The two rising self-hosted agent harnesses
Hermes and OpenClaw are two open-source, self-hostable AI agent harnesses attracting growing attention in 2026. Both let you run an agent that talks to humans across messaging channels, calls tools, and maintains state — but they differ in language, philosophy, and ecosystem.
| Dimension | Hermes | OpenClaw |
|---|---|---|
| Repo | NousResearch/hermes-agent | openclaw/openclaw |
| Version (2026-07-31) | 0.16.0 | 2026.7.1-2 |
| Language | Python | Node.js / TypeScript |
| License | MIT | MIT |
| One-liner | The self-improving AI agent | Multi-channel AI gateway |
| Skills | skills system — creates + refines skills from experience | skills marketplace (community) |
| MCP | built-in MCP serve | MCP client + serve |
| Channels | ~18 adapters across ~30 modules: Telegram, WeChat, Feishu, DingTalk, WeCom, QQ, WhatsApp, Slack, Matrix, Signal, SMS, iMessage (BlueBubbles), Tencent Yuanbao, MS Graph webhook, + | 24+ channels: Telegram, WhatsApp, Discord, Slack, Signal, iMessage, Feishu, Matrix, MS Teams, Mattermost, Line, Zalo, Nostr, IRC, Twitch, Synology Chat, + |
| Config | ~/.hermes/config.yaml + .env | ~/.openclaw/openclaw.json + .env |
| Self-host | systemd / foreground process | systemd user service (openclaw-gateway) |
What they share
Both are open-source, self-hosted, MIT-licensed, and built around the same core loop: a human messages a channel → the agent reasons, calls tools/skills, and replies. Both speak MCP, both expose Telegram/Slack/Matrix/Signal/SMS/Feishu, and both keep config + secrets local (.env). Neither locks you into a cloud.
Where they differ
Hermes (Python, Nous Research) leans research + self-improvement. Its signature is a skills system that creates and refines skills from experience — the agent learns as it runs. Its platform coverage is strongest in the China ecosystem (WeChat, Feishu, DingTalk, WeCom, QQ, Tencent Yuanbao), which fits a Nous Research project.
OpenClaw (Node/TS) leans gateway + breadth. Its signature is a mature skills marketplace plus the widest channel coverage (24+ adapters including iMessage, Nostr, MS Teams, Mattermost, Line, Zalo, Twitch). It reads more as “multi-channel gateway” than “self-improving agent.”
Choosing
- You want a self-improving agent that learns skills and you live in the WeChat/Feishu ecosystem → Hermes.
- You want the broadest channel coverage + a skills marketplace and you prefer Node/TS → OpenClaw.
- You want both worlds → they coexist (migration guide coming).
A note on this comparison
Every figure above was verified against live sources on 2026-07-31 (Hermes hermes-agent v0.16.0; OpenClaw v2026.7.1-2 — both repos + running instances). Both projects move fast — treat the version column as a snapshot, and re-check before deploying.