Skip to content
Back to search

metaharness

πŸ› οΈ The meta-harness for AI agents β€” scaffold your own focused, branded agent harness with its own npx CLI, MCP server, memory, learning loop, and witness-signed releases. Works with Claude Code, Codex, pi.dev, Hermes, OpenClaw, and RVM (hardware-isolated sandbox).

Safety scan

A static text scan, not a guarantee. Select the badge to see what was scanned and every finding.

Checked today

Check it deeper on your machine

It has changed since its scan on 11 Nov 2026. Its scripts were only partly scanned.

As a Claude Code plugin, install it and then ask Claude Code:

/plugin marketplace add keithmackay/mackayi/plugin install scanskillsafety@mackayiDo a deep scan of https://github.com/ruvnet/metaharness with scanskillsafety

Or from a terminal, clone it and run it:

git clone https://github.com/keithmackay/scanskillsafety-skill ~/.claude/skills/scanskillsafetynode ~/.claude/skills/scanskillsafety/cli.cjs --deep 'https://github.com/ruvnet/metaharness'

It checks the exact commit: every file, its packages against OSV (api.osv.dev), how it compares with this page's scan (findsafeskills.com receives the repository's address), and, if you agree, Claude reading every file with your own Claude subscription. It never runs the code.

Badge for your README
findsafeskills safety rating

Markdown

[![findsafeskills safety rating](/api/listings/ruvnet--metaharness/badge.svg)](/listings/ruvnet--metaharness)

HTML

<a href="/listings/ruvnet--metaharness"><img src="/api/listings/ruvnet--metaharness/badge.svg" alt="findsafeskills safety rating" /></a>

Install

i

What was scanned

The files the safety scan read, as written by the repository's author and shown as plain text. Lines with a finding are marked.

.claude-plugin/marketplace.json Β· Manifest Β· 15 lines

View this file on GitHub (read at the repository's current version: the scan did not record which commit it read)

This file has changed since it was scanned, so this is not exactly the text that was scanned, and no line is marked. The findings are listed with the safety scan above.

  1. {
  2. "name": "metaharness",
  3. "description": "MetaHarness Marketplace: mint a custom, branded AI agent harness from any repo β€” runs on Claude Code, OpenAI Codex, pi.dev, Hermes, OpenClaw, RVM, Copilot, OpenCode, or GitHub Actions. WASM kernel, witness-signed releases, npx CLI.",
  4. "owner": {
  5. "name": "ruvnet",
  6. "url": "https://github.com/ruvnet"
  7. },
  8. "plugins": [
  9. {
  10. "name": "metaharness",
  11. "source": "./",
  12. "description": "Scaffold your own focused AI agent harness with its own npx CLI, MCP wiring, and host adapter. 14 skills (create/validate/publish/diag/score/threat-model/…), 9 host integrations, Ed25519 witness provenance, and GCP Secret Manager validation. Install, then run /create-harness or `npx metaharness <name>`."
  13. }
  14. ]
  15. }
README.md Β· README Β· 534 lines

View this file on GitHub (read at the repository's current version: the scan did not record which commit it read)

This file has changed since it was scanned, so this is not exactly the text that was scanned, and no line is marked. The findings are listed with the safety scan above.

  1. <div align="center">
  2. # MetaHarness
  3. <!-- readme-motion: hero -->
  4. <a href="docs/visual-walkthrough.md"><img src="docs/assets/visuals/hero.svg" alt="MetaHarness: your repo, your agent, your harness. A layered rotating torus and orange trajectory introduce a factory for custom agent harnesses." width="100%"></a>
  5. <!-- /readme-motion: hero -->
  6. ### Mint a custom AI agent harness from any repo.
  7. `npx metaharness` Β· [open the Studio β†’](https://ruvnet.github.io/metaharness/)
  8. <sub>(Repo: `ruvnet/metaharness` Β· CLI: `metaharness` Β· Library: `@ruvnet/agent-harness-generator`)</sub>
  9. [![Open the Studio](https://img.shields.io/badge/Studio-open_in_browser_β†—-7c5cff?style=for-the-badge&logo=githubpages&logoColor=white)](https://ruvnet.github.io/metaharness/)
  10. [![User guide](https://img.shields.io/badge/User_guide-plain_language-22c55e?style=for-the-badge)](docs/USERGUIDE.md)
  11. [![Tests β€” 2,254 passing](https://img.shields.io/badge/tests-2%2C254%20passing-22c55e?style=for-the-badge)](docs/ARCHITECTURE.md)
  12. [![npm publish smoke](https://github.com/ruvnet/metaharness/actions/workflows/published-smoke.yml/badge.svg)](https://github.com/ruvnet/metaharness/actions/workflows/published-smoke.yml)
  13. [![License MIT](https://img.shields.io/badge/License-MIT-yellow?style=for-the-badge)](LICENSE)
  14. [![Agent Harness Studio](docs/web-ui/screenshot-desktop.png)](https://ruvnet.github.io/metaharness/)
  15. </div>
  16. <!-- readme-motion: trailer -->
  17. <a href="docs/visual-walkthrough.md"><img src="docs/assets/visuals/trailer.svg" alt="MetaHarness in motion: four animated architecture and workflow chapters. Open the complete walkthrough." width="100%"></a>
  18. [**Explore the full animated walkthrough β†’**](docs/visual-walkthrough.md) Β· 12 chapters Β· 96-second full tour Β· 32-second preview
  19. [Earlier 20-chapter capability tour](docs/assets/metaharness-walkthrough.svg) Β· [Open the Studio](https://ruvnet.github.io/metaharness/)
  20. <table>
  21. <tr><td width="50%"><a href="docs/visual-walkthrough.md"><img src="docs/assets/visuals/icon-generate.svg" width="48" height="48" alt="Generate: Generate an owned harness from project choices."></a><br><strong>Generate</strong><br>A harness for your repo</td><td width="50%"><a href="docs/visual-walkthrough.md"><img src="docs/assets/visuals/icon-govern.svg" width="48" height="48" alt="Govern: Govern tools with explicit policy and audit."></a><br><strong>Govern</strong><br>Tools with explicit policy</td></tr>
  22. <tr><td width="50%"><a href="docs/visual-walkthrough.md"><img src="docs/assets/visuals/icon-evolve.svg" width="48" height="48" alt="Evolve: Evaluate changes through Darwin Mode."></a><br><strong>Evolve</strong><br>Evaluate harness changes</td><td width="50%"><a href="docs/visual-walkthrough.md"><img src="docs/assets/visuals/icon-ship.svg" width="48" height="48" alt="Ship: Publish your customized and validated package."></a><br><strong>Ship</strong><br>Your name. Your package.</td></tr>
  23. </table>
  24. <!-- /readme-motion: trailer -->
  25. ---
  26. ## What this is
  27. <a href="docs/visual-walkthrough.md"><img src="docs/assets/visuals/header-create.svg" alt="MetaHarness generates an owned, customized agent harness." width="100%"></a>
  28. **Every serious repo deserves its own agent.** A repo-aware CLI, a repo-aware coding agent, a local MCP server, memory scoped to the project, skills generated from the actual file layout, governance policy, release verification, witness-signed provenance.
  29. `metaharness` mints those, on demand, from a GitHub URL or a blank slate. **It is not another agent framework. It is a factory for agent frameworks.**
  30. The model is replaceable. The harness is the product.
  31. ## What it gives you
  32. In under 60 seconds, in your browser, with nothing leaving your machine:
  33. - A custom AI agent harness for your repo (or any repo)
  34. - Recommended agents, skills, slash commands, MCP tools
  35. - A scoped memory namespace + governance policy
  36. - Witness-signed provenance + release gates
  37. - Drops into Claude Code, OpenAI Codex, pi.dev, Hermes, OpenClaw, RVM, or Prime Agent β€” pick one or all
  38. Output is an npm-publishable `.zip` with **your name on it, your branding, your `npx <your-name>` CLI**.
  39. ### New
  40. * Bounded contrastive harness development: `@metaharness/darwin/contrastive` adds scoped proposals, actual probes, selective composition and metered experiment memory behind immutable caller policy. [API guide](packages/darwin-mode/CONTRASTIVE.md). Independent paired confirmation adds unseen-task, statistical and regression checks. Public intent-classification evidence remains separate from general agent performance claims.
  41. [Watch Darwin evolve a harness](docs/visual-walkthrough.md#evolve) Β· [Explore model routing](docs/visual-walkthrough.md#routing) Β· [Review experimental extensions](docs/visual-walkthrough.md#extensions)
  42. - **Run an experimental ARC-AGI-3 harness from the ChatGPT UI.**
  43. [`@metaharness/arc-agi-3`](packages/arc-agi-3/) owns exact observations,
  44. persistent evidence-backed memory, belief-state exploration, guarded actions,
  45. supervisor interventions, checkpoints, and hash-chained receipts.
  46. [`@metaharness/arc-agi-3-chatgpt`](packages/arc-agi-3-chatgpt/) exposes that
  47. controller as a remote MCP server and MCP Apps canvas for ChatGPT Developer
  48. Mode. ChatGPT is the OpenAI reasoning host: neither package uses the OpenAI
  49. API or an `OPENAI_API_KEY`. An opt-in ARC-specific AVO loop adds governed
  50. candidate selection, lineage, memory, blocking supervision, and an explicitly
  51. separate retrodiction arm. The frozen synthetic mechanism gate passes, and
  52. one actor-declared-clean single-game smoke favored AVO 3.2676 to 0.3968, but that
  53. non-competition result is not claim-eligible. All packages remain private and
  54. experimental; no ARC performance claim is made until an official closed
  55. scorecard satisfies the frozen controlled-ablation gate in
  56. [ADR-254](docs/adrs/ADR-254-arc-avo-controlled-ablation.md).
  57. - **Governed autonomous variation for difficult engineering work.**
  58. [`@metaharness/avo`](packages/avo/) lets an agent repeatedly inspect, edit,
  59. execute real tools, evaluate, repair/revert, branch, consult structured RVF
  60. memory, and commitβ€”while MetaHarness retains immutable capabilities, budgets,
  61. promotion, quarantine, rollback, and signed replay receipts. Simple work stays
  62. on Darwin's fast path. The runtime has a deterministic 205-action RVF
  63. interruption proof; the stronger β€œAVO-class” claim remains blocked on the
  64. preregistered 100-task unseen SWE-bench gate in [ADR-251](docs/adrs/ADR-251-governed-autonomous-variation-runtime.md).
  65. [ADR-276](docs/adrs/ADR-276-avo-release-claim-evidence-gate.md) now enforces
  66. that boundary at publication against the exact tag SHA and npm tarball,
  67. protected claim semantics, measured cost, lineage roots, and two independent
  68. graders.
  69. - **Run your harness on Prime Agent β€” and borrow its best ideas.** The 10th host
  70. ([`@metaharness/host-prime-agent`](packages/host-prime-agent/), `--host prime-agent`) emits your
  71. tools as project-scoped, Python-backed Prime Agent skills (`.prime/agent/skills/` β€” the host has
  72. no MCP) plus an install runbook. **Fail-closed**: Prime Agent can't enforce a deny-list itself,
  73. so a non-empty deny-list ships a prominent `SANDBOX-REQUIRED.md` instead of silently dropping
  74. your posture. From its design we also shipped **RefineMutator** β€” an evidence-backed proposer in
  75. [`@metaharness/darwin`](https://www.npmjs.com/package/@metaharness/darwin) that must cite the
  76. failing traces behind each edit or propose nothing β€” and an opt-in `--sessions` crash-recoverable
  77. JSONL session log whose Rust and TS (wasm) replays produce identical state hashes. Its PTC
  78. ("kernel as the only tool") claim is honestly *deferred* behind a pre-registered A/B in
  79. `evals-toolcall`. See [ADR-246](docs/adrs/ADR-246-prime-agent-continual-harness-refine.md) /
  80. [ADR-247](docs/adrs/ADR-247-host-prime-agent.md). ($0)
  81. - **Score any repo before you build it.** `npx metaharness score <repo>` reads
  82. the repo (never runs it) and prints a one-screen report card β€” how well a
  83. harness fits, how likely it is to build, how safe the tools are, and the
  84. rough **cost per run** β€” so you know what you'll get before scaffolding.
  85. - **Pick the cheapest model that's good enough.** [`@metaharness/router`](https://www.npmjs.com/package/@metaharness/router)
  86. routes each request to the right model from your own results β€” same quality,
  87. far less spend. Works out of the box with zero native deps; **train it on your
  88. data** for a sharper fit (`npm i @metaharness/router`). Add the optional
  89. [`@ruvector/tiny-dancer`](https://www.npmjs.com/package/@ruvector/tiny-dancer)
  90. to train a fast native model instead β€” same training data, no API change.
  91. - **Route through a learned field without replaying predecessor context.**
  92. [`@metaharness/field-memory`](packages/field-memory/) is an experimental,
  93. single-process-by-default packed-attractor layer with verified outcome caps,
  94. support quarantine, decay, cost-aware choice, and no raw-episode API. Its
  95. RuVector adapter fails closed unless a corrected flat-index wrapper is used.
  96. - **Experiment with bounded constraint planning.**
  97. [`@metaharness/energy`](packages/energy/) provides certified integer task
  98. assignment through exact matching. Its generic `searchEnergy` API optimizes
  99. pairwise models using exact tree or bounded cyclic search and a seeded local
  100. fallback. The package also supports local pseudolikelihood learning. See
  101. [ADR 324](docs/adrs/ADR-324-energy-constraint-module.md).
  102. - **Let your harness improve itself.** Every scaffold now ships with **Darwin Mode**
  103. ([`@metaharness/darwin`](https://www.npmjs.com/package/@metaharness/darwin)) wired in β€”
  104. run `npm run evolve` and the harness mutates its own config, tests each change in a
  105. sandbox, and keeps only what *measurably* improves. The model stays frozen; the harness
  106. evolves. Safe by default (no network, no API key; pure refactor/tuning behind a safety
  107. gate). Validated on real **SWE-bench Lite** bug-fixing. `--no-darwin` to skip.
  108. - **Opt into field memory without copying its implementation.** Add
  109. `--field-memory` to scaffold a small bootstrap around the forthcoming
  110. `@metaharness/field-memory@^0.1.0` package. The generated harness uses packed
  111. attractors, three distinct verifier-derived principals for support, no
  112. hysteresis, and a bounded drift window. Principal independence depends on the
  113. verifier's identity system.
  114. It will not open storage or accept shared updates until the deployment
  115. supplies an absolute storage path, a compatible adapter, and a principal
  116. verifier backed by deployment authentication. Package influence caps are
  117. centroid-local; fleet-wide admission and rate limits belong in that verifier.
  118. The RuVector adapter requires a configuration-verified, in-place FlatIndex
  119. with cosine distance and rejects HNSW or unverified legacy stores.
  120. Minimum support quarantines routing but does not make stored singleton
  121. aggregates confidential; protect field snapshots and registry storage.
  122. A deployment-secret identity hash key of at least 32 bytes is also required
  123. and must survive state restore. Conflicting package declarations or an
  124. existing field module fail before disk writes, and `harness upgrade` reapplies
  125. the opted-in overlay only after validating its exact, versioned manifest
  126. contract. Unsupported overlay schemas require an explicit migration.
  127. `storage.writerScope: process` requires one writer service; multi-process or
  128. fleet use requires `distributed`. The flag is off by default, so existing
  129. scaffold output is unchanged.
  130. - **Distil the cheap tier instead of escalating to a frontier model.** **Weight-EFT**
  131. ([`@metaharness/weight-eft`](packages/weight-eft/), `metaharness weight-eft`) takes the
  132. *complementary* lever to Darwin's gradient-free evolution: it exports the harness's gold-resolved
  133. archive into standard SFT/DPO sets and LoRA-tunes the open cheap tier (GLM/Qwen), so the
  134. cost-cascade escalates to Opus/GPT **less often**. It attacks **cost** (fewer $0.50 escalations),
  135. not the frontier ceiling β€” and stays honest about it. Strict train/eval-disjointness +
  136. reward-hacking filters keep the lift real; the tune is a *gene* Darwin can prune if it overfits.
  137. See [ADR-198](docs/adrs/ADR-198-weight-eft-hybrid-policy-weight-evolution.md). ($0 / GPU-gated.)
  138. ## Tune it to your project β€” then ship it as your own npm
  139. A generated harness is a **starting point you own**, not a fixed framework. Open
  140. it and make it yours:
  141. - **Keep only what your repo needs.** Delete the agents, skills, slash commands,
  142. and MCP servers you won't use β€” the scaffold ships a recommended set, but a
  143. payments service and a docs site want very different harnesses. A smaller,
  144. targeted harness is faster, cheaper, and easier to reason about. `harness
  145. doctor` / `harness validate` keep it healthy as you trim.
  146. - **Optimize the model routing for your work.** Swap the per-task model tiers,
  147. tighten the governance policy, point the memory namespace at your domain. The
  148. harness is config you control, not a black box.
  149. - **Publish it as your own package for the whole org.** Rename it, set your
  150. scope, and `npm publish` β€” now anyone on your team runs
  151. `npx @your-org/your-harness` and gets the *same* repo-tuned agent. One
  152. command, org-wide, versioned like any other dependency. (The 19
  153. [`@metaharness/*`](examples-packages/) examples are exactly this pattern,
  154. published live.)
  155. **Make older, cheaper models punch like frontier ones.** The right harness
  156. isn't a pile of extra steps bolted onto an expensive model β€” it's putting the
  157. *right* model on each task and getting out of the way. Our [DRACO](packages/bench/draco/)
  158. benchmark proves it: a small, cheap model delivers **frontier-quality** research
  159. at roughly **one-tenth the cost**, and a smart router squeezes out the rest.
  160. Stop paying frontier prices for work a $0.10 model does just as well.
  161. That router ships as [`@metaharness/router`](https://www.npmjs.com/package/@metaharness/router)
  162. β€” `route(query)` returns the cheapest model predicted to clear your quality bar,
  163. learned from your own eval logs. `npm i @metaharness/router`.
  164. ## Try it in 30 seconds
  165. <a href="docs/visual-walkthrough.md"><img src="docs/assets/visuals/header-start.svg" alt="Start with the Studio, the wizard, or the CLI." width="100%"></a>
  166. ```bash
  167. # In the browser β€” zero install, nothing leaves the page
  168. open https://ruvnet.github.io/metaharness/
  169. # Or in the terminal β€” the same harness (behaviourally equivalent output)
  170. npx metaharness my-bot --template vertical:coding --host claude-code
  171. cd my-bot && npx . --help
  172. ```
  173. **Don't know what to pick?** Run the wizard:
  174. ```bash
  175. npx metaharness --wizard
  176. ```
  177. **Already have a repo you want a harness for?**
  178. ```bash
  179. harness analyze-repo . # local β€” deterministic analysis only
  180. harness analyze-repo . --scaffold my-bot # materialise the recommended harness
  181. ```
  182. No repository code is executed. Inferred build/test commands are emitted as `trust: inferred Β· execution: disabled`.
  183. πŸ“– **[Read the plain-language user guide β†’](docs/USERGUIDE.md)**
  184. ---
  185. ## Hosts
  186. <a href="docs/visual-walkthrough.md"><img src="docs/assets/visuals/header-hosts.svg" alt="Connect your harness to Claude Code, Codex, RVM, and other hosts." width="100%"></a>
  187. <a href="docs/visual-walkthrough.md#hosts"><img src="docs/assets/visuals/04-hosts.svg" alt="Bring your preferred host. The same generated harness connects through distinct host adapters." width="100%"></a>
  188. The same harness output runs on **ten** agent hosts β€” nine interactive, plus GitHub Actions (CI/CD):
  189. | Host | What ships | Notes |
  190. |---|---|---|
  191. | [**Claude Code**](https://code.claude.com/docs/en/mcp) | MCP server + hooks + 3-scope settings | Richest surface; Ruflo-native |
  192. | [**OpenAI Codex**](https://developers.openai.com/codex) | MCP via `~/.codex/config.toml` | TOML, no hooks |
  193. | [**pi.dev**](https://pi.dev/) | Pi extension via `pi.registerTool()` | No MCP by design |
  194. | [**Hermes**](https://hermes-agent.nousresearch.com/docs/) | MCP runtime, `<think>` scrubbing | Per Hermes issue #741 |
  195. | [**OpenClaw**](https://github.com/openclaw/openclaw) | `~/.openclaw/openclaw.json` + workspace skills | Personal-AI gateway |
  196. | [**RVM**](https://github.com/ruvnet/rvm) | Bare-metal microhypervisor + capability tokens | Hardware isolation for untrusted peers |
  197. | [**GitHub Copilot**](https://code.visualstudio.com/docs/copilot/mcp) | MCP via `.vscode/mcp.json` | VSCode 1.99+ (ADR-032) |
  198. | [**OpenCode**](https://opencode.ai/) | MCP via `.opencode/opencode.json` | sst/opencode TUI (ADR-036) |
  199. | [**Prime Agent**](packages/host-prime-agent/) | `.prime/agent/` skills (Python-backed, no MCP) + `install-prime-agent.md` | Fail-closed sandbox posture (ADR-247) |
  200. | [**GitHub Actions**](https://docs.github.com/actions) | `.github/workflows/` + composite `action.yml` | **Non-interactive** CI/CD; default-deny via `permissions:` (ADR-033) |
  201. See [ADR-004 β€” Host integration model](docs/adrs/ADR-004-host-integration-model.md) and [ADR-033 β€” GitHub Actions host](docs/adrs/ADR-033-host-github-actions.md).
  202. ---
  203. ## MCP β€” modular, default-deny
  204. <a href="docs/visual-walkthrough.md"><img src="docs/assets/visuals/header-policy.svg" alt="Modular MCP, explicit policy, and restrictive defaults." width="100%"></a>
  205. MCP is included as a first-class **adapter surface, not the identity**. It is gated and default-deny ([ADR-022](docs/adrs/ADR-022-mcp-primitive.md)):
  206. - Modes: `off` Β· `local` (stdio) Β· `remote` (HTTPS + auth)
  207. - Emits `src/mcp/{server,tools,resources,prompts,policy,audit}.ts` + a scannable `.harness/mcp-policy.json`
  208. - Safe defaults: no network, no shell, no file-write, approve-dangerous, 30s timeout, 8 calls/turn, audit on
  209. - `harness mcp-scan <path>` β€” *"npm audit for agent tools"*: static-only scan flagging shell/network grants, missing audit/timeouts, wildcard permissions, unguarded secrets, unpinned deps. Exits 1 on any HIGH.
  210. ---
  211. ## Verticals (19 quick-start templates)
  212. ```bash
  213. npx metaharness --list
  214. npx metaharness my-bot --template vertical:coding
  215. ```
  216. | Category | Templates |
  217. |---|---|
  218. | Starter / Operations | `minimal`, `vertical:devops` |
  219. | Engineering | `vertical:coding`, `vertical:ai`, **`vertical:repo-maintainer`** (iter 113) |
  220. | Knowledge | `vertical:research`, `vertical:ruview`, `vertical:education` |
  221. | Finance / Pro | `vertical:trading`, `vertical:legal`, `vertical:health` |
  222. | Customer / Growth | `vertical:support`, `vertical:crm`, `vertical:marketing`, `vertical:advertising`, `vertical:sales` |
  223. | Business / Frontier | `vertical:business`, `vertical:agentics`, `vertical:gaming`, `vertical:exotic` |
  224. Each ships bespoke domain agents (with system prompts), skills, commands, and per-host settings β€” all default-deny.
  225. ---
  226. ## One-command examples
  227. Don't want to pick flags? Each host and vertical has a dedicated
  228. `@metaharness/*` wrapper β€” **published, one `npx` away**, no template/host
  229. flags to remember. A scaffold from a wrapper is byte-identical to the
  230. equivalent `metaharness` invocation.
  231. **Host integrations**
  232. | Package | Scaffolds | npm |
  233. |---|---|---|
  234. | `npx @metaharness/claude-code my-bot` | Claude Code workspace + plugin | [![npm](https://img.shields.io/npm/v/@metaharness/claude-code?color=cb3837&logo=npm)](https://www.npmjs.com/package/@metaharness/claude-code) |
  235. | `npx @metaharness/codex my-bot` | OpenAI Codex | [![npm](https://img.shields.io/npm/v/@metaharness/codex?color=cb3837&logo=npm)](https://www.npmjs.com/package/@metaharness/codex) |
  236. | `npx @metaharness/hermes my-bot` | Hermes cli-config | [![npm](https://img.shields.io/npm/v/@metaharness/hermes?color=cb3837&logo=npm)](https://www.npmjs.com/package/@metaharness/hermes) |
  237. | `npx @metaharness/pi-dev my-bot` | pi.dev AGENTS.md | [![npm](https://img.shields.io/npm/v/@metaharness/pi-dev?color=cb3837&logo=npm)](https://www.npmjs.com/package/@metaharness/pi-dev) |
  238. | `npx @metaharness/openclaw my-bot` | OpenClaw `.openclaw/` | [![npm](https://img.shields.io/npm/v/@metaharness/openclaw?color=cb3837&logo=npm)](https://www.npmjs.com/package/@metaharness/openclaw) |
  239. | `npx @metaharness/rvm my-bot` | RVM deployment partition | [![npm](https://img.shields.io/npm/v/@metaharness/rvm?color=cb3837&logo=npm)](https://www.npmjs.com/package/@metaharness/rvm) |
  240. | `npx @metaharness/copilot my-bot` | VSCode / Copilot `mcp.json` | [![npm](https://img.shields.io/npm/v/@metaharness/copilot?color=cb3837&logo=npm)](https://www.npmjs.com/package/@metaharness/copilot) |
  241. | `npx @metaharness/opencode my-bot` | OpenCode `.opencode/` | [![npm](https://img.shields.io/npm/v/@metaharness/opencode?color=cb3837&logo=npm)](https://www.npmjs.com/package/@metaharness/opencode) |
  242. | `npx @metaharness/github-actions my-bot` | GitHub Actions CI/CD (non-interactive) | [![npm](https://img.shields.io/npm/v/@metaharness/github-actions?color=cb3837&logo=npm)](https://www.npmjs.com/package/@metaharness/github-actions) |
  243. **Vertical workflows** (ready-made multi-agent pods)
  244. | Package | Scaffolds | npm |
  245. |---|---|---|
  246. | `npx @metaharness/devops my-bot` | incident response | [![npm](https://img.shields.io/npm/v/@metaharness/devops?color=cb3837&logo=npm)](https://www.npmjs.com/package/@metaharness/devops) |
  247. | `npx @metaharness/research my-bot` | multi-source dossier | [![npm](https://img.shields.io/npm/v/@metaharness/research?color=cb3837&logo=npm)](https://www.npmjs.com/package/@metaharness/research) |
  248. | `npx @metaharness/trading my-bot` | quant trading (paper-by-default) | [![npm](https://img.shields.io/npm/v/@metaharness/trading?color=cb3837&logo=npm)](https://www.npmjs.com/package/@metaharness/trading) |
  249. | `npx @metaharness/support my-bot` | customer support | [![npm](https://img.shields.io/npm/v/@metaharness/support?color=cb3837&logo=npm)](https://www.npmjs.com/package/@metaharness/support) |
  250. | `npx @metaharness/legal my-bot` | contract redline (drafts only) | [![npm](https://img.shields.io/npm/v/@metaharness/legal?color=cb3837&logo=npm)](https://www.npmjs.com/package/@metaharness/legal) |
  251. | `npx @metaharness/repo-maintainer my-bot` | OSS repo maintainer | [![npm](https://img.shields.io/npm/v/@metaharness/repo-maintainer?color=cb3837&logo=npm)](https://www.npmjs.com/package/@metaharness/repo-maintainer) |
  252. | `npx @metaharness/coding my-bot` | engineering pod | [![npm](https://img.shields.io/npm/v/@metaharness/coding?color=cb3837&logo=npm)](https://www.npmjs.com/package/@metaharness/coding) |
  253. | `npx @metaharness/education my-bot` | tutor pod | [![npm](https://img.shields.io/npm/v/@metaharness/education?color=cb3837&logo=npm)](https://www.npmjs.com/package/@metaharness/education) |
  254. | `npx @metaharness/sales my-bot` | sales pipeline pod | [![npm](https://img.shields.io/npm/v/@metaharness/sales?color=cb3837&logo=npm)](https://www.npmjs.com/package/@metaharness/sales) |
  255. | `npx @metaharness/gaming my-bot` | game-design pod | [![npm](https://img.shields.io/npm/v/@metaharness/gaming?color=cb3837&logo=npm)](https://www.npmjs.com/package/@metaharness/gaming) |
  256. All 18 are live on npm under [`@metaharness`](https://www.npmjs.com/org/metaharness). Source + per-package README:
  257. [`examples-packages/`](examples-packages/) Β· plain-language deep-dive gists:
  258. [`examples-packages/GISTS.md`](examples-packages/GISTS.md).
  259. ---
  260. ## Day-to-day commands
  261. After scaffolding, every harness has a `harness` CLI:
  262. | You're trying to … | Subcommand |
  263. |---|---|
  264. | Smoke-check the scaffold | `harness doctor` |
  265. | Run every release gate | `harness validate` |
  266. | Check kernel ↔ harness compatibility | `harness diag` |
  267. | Score the harness 0-100 with badges | `harness score` |
  268. | Pre-scaffold: is this REPO ready for an agent? | `harness genome <repo>` |
  269. | Pre-scaffold: fit/cost/safety report card for a repo | `metaharness score <repo>` |
  270. | MCP threat-model artifact for a PR review | `harness threat-model` |
  271. | Declare OIA v0.1 layer alignment | `harness oia-manifest` |
  272. | File a useful support ticket | `harness diag --bundle > bundle.json` |
  273. | Diff two harnesses | `harness compare a/ b/` |
  274. | Share MCP + Bash + claims config for review | `harness export-config` |
  275. | Run npm-audit per-harness | `harness audit --bundle > audit.json` |
  276. | Emit SPDX-2.3 SBOM | `harness sbom` |
  277. | Drift-detect against the latest template | `harness upgrade` |
  278. | Sign / verify the witness | `harness sign` Β· `harness verify` |
  279. | Pin the manifest to IPFS | `harness publish --confirm` |
  280. | Recommend a harness from a repo | `harness analyze-repo` |
  281. 21 subcommands total. Every one respects `--help` / `-h`. Shell completion: `harness completions bash | zsh | fish`.
  282. πŸ“– Full reference: [docs/USAGE.md](docs/USAGE.md)
  283. ---
  284. ## Status
  285. **v0.1.x beta** β€” published and usable, with the credibility/doc reconciliation in
  286. issue #4 / ADR-042 in progress. The release *pipeline* is mature: CI matrix green
  287. across Rust Γ— 3 OS + WASM Γ— 3 OS + Node 20+22 Γ— 3 OS + Bench + pack+install Γ— 3 OS
  288. + a CI-passed aggregator; single-command releases (`node scripts/release.mjs <bump>
  289. --push`) atomically bump 15 sources, run all gates, and tag.
  290. | Layer | Status |
  291. |---|---|
  292. | Rust kernel (WASM + NAPI-RS) | Shipped β€” 7 subsystems |
  293. | 10 host adapters | claude-code Β· codex Β· pi-dev Β· hermes Β· openclaw Β· rvm Β· copilot Β· opencode Β· github-actions Β· prime-agent |
  294. | 17 `harness` subcommands | Shipped |
  295. | 7 Codex skills | Shipped |
  296. | Claude marketplace plugin | Shipped + schema-validated |
  297. | Witness signing (Ed25519) | Shipped + tamper-tested |
  298. | MCP tool dispatch | 11 end-to-end cases |
  299. | Test suite | **2,254 passing** across 246 files (CI green on `main`) |
  300. | CI matrix | 16 jobs green |
  301. | Security pipeline | cargo-audit Β· cargo-deny Β· npm-audit Β· CodeQL Β· SBOM (SPDX-2.3) |
  302. | Publish pipeline | GCP WIF + 2 gates + 11 packages + IPFS pin |
  303. | Agent Harness Studio | Live at <https://ruvnet.github.io/metaharness/> |
  304. ---
  305. ## Architecture in 30 seconds
  306. <a href="docs/visual-walkthrough.md"><img src="docs/assets/visuals/header-architecture.svg" alt="Shared Rust primitives behind customized harnesses and host adapters." width="100%"></a>
  307. <a href="docs/visual-walkthrough.md#kernel"><img src="docs/assets/visuals/03-kernel.svg" alt="One shared kernel. Generated harnesses consume a shared Rust kernel through host adapters." width="100%"></a>
  308. <details>
  309. <summary>Text architecture</summary>
  310. ```
  311. You (harness author)
  312. β””β†’ agent-harness-generator ← the factory
  313. β””β†’ Your harness (.zip) ← what you ship
  314. β”œ npx <your-name> ← your identity
  315. β”œ <your agents> ← your content
  316. β”” @metaharness/kernel ← shared primitives (Rust + WASM + NAPI-RS)
  317. β””β†’ Host adapter (Claude Code / Codex / pi.dev / Hermes / OpenClaw / RVM)
  318. β””β†’ LLM providers
  319. ```
  320. </details>
  321. You operate the factory. The factory produces your harness. Your users never see the factory β€” only the brand and CLI you ship. The kernel ships as `@metaharness/kernel` (Rust β†’ wasm-pack + NAPI-RS); your content stays yours.
  322. πŸ“– Deeper: [docs/ARCHITECTURE.md](docs/ARCHITECTURE.md) Β· [docs/adrs/INDEX.md](docs/adrs/INDEX.md) (220 ADRs)
  323. ---
  324. ## Quality gates
  325. <a href="docs/visual-walkthrough.md"><img src="docs/assets/visuals/header-quality.svg" alt="Quality gates, tests, provenance, and reviewable releases." width="100%"></a>
  326. | Concern | Where |
  327. |---|---|
  328. | **CI** | [`ci.yml`](.github/workflows/ci.yml) β€” Rust 3-platform Γ— fmt/clippy/test/doc + WASM build + size budget + Node 20/22 Γ— 3-platform |
  329. | **Publish** | [`publish.yml`](.github/workflows/publish.yml) β€” GCP WIF β†’ Secret Manager β†’ smoke β†’ `npm publish --provenance` (SLSA L2) |
  330. | **Security** | [`security.yml`](.github/workflows/security.yml) β€” cargo-audit + cargo-deny + npm-audit + CodeQL + weekly cron |
  331. | **Provenance** | [ADR-011](docs/adrs/ADR-011-witness-and-provenance.md) β€” Ed25519-signed witness manifest, byte-deterministic across runners |
  332. | **Studio liveness** | [`pages-monitor.yml`](.github/workflows/pages-monitor.yml) β€” daily HTTP probe of live Studio |
  333. | **Research quality (DRACO)** | [`draco.yml`](.github/workflows/draco.yml) β€” cross-domain deep-research benchmark ([ADR-037](docs/adrs/ADR-037-draco-deep-research-benchmark.md)). Deterministic subset gates the scorer/runner machinery on every push (offline); a weekly judged cadence runs the real OpenRouter-fusion score. 5 dimensions (grounding/coverage/balance/cleanliness/faithfulness); the verifier + judge are different model families than the synthesizer (fusion). See [`packages/bench/draco/`](packages/bench/draco/). |
  334. ---
  335. ## Developer quick-start
  336. ```bash
  337. git clone https://github.com/ruvnet/metaharness
  338. cd metaharness
  339. cargo test --workspace
  340. cargo clippy --workspace --all-targets -- -D warnings
  341. npm install
  342. npm run build:wasm
  343. npm test
  344. node scripts/healthcheck.mjs
  345. ```
  346. See [CONTRIBUTING.md](CONTRIBUTING.md).
  347. ---
  348. ## Related
  349. - [**ruflo**](https://github.com/ruvnet/ruflo) β€” the meta-harness this generator factors apart
  350. - [**ruvector**](https://github.com/ruvnet/ruvector) β€” vector + agentic database (memory backend)
  351. - [**@ruvector/emergent-time**](https://www.npmjs.com/package/@ruvector/emergent-time) β€” memory-decay clock the kernel uses
  352. ## Cost-Pareto Leaderboard
  353. **[β†’ Live leaderboard β€” the cost-performance Pareto frontier](https://ruvnet.github.io/metaharness/cost-pareto.html)**
  354. Which harness resolves the most benchmark per **dollar**, not just the highest raw score β€” ranked by a tunable
  355. **Value Score** that blends capability and price. Tabs: SWE-bench **Lite** / **Verified** / **Pro** / **DRACO** /
  356. **SWE-ultralite** (a fast n=25 modelΓ—mode scouting map). Every Darwin row is fully conformant (the solver never
  357. sees gold tests during solving) and measured with the official `swebench` Docker harness, with Wilson 95% CIs.
  358. **Latest validated, conformant results (2026-06-26):**
  359. | benchmark | config | resolve | n | Wilson 95% CI | cost |
  360. |---|---|---|---|---|---|
  361. | **SWE-bench Verified** | GLM→Opus empty-patch cascade | **55.6%** (278/500) | 500 | **[51.2, 59.9]** | ~$0.15/inst (est.) |
  362. | SWE-bench Lite | GLM→Opus empty-patch cascade | 51.3% | 300 | — | ~$0.27/inst |
  363. | **LiveCodeBench** (release_v5 β‰₯2024-12-01) | single-shot | **44%** | 100 | β€” | β€” |
  364. | **LiveCodeBench** (release_v5 β‰₯2024-12-01) | cost-cascade | **62%** | 100 | β€” | β€” |
  365. The same cheap→frontier **empty-patch cascade** (escalate only the empties to Opus) measures **55.6% on the full
  366. SWE-bench Verified (500)** β€” official `swebench` gold eval, conformant, ~56Γ— cheaper than frontier-only β€” which
  367. **beats** the Lite cascade (51.3%); the pattern generalizes across both splits. SWE-bench Verified cost is an
  368. **estimate** (per-instance cost not captured in predictions). LiveCodeBench (n=100, contamination-resistant
  369. post-cutoff window) lifts **44% β†’ 62%** via cost-cascade; caveats are honest: deepseek snapshot cutoff unpinned,
  370. part of the lift is run-to-run variance (clean attributable lift +8 on the escalated tail), n=100 is directional
  371. (not 1:1 with the official whole-release ~34%). Details in
  372. [`packages/darwin-mode/bench/results/RESULTS.md`](packages/darwin-mode/bench/results/RESULTS.md) (Β§33) and
  373. [`LEARNINGS.md`](packages/darwin-mode/LEARNINGS.md) (Β§46b/Β§47).
  374. **Run your own harness and put it on the board** β€” fork, add a measured row, open a PR.
  375. Conformance rules, the entry schema, and a step-by-step walkthrough are in **[SUBMISSIONS.md](SUBMISSIONS.md)**.
  376. ## License
  377. MIT β€” see [LICENSE](LICENSE).
  378. ## FAQ
  379. ### What is MetaHarness?
  380. MetaHarness is a CLI and browser Studio that turns any GitHub repo (or a
  381. blank slate) into a custom AI agent harness. The output is a branded,
  382. npm-publishable package with its own `npx <name>` CLI, MCP server, memory,
  383. governance policy, and Ed25519 witness-signed releases. Runs on ten hosts:
  384. Claude Code, Codex, pi.dev, Hermes, OpenClaw, RVM, Copilot, OpenCode,
  385. GitHub Actions, and Prime Agent.
  386. ### How is MetaHarness different from an agent framework?
  387. Frameworks help **developers** build agents. MetaHarness helps
  388. **repositories** ship agents. The model is replaceable; the harness is
  389. the product.
  390. ### Do I need to run a server?
  391. No. The Studio is 100% client-side (GitHub Pages). The CLI runs locally.
  392. There is no MetaHarness account, no hosted backend, no telemetry.
  393. ### Does it execute my code during analysis?
  394. No. `metaharness analyze` and `metaharness genome` are deterministic
  395. static-analysis only. Inferred build/test commands are marked
  396. `trust: inferred Β· execution: disabled`.
  397. ### Which agent runtimes does it support?
  398. Ten today: Claude Code, Codex, pi.dev, Hermes, OpenClaw, RVM, Copilot,
  399. OpenCode, GitHub Actions, Prime Agent.
  400. ### What languages does it understand?
  401. Rust, TypeScript / JavaScript, Python, and Go are detected
  402. deterministically via lockfile and manifest probing. Lexical scoring is
  403. the default; optional in-browser MiniLM embeddings via Transformers.js
  404. boost recall for unusual repos.
  405. ### Is the output really npm-publishable?
  406. Yes β€” the generated harness ships with `package.json`, `bin`, a working
  407. CLI, and `harness validate` to gate releases. `harness sign` adds the
  408. Ed25519 witness; `harness sbom` emits SPDX-2.3.
  409. ---
  410. > **Keywords:** metaharness, AI agent CLI, AI agent scaffold, AI agent generator, repo to agent, GitHub repo to AI agent, agent harness, agent harness generator, agent framework alternative, agentic AI, agentic workflow, autonomous AI agents, multi-agent framework, multi-agent system, MCP, MCP server, model context protocol, Claude Code plugin, OpenAI Codex plugin, Anthropic agents, GPT agent, Codex agent, pi.dev extension, hermes agent, Nous Research, OpenClaw, RVM agent, vertical AI agents, custom AI CLI, npx metaharness, npm create AI agent, Rust WASM agent kernel, NAPI-RS, wasm-bindgen, agent memory, ReasoningBank, HNSW vector search, emergent time, witness manifest, Ed25519 signed, provenance, SBOM, SPDX, SLSA, plugin marketplace, IPFS registry, drift detection, anti-slop, TDD agents, self-evolving agents, federated agents, swarm intelligence, GCP Workload Identity Federation, Secret Manager, npm provenance, repo-aware AI, repo-native CLI, repo factory.

Ratings and comments

Sign in to rate this skill.

Sign in

Free to use. If it’s helped, pass it on.