ruvnet/metaharness

Metaharness gives you the plumbing layer for a production AI agent: a scaffolded npx CLI, an MCP server, memory and a learning loop, plus witness-signed releases baked in from the start. The part that actually differentiates it is the witness-signed release path — that is not something you get from a weekend wrapper project, and it matters the moment you are shipping agents to clients who need an audit trail. It also targets hardware-isolated sandbox environments (RVM), which is a real requirement for anyone running agents against sensitive infrastructure. Compatibility list is broad: Claude Code, Codex, Hermes, and a few others, so you are not locked into one provider. Reservation: the scaffolding is opinionated, and if your agent architecture does not map onto its assumptions, you will spend time unwinding defaults rather than building features. The docs will tell you quickly whether the fit is real. -> Best for: AI engineer or SaaS team of 2-5 building branded agent tooling for clients