from config, and it discovers a SKILL.md up to six levels deep inside any of them. Its optional block is metadata.openclaw.

The shared .agents/skills path is the useful detail: a skill in there is in a directory both of them look at.

Portability cuts both ways, so one caution before you move a skill you did not write. OpenClaw's own skills documentation says "Treat third-party skills as untrusted code. Read them before enabling." Work through our skill security audit first. ClawDocx does not vet or endorse third-party skills, and neither project's compatibility with the other says anything about whether a given skill is safe to run.

On MCP, Hermes Agent documents full support and says MCP support ships with the standard install and needs no extra step. Servers go in ~/.hermes/config.yaml under mcp_servers, both stdio subprocesses and HTTP endpoints are supported, there is a curated catalog behind hermes mcp, and discovered tools are registered as mcp_<server_name>_<tool_name>. OpenClaw's skills and model-provider pages do not discuss MCP, so I am not claiming anything about it here in either direction.

Models are the reverse asymmetry. OpenClaw's model-providers page names OpenAI, Anthropic, Google Gemini, Google Vertex, Z.AI, Vercel AI Gateway, OpenRouter, Kilo Gateway, MiniMax, NVIDIA and xAI as bundled providers, custom base-URL providers including Moonshot, Volcano Engine and BytePlus, and local runtimes: "llama.cpp, llmman, LM Studio, Ollama, vLLM, SGLang". If running a model on your own box matters to you, that page is documentation you can act on today, and our free local setup walkthrough follows it.

#What does each one cost?

The software is free in both cases. Both are MIT licensed, both are downloadable without an account, and in both cases the recurring cost is the model.

OpenClaw's model-providers documentation describes no OpenClaw-run paid subscription or hosted service. You bring an API key, or you point it at a local model and pay nothing beyond electricity.

Nous Research sells an optional hosted service alongside Hermes Agent, called Nous Portal, and the Hermes Agent homepage lists four tiers. Free is "$0" with "$0 monthly credits" and free models only. Plus is "$20 Per month" with "$22.00 monthly credits", Super is "$100 Per month" with "$110.00 monthly credits", and Ultra is "$200 Per month" with "$220.00 monthly credits", each paid tier carrying a 10% credit bonus, "200+ Models", hosted tool usage and high rate limits. The homepage also lists Nous Portal as a cloud deployment target, so it is both a model source and somewhere to run the agent. Hermes Desktop itself is free, and its page footer credits Nous Research and the MIT license.

#How hard is it to switch later?

In one direction, easy and well documented. In the other, undocumented.

Hermes Agent ships hermes claw migrate, which its guide describes as importing "your OpenClaw (or legacy Clawdbot/Moldbot) setup into Hermes". What comes across:

• Persona, memory and instructions: SOUL.md, AGENTS.md, MEMORY.md, USER.md
• Skills from four sources: workspace, managed or shared, personal cross-project, and project-level
• Model and provider configuration, including default models and custom providers
• Agent behaviour settings: max turns, verbose mode, reasoning effort, compression
• MCP servers with their command, arguments, environment and tool filtering
• Text-to-speech configuration and assets
• Messaging credentials for Telegram, Discord, Slack, WhatsApp, Signal, Matrix and Mattermost
• Browser settings, approval mode, command allowlists and terminal configuration

What does not come across matters more, because it is the part you will rebuild by hand. Identity files, TOOLS.md, HEARTBEAT.md, cron jobs, plugins, webhooks and memory backend configuration are archived for manual recreation rather than migrated. Timers are dropped by design: "Idle and daily reset timers are not imported: Hermes conversations persist until an explicit /new or /reset." Imported skills land in ~/.hermes/skills/openclaw-imports/.

API keys stay behind unless you ask for them with --migrate-secrets. Run it with --dry-run first; the other documented flags are --preset full or --preset user-data, --overwrite, --no-backup, --source <path>, --workspace-target <path> for where AGENTS.md goes, --skill-conflict <mode> with modes skip (the default), overwrite and rename, and --yes.

Going the other way, OpenClaw's install page documents no import from Hermes Agent, and I found no OpenClaw-side importer on the pages cited here. If you are planning a move in either direction, our migration guide covers the general shape, and security hardening covers what to check before you point a fresh agent at your accounts.

#So which one should you run?

Run OpenClaw if your Mac is Intel, which Hermes Agent does not support, you want native Windows without WSL2, you want documented local models today, you want to pick between a fast release line and an extended-stable one, or you simply want the larger community behind you when something breaks.

Run Hermes Agent if you want the learning loop that auto-generates skills from solved problems, you want subagents isolated across local, Docker, SSH, Singularity or Modal backends, you want an agent on an Android phone through Termux, or you want MCP support that the vendor documents explicitly. Your Mac needs to be Apple Silicon.

Either is fine if you mostly want a self-hosted agent in your chat apps with portable SKILL.md skills. The shared standard and the one-way importer mean the first choice is not the last one. If you are weighing self-hosted against a vendor-run agent instead, Claude Cowork against OpenClaw is the comparison to read next.

One last caution on everything above: this article compares what the two projects document, not how well they work. Nobody at ClawDocx has run both of these side by side on the same tasks, so there is no verdict here on reliability, speed or output quality, and you should distrust any comparison that offers one without saying how it tested.