// the find
abhigyanpatwari/GitNexus
GitNexus: The Zero-Server Code Intelligence Engine
GitNexus parses a repo with tree-sitter into a knowledge graph (calls, clusters, execution flows) and serves it to AI coding agents over MCP, instead of relying on embedding search or raw grep. It's aimed at people driving Cursor/Claude Code/Codex against large codebases who are tired of agents missing a caller and silently breaking something three files away. There's also a browser-only WASM mode for quick one-off exploration without installing anything.
The MCP surface actually goes past the usual 'search + read file' pair most of these tools ship — impact analysis with blast-radius depth, shortest-path tracing between symbols, route/tool maps for API handlers, and an optional PDG layer for taint and data-dependence queries. The tree-sitter story is handled properly: vendored prebuilt binaries for Swift/Kotlin/Dart/Proto mean no C++ toolchain is required for most installs, and there's an explicit env var escape hatch for CI boxes without one. The Claude Code/Codex hooks that detect a stale index after a commit and prompt a reindex is a small but real bit of polish — most competitors leave staleness as a silent correctness bug.
License is PolyForm Noncommercial, which directly contradicts the 'context engine for Enterprise Codebases' framing in the README — most companies can't actually adopt this without a separate agreement. The documented Render deploy is honest about being weakly secured: a single bearer token is the only gate, Origin is stripped so CSRF protection is a no-op, and anyone holding the token can read every indexed repo — fine solo, not something to hand a team without reading SECURITY.md first. The flag surface is enormous (dozens of analyze options, a separate auto-sync YAML config, Spring Actuator and AsyncAPI integrations) for a tool pitched as 'zero-server' simplicity, which means real surface area for misconfiguration. Embeddings are opt-in and pull in onnxruntime-node/transformers with a hard Node >=22.15 floor, so semantic search silently degrades to keyword-only on anything but a current toolchain.