Weft
Author a plugin once, then compile it to every coding-agent harness: Claude Code, OpenAI Codex, Cursor, GitHub Copilot, and OpenCode. Evals and trust come built in. Weft is a compiler plus conventions, not a hosted platform.
weft init my-plugin # author once: weft.yaml + a SKILL.md
weft build my-plugin # compile to every harness (zero hand-written JSON)
weft install my-plugin # place into the harnesses you actually have
weft eval my-plugin # drive the REAL headless harness and assert
weft publish my-plugin # deterministic trust gate (valid + scan + evals)
Documentation
- Concepts. Start here: plugin vs marketplace.
- Demo. Five real-world use cases, end to end.
- Getting started. A full CLI walkthrough.
- CLI reference. Every command, generated from the tree.
- Architecture. The compile pipeline and the adapter seam.
- Authoring plugins.
weft.yaml/marketplace.yaml/cases.yaml. - Writing an adapter. The public contract.
- Harness research. Verified per-harness facts.
- Roadmap. Phased plan and acceptance status (all phases done).
Why
Every harness now supports the same primitives (skills, MCP servers, sub-agents, hooks,
commands) bundled as a plugin and distributed through a marketplace, but each has its own
manifest format. Weft lets you write the canonical, upstream-standard sources once
(SKILL.md, MCP server.json) and generates every harness’s native manifests, the
marketplace catalog, eval results, and trust badges. Distribution stays git-first.