Skip to main content
Back

Context+ Gives AI Coding Agents a Map to Navigate Massive Codebases

Context+ is a free, open-source MCP server that turns any large codebase into a semantic graph your AI agent can actually navigate accurately.

  • What it is: Context+ is a free, open-source MCP server that indexes your codebase into a semantic graph for AI agents.
  • Core tech: Combines Tree-sitter AST parsing (43 file extensions), spectral clustering, RAG memory graph, and Obsidian-style wikilinks into 17 MCP tools.
  • Key tools: Semantic code search, blast radius analysis, shadow restore points, and a gated propose_commit that validates before writing.
  • Zero install: Add a JSON block to your IDE MCP config; works with Claude Code, Cursor, VS Code, Windsurf, and OpenCode via npx or bunx.
  • Embedding options: Runs fully offline with local Ollama models, or connects to OpenAI, Gemini (free tier), Groq, or any OpenAI-compatible endpoint.
  • Cost: MIT licensed and completely free; compute cost depends on your chosen embedding backend.

Context+ is an open-source MCP server that tackles one of the most frustrating problems in AI-assisted coding: the bigger your codebase, the less useful your AI agent becomes. When a project grows past a few hundred files, agents start hallucinating file paths, missing dependencies, and making changes that silently break things elsewhere. Context+ approaches this by indexing your entire codebase into a searchable, hierarchical feature graph before your agent ever touches a file.

The core problem it solves

Most AI coding agents navigate codebases the same way a junior dev would on day one: they grep for keywords, read full files, and hope they found the right thing. A typical agent reads full file contents to understand code structure -- a 500-line file gets read entirely just to find 3 relevant function signatures, wasting tokens and limiting effective coverage of large codebases. Context+ flips this around by giving the agent structural awareness upfront, so it never needs to blindly read files to understand what's in them.

Context+ guarantees minimal context bloat. It gives your agent deep semantic understanding of your codebase, from AST parsing and symbol navigation to blast radius analysis and commit validation.

How it's built under the hood

The server has three layers built with TypeScript over stdio using the Model Context Protocol SDK: a Core layer handling multi-language AST parsing (tree-sitter, 43 extensions), gitignore-aware traversal, Ollama vector embeddings with disk cache, wikilink hub graph, and in-memory property graph with decay scoring; a Tools layer exposing 17 MCP tools; and a Git layer providing a shadow restore point system for undo without touching git history.

A few key techniques power the whole thing:

  • Tree-sitter AST parsing -- Tree-sitter is a high-performance incremental parser (originally built at GitHub) that produces concrete syntax trees for any supported language. Context+ uses it for multi-language symbol extraction with regex fallback, supporting 14+ languages and 43 file extensions via WASM grammar loaders.