ruvnet/metaharness

What it solves

MetaHarness is a "factory for agent frameworks," designed to solve the problem of creating custom, repo-aware AI agents for specific software projects. Instead of using a generic agent, it allows developers to quickly generate a tailored "harness"—a package containing agents, skills, slash commands, and governance policies—that is specifically tuned to a project's file layout and needs.

How it works

The tool analyzes a GitHub repository or a blank slate to "mint" a custom agent harness. This harness is delivered as an npm-publishable package that the user owns and can brand. The system uses a Rust-based kernel (via WASM and NAPI-RS) for shared primitives and provides adapters for ten different agent hosts, such as Claude Code, OpenAI Codex, and GitHub Actions.

Key technical components include:

• Model Routing: A router that selects the cheapest model capable of completing a task to reduce costs.
• Darwin Mode: An evolutionary mechanism where the harness mutates its own configuration and tests changes in a sandbox to improve itself.
• Weight-EFT: A process that exports gold-resolved archives into SFT/DPO sets to fine-tune cheaper open-tier models.
• MCP Integration: A default-deny Model Context Protocol (MCP) implementation for secure tool dispatch.

Who it’s for

Software engineers, OSS maintainers, and organizations that want to deploy project-specific AI agents that are versioned, governed, and deployable via npx across a team.

Highlights

• Multi-Host Support: Compatible with 10 different hosts including Claude Code, Copilot, and Prime Agent.
• Rapid Scaffolding: Generates a custom harness in under 60 seconds via a browser studio or CLI.
• Repo Analysis: Includes a score command to evaluate a repository's suitability for an agent before building.
• Self-Evolution: Features "Darwin Mode" for automated configuration tuning and "Weight-EFT" for model distillation.
• Security-First: Implements a default-deny MCP policy and a mcp-scan tool for threat modeling agent tools.