A local code-graph context layer for coding agents. Structure over guesswork: no embeddings, no telemetry.
Your coding agent doesn't have a context problem. It has a map problem. Every agent session starts the same way: re-exploring your codebase file by file, burning tokens to rediscover what your team already knows. Vector search helps, but embeddings go stale and retrieval is guesswork.
Graft takes a different route. It builds a local, linked code graph of your repo and lets the agent navigate structure instead of re-reading files. No vector embeddings, no telemetry, runs 100% on your machine.
Why it matters: discovery is the most expensive part of agentic coding. The project claims up to 4x cheaper and 3x faster. Those are their numbers, not independently verified by me, but the mechanism is sound. A deterministic map beats probabilistic retrieval for navigation.
The debatable part: if code graphs beat embeddings for agent context, a lot of RAG infrastructure built for code assistance is solving the wrong problem.