AI coding assistants are flying blind. This is their map.

What it does

CodeGraph ingests your codebase and builds a hybrid knowledge graph—AST nodes, LSP-resolved symbols, module links, and optional dataflow edges—stored with embeddings for semantic search. AI agents query it through four consolidated tools (agentic_context, agentic_impact, agentic_architecture, agentic_quality) that return synthesized answers rather than raw file lists. It runs as an MCP server, so Claude Code, Cursor, and similar agents can call it directly.

The interesting bit

The system auto-selects its reasoning strategy—LATS tree search for open-ended analysis, ReAct for direct lookups, Reflexion for self-correction—based on which agentic tool you invoke and your configured LLM context window. It also tiers its own indexing from fast (bare AST) to full (dataflow + architecture violations), so you trade storage for graph richness depending on the job.

Key highlights

• Four agentic tools that plan, search, and synthesize answers instead of returning file lists.
• Tiered indexing (fast, balanced, full) that toggles LSP resolution, dataflow edges, and architecture boundary checks.
• Context-overflow guards that truncate per-tool results and fail fast if accumulated output exceeds a safe threshold.
• Hybrid search mixing vector similarity, lexical matching, and graph traversal.
• Automatic reasoning-mode switching via the Rig backend, with legacy ReAct and LATS still available.

Caveats

• LSP-backed indexing tiers require manually installed language servers; the tool fails fast if rust-analyzer, gopls, or others are missing, and rustup shims specifically can trip it up.
• The agent is stateless across calls, so your client—not CodeGraph—must manage multi-turn context accumulation.

Verdict

If you’re feeding a large codebase to Claude Code or Cursor and tired of watching it grep in circles, CodeGraph is worth a look. If you just need occasional file search or don’t run an MCP-compatible agent, it’s probably overkill.