MetaHarness builds custom AI agent setups for each repo
MetaHarness is an open-source tool that creates a custom AI agent setup from a or a blank project. The generated setup can include a repo-aware , a local MCP server, project-scoped memory, skills based on the file layout, governance rules, release checks, and signed proof of origin. It can run in the browser without installation, or from the terminal with the npx metaharness command.
It exports settings for Claude Code, , pi.dev, Hermes, OpenClaw, RVM, , OpenCode, and . The output can be renamed, branded, and published as an npm package so a team can run the same repo-tuned agent with one npx command. The cost angle is that teams can remove agents, skills, commands, and MCP servers they do not need, leaving a smaller setup that should be faster, cheaper, and easier to understand.
It also lets teams adjust model tiers by task, memory scope, and governance rules. Its default safety settings block network access, shell access, and file writes, require approval for risky actions, and add time limits, call limits, and audit logs.
Key points
- It creates a custom AI agent setup from a GitHub repo or a blank project.
- The generated setup can include a CLI, MCP server, memory, skills, governance rules, and release checks.
- It supports Claude Code, , , , and several other agent hosts.
- Teams are encouraged to delete unneeded agents and tools so the setup stays smaller and cheaper.
- Default safety rules block network, shell, and file-write access unless allowed.