The two biggest open-source personal agents, compared on what their docs actually say: licence, chat apps, skills, safety defaults, and the one-command Ollama setup for each.
"OpenClaw vs Hermes" is the top comparison search for both projects right now, and it's easy to see why. They're the two most-starred open-source personal agents on GitHub, both ship weekly, and Ollama now sets up either one in a single command. They sound alike and they aren't quite. I've compared them only on what each project documents: licence, how they connect to your local model, what they connect to, and how they handle permissions. No "feels smarter", because on a local model the model does the thinking, not the wrapper.
Two personal agents, one local model server, and your chat apps in between. · Illustration generated with Higgsfield
What each one is (in its own words)
OpenClaw vs Hermes Agent, from their docs
OpenClaw
MIT · OpenClaw Foundation
vs
Hermes Agent
MIT · Nous Research
A personal AI assistant with a central gateway to chat apps
Described as
A self-improving agent with skills and memory
MIT
Licence (LICENSE file)
MIT
ollama launch openclaw
One-command Ollama setup
ollama launch hermes
WhatsApp, Telegram, Slack, Discord, iMessage, and more
Chat apps (per Ollama docs)
Telegram, Discord, Slack, WhatsApp, Signal, Email
Gateway bridging chat apps to coding agents; bundled Ollama web search
approvals.mode: smart by default (LLM-assisted risk check)
1,000+ (June & Sept 2026)
Published security advisories (GitHub)
None published
OpenClaw wins 0wins 1 Agent
Descriptions come from Ollama's OpenClaw and Hermes integration pages. Licences are read from each project's LICENSE file: OpenClaw's is MIT, copyright the OpenClaw Foundation. GitHub's summary shows "NOASSERTION" only because the file adds a third-party-notices line. Hermes Agent's is MIT, copyright Nous Research.
Setting each one up on a local model
# Ollama v0.34.4 docs, accessed 25 Sep 2026
ollama launch openclaw # installs OpenClaw if needed, shows a security notice,
# sets your model, starts the gateway + TUI
ollama launch hermes # installs Hermes if needed, points it at http://127.0.0.1:11434/v1
ollama launch hermes-desktop # the Hermes desktop app, same config and memory
OpenClaw: Ollama's page lists the steps as install, security notice ("explains the risks of tool access"), model, onboarding (it "installs the gateway daemon" and enables Ollama web search), then the gateway starts. Chat apps connect with openclaw configure --section channels.
Hermes: install, model, onboarding (points Hermes at http://127.0.0.1:11434/v1), then an optional messaging gateway (hermes gateway setup).
Local models Ollama suggests: for OpenClaw, gemma4 (~16GB VRAM) or qwen3.5 (~11GB). For Hermes, gemma4 (~16GB) or qwen3.6 (~24GB). Those are Ollama's own figures. I size them against real cards, with long-context memory, in running OpenClaw and Hermes on a local model.
Which to pick
You live in iMessage or want the broadest chat-app reach: OpenClaw. iMessage is on its list and not on Hermes's.
You want Signal or email access: Hermes. Both are on its list and not on OpenClaw's (per Ollama's pages).
You want an assistant that builds up skills and remembers across sessions: Hermes. That's its core design, with 70+ skills bundled.
You want a desktop app rather than a terminal: Hermes Desktop.
You mainly want to drive coding agents from your phone: OpenClaw's gateway is built for exactly that.
Security posture: both give you real controls. Hermes turns on an LLM-assisted approval check by default. OpenClaw ships loopback-only, pairing-first defaults and an openclaw security audit command, plus NVIDIA's NemoClaw sandbox as an option. Whichever you pick, connect chat apps only after reading the security page.
What's the difference between OpenClaw and Hermes Agent?
OpenClaw is a personal assistant built around a gateway that bridges chat apps (WhatsApp, Telegram, Slack, Discord, iMessage and more) to AI coding agents. Hermes Agent, from Nous Research, is built around skills (70+ bundled, plus automatic skill creation) and cross-session memory, and has a desktop app. Both are MIT-licensed and both run on local models through Ollama.
Can OpenClaw and Hermes run on a local model?
Yes. ollama launch openclaw and ollama launch hermes configure each one to use a model served by Ollama on your machine (Ollama docs, accessed 25 Sep 2026). Ollama suggests gemma4 (~16GB VRAM), qwen3.5 (~11GB) or qwen3.6 (~24GB) as local options.
Is OpenClaw open source?
Yes. Its LICENSE file is the MIT License, copyright the OpenClaw Foundation. GitHub shows "NOASSERTION" in its summary because the file also references third-party notices, but the licence text itself is MIT.
Is OpenClaw safe to use?
It gives you real safety controls (loopback-only by default, DM pairing, exec approvals, a Docker sandbox and an openclaw security audit command), and it publishes security advisories with patched versions. The core risk of any messaging-connected agent is that people who can message it can try to steer it. See our OpenClaw security guide before connecting chat apps.
Is Hermes Agent free?
Hermes Agent is open source under the MIT License (copyright Nous Research). Running it on a local Ollama model costs nothing beyond your hardware. Cloud models cost whatever their provider charges.
