MCP Code Indexer

What Nishant-Chaudhary5338/mcp-code-indexer MCP server does

Nishant-Chaudhary5338/mcp-code-indexer MCP server turns a TypeScript, React, or Next.js repository into a graph of code entities and relationships. The graph can represent repositories, applications, packages, folders, files, components, functions, and external entities. Relationships include containment, imports, calls, renders, references, and package dependencies.

The server is intended for code navigation and impact analysis. It can answer reverse queries such as which components render a target or which functions call it, locate references, calculate a change blast radius, identify dependency cycles, and find orphaned code. Node records also include metrics such as lines of code and export counts, status information, and Git metadata.

How it works

Analysis uses ts-morph and the TypeScript compiler rather than text-only matching. An edge is retained only when it resolves to an indexed node, which makes the graph conservative about unresolved relationships. Workspace discovery handles both single-package repositories and monorepos using pnpm, Turbo, or Lerna conventions.

The MCP surface runs over stdio and provides 14 tools: index_repo, get_graph, get_node, who_renders, who_calls, find_references, blast_radius, find_cycles, find_orphans, search_nodes, get_context_pack, build_embeddings, semantic_search, and open_explorer. The same engine also supports a CLI, an HTTP and WebSocket service, and a bundled 3D explorer with a chatbot, but those are separate ways to consume the index.

Semantic search uses the local Xenova/all-MiniLM-L6-v2 model through transformers.js. The README states that the model runs locally and that lexical search is used if the model is unavailable, so an external API key is not required for this feature.

Setup and configuration

For source development, Nishant-Chaudhary5338/mcp-code-indexer MCP server requires Node 20.19 or newer and pnpm 10. The documented source workflow is:

• Run pnpm install.
• Run pnpm build; the build must finish before running the source CLI or server.
• Add the built MCP entry point to an MCP client using the repository’s documented Node command: node "$(pwd)/tools/code-indexer/build/code-indexer/src/index.js".

The published npm package is named code-graph-indexer. Its documented package-level quickstart uses npx code-graph-indexer ui --root . for the explorer. The provided material does not specify a separate published-package command for launching only the MCP stdio server, so the source invocation is the grounded MCP setup path here.

Limitations and notes

The analyzer is described for TypeScript, React, and Next.js repositories; the material does not establish support for other languages. Source execution depends on a completed build. Semantic embeddings are local, but they require the model to be installed or available locally; otherwise semantic search falls back to lexical search.

The HTTP service binds to 127.0.0.1 and has unauthenticated, mutating endpoints. It should not be exposed beyond the local host. The MCP interface itself is presented as a stdio server, and no credentials are listed as required.

The graph can become stale when files change unless the repository is reindexed. The separate HTTP/WebSocket runtime includes file watching and live graph patches, while the MCP tool list includes index_repo for indexing through the agent interface.