# metaharness [Health: Active]

**Category:** 🤖 Coding Agents  
**Repository:** https://github.com/ruvnet/metaharness  
**GitHub Stars:** 722  
**Views:** 0  
**Installs:** 0  
**Upvotes:** 0  
**Directory Page:** https://allmcps.com/mcp/metaharness

## Description
The meta-harness for AI agents — 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).

## Claude Desktop Quick Installation
Install path detected from listing signals. Uses `npx` (confidence: high):

```json
"mcpServers": {
  "metaharness": {
    "command": "npx",
    "args": ["-y","metaharness"]
  }
}
```

## Documentation

## What metaharness does

metaharness is a factory for creating repository-specific agent harnesses. It accepts a GitHub URL or starts from a blank slate, then produces a customized package with your naming and branding, an `npx <your-name>` command, and the components needed to operate an agent around a project.

The generated output can include a repo-aware coding agent, recommended agents and skills, slash commands, MCP tools, project-scoped memory, governance rules, release checks, and witness-signed provenance. The result is packaged as an npm-publishable ZIP rather than used as a hosted agent service.

## How it works

You can begin with the command-line interface or open the web-based Agent Harness Studio. The CLI also provides a `score` command for examining a repository before generation. Scoring reads the repository without running it and reports harness fit, build likelihood, tool-safety information, and an estimated cost per run.

After configuration, metaharness generates a standalone harness for a selected host. Supported host choices named by the project include Claude Code, OpenAI Codex, pi.dev, Hermes, OpenClaw, RVM, and Prime Agent. Prime Agent receives project-scoped Python-backed skills rather than MCP tools, along with an installation runbook; its generated output includes a sandbox warning when a deny-list cannot be enforced by the host.

The project also includes experimental evolution components. Darwin Mode supports bounded proposals, probes, selective composition, and metered experiment memory under caller policy. A separate autonomous variation runtime can inspect, edit, execute tools, evaluate changes, repair or revert work, branch, use structured memory, and commit while applying budgets, promotion, quarantine, rollback, and signed replay receipts.

## Setup and configuration

Run the generator through the package command:

```bash
npx metaharness
```

The Studio is available from the project’s GitHub Pages site. Generation can use a GitHub repository URL or a blank project. The resulting archive is intended for publication under the owner’s name and can expose a custom `npx` command.

Model selection is replaceable. The optional `@metaharness/router` package can route requests using results supplied by the user, and the project describes an optional native training dependency for that router. These routing packages are extensions to the generated harness rather than prerequisites for using the generator.

## Tools and capabilities

- Generate a branded harness from a repository or blank slate.
- Produce a custom CLI and npm-publishable ZIP package.
- Create project-scoped memory and governance policy.
- Recommend agents, skills, slash commands, and MCP tools from repository structure.
- Add release verification and witness-signed provenance.
- Score a repository before scaffolding it.
- Route requests among models using user-provided results.
- Evaluate harness changes through Darwin Mode.
- Export Prime Agent skills and an installation runbook.

## Limitations and notes

Prime Agent does not provide MCP support and cannot enforce a deny-list itself. When a deny-list is present, the Prime Agent host emits a `SANDBOX-REQUIRED.md` notice instead of silently removing that policy.

Several advanced features are explicitly experimental. The ARC-AGI-3 packages are private and experimental, and the project makes no ARC performance claim without its stated evaluation gate. Stronger claims about autonomous variation and model-routing behavior depend on separate evaluation conditions. Treat these packages as research extensions rather than established requirements for the basic generator.

The generated harness is intended to be owned and published by the user. It is not presented as a single hosted agent or as a replacement for the model used by the harness.

_Full upstream README: https://allmcps.com/mcp/metaharness/readme_