What is an agent harness?
An agent harness is the runtime that hosts an AI agent — the loop, the tool-calling layer, the memory store, the platform integrations, and the supply chain that ships it. Three open-source harnesses now dominate the OpenClaw-adjacent ecosystem: OpenClaw (375K⭐, the original), Hermes Agent (171K⭐, fastest-growing), and Pi (57K⭐, the supply-chain-hardened newcomer). This post compares them and tells you which one to pick for which job.
All three are MIT-licensed, actively maintained, model-agnostic. Star counts verified against GitHub on 2026-05-29.
The three harnesses at a glance
| Project | Stars | License | Best for | Maintainer |
|---|---|---|---|---|
| OpenClaw | 375,222 | MIT | Personal AI assistant on your devices, multi-platform messaging | openclaw |
| Hermes Agent | 170,974 | MIT | Self-improving agent that lives on Telegram / Slack / WhatsApp / etc. | Nous Research |
| Pi | 57,298 | MIT | Coding agent CLI + reusable agent runtime + supply-chain hardening | earendil-works (Mario Zechner / badlogic) |
Star count is not a proxy for quality here. OpenClaw's 375K reflects nine months of head start. Hermes's 171K reflects ten weeks of acceleration. Pi's 57K reflects a smaller but more focused scope (coding agent + library).
Which agent harness should I use?
Decision tree:
- I want to talk to an agent from WhatsApp / Telegram / Slack while it works on a VPS → Hermes (it has the strongest gateway story)
- I want a personal AI assistant with voice + canvas + 25+ messaging channels → OpenClaw (the most channels, the most polish)
- I want to embed an agent loop into my own software → Pi (it ships as libraries — pi-ai, pi-agent-core, pi-tui — that other projects can adopt)
- I want a self-improving agent that creates skills from experience → Hermes (it has the closed learning loop)
- I am specifically building a coding agent → Pi (pi-coding-agent is the headliner)
- I value supply-chain hardening (pinned deps, min-release-age, signature audits) → Pi (most explicit posture on this)
- I am migrating from OpenClaw and want compatibility → Hermes (it ships
hermes claw migrate)
OpenClaw — the original platform
OpenClaw is the original OpenClaw-ecosystem harness — a personal AI assistant that runs on your own devices and connects 25+ messaging platforms. 375,222 stars (the highest in the ecosystem by a wide margin). TypeScript. Voice on macOS / iOS / Android. Live Canvas rendering. Gateway control plane.
The defining design choice is breadth of integration. WhatsApp, Telegram, Slack, Discord, Signal, iMessage, Teams, Matrix, Feishu, LINE, WeChat, QQ, and more — all from one process. If your work touches a non-trivial cross-section of these, no other harness gets you there faster.
The trade-off is opinionation. OpenClaw is a platform, not a library — you adopt it, you don't embed it. If your goal is to wire an agent loop into something you already have, the harness becomes a constraint.
Best for: someone who wants a polished personal-AI experience across many channels with minimal setup.
Hermes Agent — the breakout
Hermes Agent is Nous Research's harness — multi-platform gateway, self-improving skill loop, 200+ model support, serverless persistence. 170,974 stars as of 2026-05-29 (up from 82K on 2026-04-30 — a +89K month). MIT-licensed.
Three features dominate the Hermes pitch:
- The multi-platform gateway: one
hermes gatewayprocess exposes the agent to Telegram, Discord, Slack, WhatsApp, Signal, email, and CLI simultaneously, with cross-platform conversation continuity. Voice memo transcription comes built-in. - The closed learning loop: agent autonomously creates skills (
SKILL.mdfiles) from experience, then improves them during use. Compatible with the agentskills.io open standard. - Seven terminal backends — local, Docker, SSH, Singularity, Modal, Daytona, Vercel Sandbox. Modal / Daytona give serverless persistence that costs nearly nothing when idle.
We covered Hermes in depth in the Hermes Agent review — the headline finding is that the architecture is solid; the trade-off is higher token usage than minimalist harnesses.
Best for: anyone who wants an agent that lives on a $5 VPS, listens on Telegram, and gets better over weeks of use.
Pi — the library-first harness
Pi is earendil-works's harness — a mono repo of small, focused packages that ship as libraries other projects can adopt. 57,298 stars. Designed and maintained by Mario Zechner (badlogic, of libgdx fame).
The mono repo contains four canonical packages:
| Package | Role |
|---|---|
pi-ai | Unified multi-provider LLM API (OpenAI, Anthropic, Google, …) |
pi-agent-core | Agent runtime with tool calling and state management |
pi-coding-agent | Interactive coding agent CLI |
pi-tui | Terminal UI library with differential rendering |
The defining design choice is library, not platform. Pi is what you reach for when you want to build an agent into your own software — feynman is built on Pi, Darwin (in our Biomedical AI Agents 2026 roundup) is built on feynman → Pi.
A second distinction: Pi takes supply-chain seriously. Direct dependencies are pinned to exact versions. npmrc sets min-release-age=2 to avoid same-day dependency releases during install resolution. CI runs npm audit signatures. The published CLI ships with a shrinkwrap to pin transitive deps. This is unusual rigor in the AI-agent space.
Best for: builders embedding an agent into their own software, and anyone whose threat model includes supply-chain attacks on dependencies.
How do they compare on the architecture axes?
| Axis | OpenClaw | Hermes | Pi |
|---|---|---|---|
| Distribution shape | Platform | Platform + gateway | Library |
| Primary entry point | OpenClaw CLI / messaging | hermes CLI + hermes gateway | pi-coding-agent CLI + npm packages |
| Messaging integrations | 25+ channels | 6 channels (TG/Discord/Slack/WA/Signal/email) | None built-in (use pi-chat) |
| Voice | macOS / iOS / Android | Telegram voice memo transcription | None |
| Skill / memory loop | Native skill system | Closed learning loop, agentskills.io | Through application code |
| Multi-model | Yes | 200+ providers via OpenRouter | Native multi-provider |
| Serverless persistence | No | Modal / Daytona backends | Through host |
| Supply-chain posture | Standard | Standard | Pinned deps, min-release-age, audit signatures |
| Direct OpenClaw migration | N/A | hermes claw migrate | None |
| Star count (2026-05-29) | 375K | 171K | 57K |
| First commit | 2025-08 (early) | 2026-03 | 2025-08 |
The three harnesses are not really competing for the same job. OpenClaw competes with phone assistants. Hermes competes with messaging-platform bots. Pi competes with agent SDKs.
What about the rest of the field?
Honorable mentions inside the OpenClaw-adjacent space:
- HKUDS/nanobot (41K⭐) — ultra-lightweight harness, 99% fewer lines than OpenClaw. Listed in our core group
- HKUDS/OpenHarness (12K+⭐) — more recent HKUDS effort, generic agent runtime
- kevinrgu/autoagent — listed in core, smaller but actively developed
- karpathy/autoresearch (83K⭐) — adjacent rather than a true harness; more an autonomous-research orchestrator
A reasonable mental model: OpenClaw / Hermes / Pi are the three most-starred general-purpose harnesses with active communities. The HKUDS efforts are tier two. Everything else is either a specialised harness or a fork.
FAQ
Are these three agent harnesses free?
Yes. All three are MIT-licensed. Free for personal, academic, and commercial use including modification. You pay only for the underlying LLM API calls, which you control.
Can I use them together?
Partially. Pi packages can be embedded into Hermes or OpenClaw flows (it ships as libraries). The two platforms (OpenClaw, Hermes) are harder to combine because each owns its own loop.
Which is most secure?
Pi has the most explicit supply-chain posture (pinned deps, signature audits, shrinkwrap). For runtime sandboxing, Hermes ships DM pairing + command approval + container isolation. OpenClaw's posture is standard for a multi-platform messenger.
Which is best for a research lab?
Hermes — it lets you dictate observations from Telegram, transcribe voice memos, schedule cron'd literature scans, and have results delivered back to a messaging app. We cover the lab-specific case in the Hermes Agent review.
Which is best for a software engineering team?
Pi — the coding agent CLI is the headline feature, and the library-first design means you can wire it into your existing tooling.
Which has the strongest community?
Hermes (Discord + dedicated subreddit + 22K+ forks) and OpenClaw (huge installed base across messaging apps) both have strong communities. Pi is smaller but more focused — the contributors are typically library authors.
Will Hermes overtake OpenClaw on stars?
Mathematically possible in 2026 if the +45K-per-month pace holds (it would take ~5 more months). Practically, OpenClaw is also growing — see our May 2026 ecosystem report. We watch this monthly.
Try it
- Personal AI assistant on many platforms → start with OpenClaw
- Self-improving agent on Telegram / Slack / etc. → start with Hermes Agent
- Embedding an agent into your software → start with Pi
- Browse all → Claw4Science core group
If you build a fourth-major-harness candidate (with real adoption signals, not just stars), submit it and we will feature it.
Last updated 2026-05-29. Star counts, package list, and feature claims verified against the three official repos on the publication date.