MetaHarness
MetaHarness scaffolds a focused, branded AI agent harness from any repository, with an npx CLI, MCP server, memory, learning loop, governance policy, and witness-signed releases. It supports Claude Code, Codex, Hermes, OpenClaw, and RVM.
Visit ruvnet/metaharnessOverview
MetaHarness is the meta-harness for AI agents and a factory for custom agent frameworks. It generates repo-aware harnesses with a CLI, coding agent, MCP server, project-scoped memory, generated skills, governance policy, release verification, and witness-signed provenance from a GitHub URL or blank slate. The output is an npm-publishable package with user branding and an npx CLI, designed to run across hosts such as Claude Code, Codex, pi.dev, Hermes, OpenClaw, RVM, and GitHub Actions, within Cognitum's sovereign, verifiable, edge-native agentic AI platform.
Key Features
- Generates a custom AI agent harness for a repository or blank slate, including recommended agents, skills, slash commands, MCP tools, scoped memory namespace, and governance policy.
- Ships default-deny MCP support with off, local stdio, and remote HTTPS/auth modes plus policy and audit files; harness mcp-scan performs static checks for risky tool grants.
- Includes Darwin Mode, which lets the harness mutate its own config, test changes in a sandbox, and keep only measurable improvements while the model stays frozen.
- Provides a model router that selects the cheapest model predicted to clear the user's quality bar from eval logs, with optional training on custom data.
- Supports ten agent hosts, including Claude Code, OpenAI Codex, pi.dev, Hermes, OpenClaw, RVM, GitHub Copilot, OpenCode, Prime Agent, and GitHub Actions.
- Outputs an npm-publishable package with custom branding, a custom npx CLI, and host-specific scaffold wrappers.
Use Cases
- Create a repo-specific coding agent harness for an engineering team and publish it org-wide as a versioned npm package.
- Scaffold domain workflows such as DevOps incident response, multi-source research, customer support, or paper-by-default trading.
- Add governed MCP tools, project-scoped memory, and generated skills to existing agent hosts such as Claude Code or Codex.
- Score a repository before scaffolding to estimate harness fit, build likelihood, tool safety, and rough cost per run.
Getting Started
- Use the browser Studio, the wizard, or the CLI to generate a harness.
- Run npx metaharness my-bot --template vertical:coding --host claude-code; then cd my-bot and run npx . --help.
- Run npx metaharness --wizard when unsure which options to choose.
- For an existing repository, run harness analyze-repo . --scaffold my-bot.
Deployment & Requirements
- GitHub Copilot host integration requires VSCode 1.99 or later.
- The Prime Agent host uses project-scoped, Python-backed skills under .prime/agent/skills and does not use MCP.
- The field memory overlay requires an absolute storage path, a compatible adapter, and a principal verifier backed by deployment authentication before it opens storage or accepts shared updates.
Before You Adopt
- License: MIT. Review its terms before using, modifying, or distributing the project.