ai-codegraph — semantic code intelligence via MCP

What it produces

Nothing written. This skill reads the code graph and returns structured answers: verbatim source grouped by file, call paths (including dynamic dispatch), and blast radius summaries. The agent uses these to answer graph questions without grepping every file.

When to use CodeGraph vs find/grep

Question	Tool	Why
"How does X work?"	codegraph_explore	Follows call chains across files
"What calls function Y?"	codegraph_explore	Graph traversal, follows polymorphism
"If I change this interface, what breaks?"	codegraph_explore	Blast radius via edge traversal
"What does this module depend on?"	codegraph_explore	Import edges in the graph
"Find all TODOs"	grep	Literal text search, faster
"Where is X defined?"	find	Simple path lookup
"Find all files importing X"	grep	Literal text match
"Show me the test for this function"	find + grep	Convention-based lookup

Rule of thumb: graph questions → CodeGraph. Text questions → find/grep.

Steps

1. Check availability — is CodeGraph installed?

   which codegraph && codegraph --version
   

   If missing: npm i -g @colbymchenry/codegraph

2. Check daemon — is the daemon running?

   test -S .codegraph/daemon.sock && echo "running" || echo "not running"
   

   If not running: codegraph serve --mcp &

3. Check index — is the project indexed?

   codegraph status
   

   If not indexed: codegraph init

4. Query — use codegraph_explore MCP tool:

   • query: natural language or symbol/file names
   • maxFiles: default 12, increase for broad questions

5. Interpret results — CodeGraph returns:

   • Verbatim source grouped by file (with line numbers)
   • Call paths (calls, imports, extends, implements)
   • Blast radius summary (what depends on the queried symbol)

Per-surface MCP configuration

CodeGraph is an MCP server, not a governance hook. Each surface reads its own config:

Surface	Config file	Key
Claude Code	~/.claude.json (global) or .mcp.json (project)	mcpServers.codegraph
Cursor	~/.cursor/mcp.json (global) or .cursor/mcp.json (project)	mcpServers.codegraph
Copilot VS Code	.vscode/mcp.json (local) or user mcp.json	servers.codegraph
Copilot CLI	~/.copilot/mcp-config.json	mcpServers.codegraph
opencode	~/.config/opencode/opencode.jsonc	mcp.servers.codegraph
OMP	Import from ~/.claude.json or add manually	—
Pi	pi-codegraph-extension or @isac322/pi-codegraph	—

Stdio launch shape (all surfaces):

{
  "type": "stdio",
  "command": "codegraph",
  "args": ["serve", "--mcp"]
}

Lifecycle

Lane: light Writes: nothing (read-only tool server) Read by: any agent that needs graph queries Dies: on session end Next: none — the agent decides when to use it based on the question type

Anti-patterns

• Do NOT use for literal text search (grep is faster)
• Do NOT use for file discovery (find/glob is faster)
• Do NOT use CodeGraph AND grep for the same question — pick one
• DO remember: CodeGraph leaves 82% more context resident after multi-turn sessions (vendor data: 67k vs 18k tokens after 3 turns). On small repos (<300 files), the cost may outweigh the benefit.

Source

• Repository: https://github.com/colbymchenry/codegraph
• License: MIT
• Stars: ~71k
• Version: 1.6.0 (measured 2026-09-22)
• Integration evidence: .codegraph/codegraph.db in this workspace

Source: ai-engineering (own), Apache-2.0.