graft

graft keeps one canonical agent definition in sync across the AI-coding tools you use. You edit an agent once; graft propagates that change to every provider — and pulls a change made in any single provider back to the canonical source and out to the rest.

Why

If you use several AI-coding assistants, each keeps its own agent/instruction files in its own layout (.claude, .codex, .cursor, …). Editing the same agent in eight places by hand drifts immediately. graft makes the canonical store under .graft/ the single source of truth and uses a git branch/worktree merge engine to keep providers equal.

What it does

• Stores a provider-neutral agent under .graft/agents/<name>/.
• Transforms canonical ⇄ provider in both directions — a change at any provider can propagate to all.
• Syncs through a deterministic, resumable, concurrency-safe merge engine.
• Reports drift: which providers are out of sync, for which agents.
• Validates schema and semantics before every sync, with per-provider providerOverrides validation.
• Manages skills: a canonical store under .agents/skills/ symlinked into supporting providers.
• Checks for and installs binary updates via graft update.

Supported platforms

macOS and Windows are both supported. On Windows, symlink-based skill linking requires Developer Mode or Administrator privileges; without those, skill commands report a capability error rather than silently falling back to a file copy.

What is not active yet

• antigravity — included in the embedded catalog but unregistered in the sync engine pending a research spike on the agent-definition format. It is not in the active provider set.
• No TUI or web UI.
• No git hooks: migration and sync run only when you invoke graft sync.

Start here

1. Quickstart — from install to your first sync.
2. Install — install, verify, upgrade, uninstall.
3. Core concepts — the canonical-first mental model.
4. CLI reference — every command.