harnessforge vs. Hermes / OpenClaw / OpenHarness
The agent-infrastructure space settled into three patterns by May 2026. harnessforge picks a fourth lane that none of them occupies.
The landscape
Hermes (Nous Research)
Standalone agent runtime with the harness built in. "First personal AI agent that ships with the harness already built in." Replaces Claude Code / Cursor for the autonomous-agent use case. Goes from 0 → 57k stars in six weeks (April 2026).
Use Hermes if: you want a personal AI agent that owns its own runtime + memory + skills, and you're not committed to a specific coding IDE.
OpenClaw
Standalone Gateway daemon. Multi-channel personal assistant — WhatsApp, Telegram, Slack, Discord, iMessage. Single-user local-first. Ships its own skills under ~/.openclaw/workspace/skills/.
Use OpenClaw if: you want a single-user personal assistant that talks across messaging platforms.
OpenHarness (HKUDS)
Runtime harness for existing agents. "The model is the agent. The code is the harness." The oh CLI runs your agent through a tool-call loop with permissions, skills, MCP, memory. Compatible with anthropics/skills.
Use OpenHarness if: you want a Python runtime that wraps your agent loop with permissions and skill loading.
OpenAI Agents SDK / Mastra
Library / framework. Primitives for building agents in code.
Use these if: you're writing custom agent code from scratch and want a SDK with batteries.
harnessforge
Provisioning harness. One command — uvx harnessforge init — inspects your repo, picks a blueprint, generates AGENTS.md + SOUL.md + TOOLS.md + MEMORY.md + SKILLS/ + validators + MCP config, then steps away. Your existing coding agent (Claude Code, Cursor, Codex, Gemini CLI, Aider, OpenHarness) reads the output.
Use harnessforge if: you use Claude Code / Cursor / Codex / Gemini CLI / Aider / OpenHarness and want them to start a project with the right ground truth, blueprint, validators, and tools without hand-authoring any of it.
The key distinction vs. OpenHarness
OpenHarness is a runtime harness — you pipe your agent through oh and it provides the loop. harnessforge is a provisioner — we generate files into your repo and exit. Your agent never knows we exist; it just sees the right ground truth on startup.
Different lifecycle, different abstraction:
| OpenHarness | harnessforge | |
|---|---|---|
| Owns the loop | Yes (oh CLI runs your agent) |
No (your IDE keeps the loop) |
| Lives in your shell | Yes (always-on oh process) |
No (one-shot harness init) |
| Files in your repo | Discovers ~/.openharness/skills/, CLAUDE.md, MEMORY.md |
Generates AGENTS.md, SOUL.md, TOOLS.md, MEMORY.md, SKILLS/, plus per-IDE files |
| Visible to your coding agent | Yes (the agent runs inside it) | No (the agent reads the generated files; we're invisible) |
| Compatible with your existing setup | Replaces your loop | Augments your existing tool |
Both are good tools for different jobs. harnessforge doesn't compete with OpenHarness's runtime; it bootstraps a repo that either tool (or any other coding agent) can then use.
The positioning sentence
If you use Hermes or OpenClaw, you don't need harnessforge — those agents ship with their own harness.
If you use Claude Code, Cursor, Codex, Gemini CLI, Aider, OpenHarness, or anything else, you do.