OpenClaw vs Hermes Agent
OpenClaw and Hermes Agent are both self-hosted, MIT-licensed personal agents that sit behind chat apps and can drive coding harnesses. OpenClaw is a TypeScript Gateway with isolated multi-agent routing, sub-agents and a JavaScript Swarm mode. Hermes is Nous Research's Python agent with delegated subagents, a Kanban board for multi-process work and self-written skills.
Facts side by side
| Fact | OpenClaw | Hermes Agent |
|---|---|---|
| Type | Harness | Harness |
| Languages / SDKs | TypeScript | Python |
| License | unknown | MIT |
| Pricing model | Open source, free | Open core |
| Orchestration pattern | Supervisor | Supervisor |
| GitHub stars | 390,844 (as of 2026-09-30) | 250,238 (as of 2026-09-30) |
| GitHub forks | 82,207 | 53,433 |
| Last push | 2026-09-30 | 2026-09-30 |
| Latest release | v2026.9.7 | v2026.9.24 |
| Repository | openclaw/openclaw | NousResearch/hermes-agent |
| Website | openclaw.ai | hermes-agent.nousresearch.com |
| Documentation | docs.openclaw.ai | hermes-agent.nousresearch.com |
| Last verified | 2026-09-30 | 2026-09-30 |
| MCP support | Yes (checked 2026-09-30) | Yes (checked 2026-09-30) |
| A2A support | Yes (checked 2026-09-30) | Yes (checked 2026-09-30) |
| AG-UI support | Unknown (checked 2026-09-30) | Unknown (checked 2026-09-30) |
| Install verified | 2026-09-30 | Not yet |
Choose OpenClaw if
- You want several isolated agents on one Gateway, each with its own workspace and credentials, routed by chat account.
- You want large fan-out from a script: Swarm launches many collector sub-agents with bounded concurrency.
- You want a project that states it has no paid tier or hosted service; OpenClaw's README says state and credentials stay on your hardware.
Choose Hermes Agent if
- You prefer Python and want tools to run on a choice of terminal backends (local, Docker, SSH, Modal, Daytona).
- You want long-lived collaboration between separate agent processes through a shared Kanban board with handoffs and review.
- You want the agent to write and reuse its own skills, and are fine with the optional paid Nous Portal for managed tools.
Migration notes
Both keep memory as Markdown files, so notes carry over by copying content: OpenClaw uses MEMORY.md, USER.md and dated daily notes in the agent workspace; Hermes uses MEMORY.md and USER.md per profile, capped at 2,200 and 1,375 characters, so long OpenClaw memories must be trimmed. Delegation maps closely: OpenClaw's sessions_spawn and Hermes' delegate_task both start a child with its own context that reports back. Nesting defaults differ: OpenClaw allows depth 5, Hermes children are leaf workers unless delegation.max_spawn_depth is raised. Approval models differ: OpenClaw requires policy, allowlist and an optional prompt to agree on host commands (sandboxing is off by default), while Hermes defaults to a smart mode in which an auxiliary LLM auto-approves low-risk commands. Review these settings before moving an always-on agent. OpenClaw's docs include its own comparison page with Hermes; treat it as the OpenClaw project's view.
FAQ
Are OpenClaw and Hermes Agent free?
Both are MIT-licensed and self-hosted. OpenClaw's README states there is no paid tier or hosted service. Hermes works with any model provider, and Nous Research sells an optional Nous Portal subscription whose Tool Gateway bundles managed web search, image generation, TTS and browser tools.
Can they drive Claude Code or Codex?
Yes, both. OpenClaw runs Claude Code, Codex, Gemini CLI and OpenCode through ACP and has an official Codex plugin. Hermes has an optional Codex app-server runtime and ships skills that run Claude Code and OpenCode through its terminal tool. Evidence links are on each tool page.
Which protocols do they support?
Both document MCP as client and server and ship an A2A plugin or channel. Neither documents AG-UI.
Which one is safer to leave running?
Neither is safe by default without configuration. OpenClaw runs main-session tools on the host until sandboxing is set up, and its plugins run in-process. Hermes' default smart approval mode lets a model auto-approve low-risk commands, and --yolo disables prompts.