Provenant
A command-line interface and MCP server that gives AI coding agents a local, cited memory layer for a Git repository, built from generated wiki pages, symbol context, git history and dead-code and risk signals.
Provenant reads a codebase and builds a local memory layer for coding agents: generated wiki pages, tracked symbols and dependencies, git history, dead-code detection and per-file risk scores, all stored under a .provenant directory rather than sent anywhere. Agents such as Claude Code, Cursor, Windsurf, Cline or Copilot query this layer through MCP tools to get cited, focused context instead of repeatedly loading raw source files.
On the SWE-bench Verified benchmark the project reports 63.8% file coverage at 5 results and 75.2% at 10 with full retrieval, using 60 to 65 times less context than loading files directly. It won first place nationally in the Microsoft Build AI 2026 hackathon for codebase memory and retrieval.