What it does

Context+ is an MCP server that plugs into Claude Code, Cursor, VS Code, and Windsurf. It parses your codebase with Tree-sitter, embeds symbols and file headers, clusters related files by meaning, and exposes 17 tools for semantic search, blast-radius tracing, static analysis, and shadow git restore points. The goal: give your coding agent structured context instead of dumping raw files into the prompt.

The interesting bit

The memory graph is the unusual part. Nodes (files, symbols, concepts) get auto-linked when cosine similarity hits ≥0.72, and edges decay over time via e^(-λt) scoring. There’s also an Obsidian-style feature hub — .md files with [[wikilinks]] that map human-readable features to code. It’s part documentation, part embedding-backed knowledge base, and it lives inside the repo.

Key highlights

• Semantic navigation over structural search: semantic_navigate clusters files by meaning using spectral clustering, not just folder hierarchy or string matching.
• Blast-radius tracing: get_blast_radius follows imports and call sites to show every file and line affected by changing a symbol.
• Shadow undo without git pollution: propose_commit creates restore points before AI writes code; undo_change reverts them without touching git history.
• Offline-by-default, cloud-optional: Ollama is the default embedding backend; OpenAI-compatible APIs (including Gemini free tier) are a config switch away.
• Realtime incremental cache: .mcp_data/ stores embeddings and refreshes only changed files/functions on the fly.

Caveats

• Requires running your own Ollama instance or supplying an API key for embeddings; not zero-config if you lack a GPU or API budget.
• The README claims “99% accuracy” but never defines the metric or cites a benchmark — treat it as aspiration, not measurement.
• Static analysis supports TypeScript, Python, Rust, and Go; other languages get AST parsing and search but not linting/dead-code detection.

Verdict

Worth trying if you’re hitting context-window limits or watching your agent hallucinate file relationships in large repos. Skip it if you want a drop-in replacement for grep or if your codebase is small enough that full-file inclusion already works fine.