Draft launches a local, deterministic code graph engine that replaces cloud-based embeddings and vector search for AI code understanding.

Draft ships a Node.js binary that builds a version-controlled knowledge graph of your codebase into plain JSONL files (module-graph.jsonl, call-index.jsonl, hotspots.jsonl, etc.), eliminating the need for external vector databases, embedding APIs, or hosted retrieval services.

The graph answers deterministic queries by name and file path—e.g., "which functions call buildGoIndex?"—rather than probabilistic semantic search; it also enables impact queries ("what depends on this file?"), hotspot ranking (files by complexity), and cycle detection, all running locally for zero per-token cost.

For regulated or air-gapped environments (finance, healthcare, defense), the entire workflow stays on-premises; the graph diffs in git PRs show structural relationship changes, and Draft's skills (/draft:implement, /draft:review, /draft:bughunt, /draft:debug, /draft:decompose) use graph queries to scope AI reasoning without transmitting code off-machine.

Available as a Claude Code plugin at https://getdraft.dev; /draft:init builds the graph during setup.