ruvnet/metaharness

What it solves

MetaHarness is a "factory for agent frameworks," allowing users to quickly generate a custom, repo-aware AI agent harness. Instead of using a generic agent framework, it creates a tailored environment—including a CLI, skills, and governance policies—that is specifically tuned to a particular codebase or project, which can then be published as a standalone npm package.

How it works

The tool analyzes a GitHub repository or a blank slate to "mint" a harness. This harness consists of a branded CLI (npx <your-name>) and a set of agents and skills powered by a shared Rust-based kernel (@metaharness/kernel). The generated output is an npm-publishable package that can be deployed across various agent hosts, such as Claude Code, OpenAI Codex, and GitHub Actions.

Key technical components include:

• Model Routing: A router that directs requests to the cheapest model capable of completing the task to reduce costs.
• Darwin Mode: A self-improvement system where the harness mutates its own configuration and tests changes in a sandbox to measurably improve performance.
• Weight-EFT: A process that exports gold-resolved archives into SFT/DPO sets to LoRA-tune cheaper open-tier models.
• MCP Integration: A modular, default-deny adapter surface for tools and resources.

Who it’s for

Developers and organizations who want to provide their team with a standardized, repo-specific AI agent (e.g., a coding agent or a repo maintainer) without building a framework from scratch or relying on a black-box solution.

Highlights

• Multi-Host Support: Compatible with 10 different hosts including Claude Code, Copilot, and Prime Agent.
• Repo Analysis: Includes a score command to evaluate a repository's suitability for an agent before scaffolding.
• Self-Evolving: Capable of autonomous configuration tuning via Darwin Mode.
• Cost Optimization: Uses a smart router to minimize reliance on expensive frontier models.
• Security-First: Implements a default-deny policy for MCP tools and provides a static-only scan for threat modeling.