aiguicai/mcp-gateway/codegraph

CodeGraph

CodeGraph builds a local semantic graph of a codebase and exposes CLI commands for indexing, symbol search, task context, file structure, call relationships, impact analysis, and affected-test discovery.

Use this tool when a repository has or should have a .codegraph/ index and the task needs structural code intelligence: "where is this symbol", "how does this area work", "who calls this", "what might break if I change it", or "which tests are affected by these files".

IMPORTANT: How to Use This Tool

This is a dedicated built-in tool named codegraph. You MUST follow these rules:

1. Do NOT use `shell_command` to run CodeGraph commands. The gateway blocks

attempts to run codegraph or npx -y @colbymchenry/codegraph through shell_command when this tool is enabled. Always use this codegraph tool.

1. Executing commands (skillToken required):

• Every exec value MUST start with the codegraph prefix.
• You MUST include the skillToken returned from reading this document.
• Example: codegraph({"exec": "codegraph status", "cwd": "D:/project", "skillToken": "<token>"})

1. The gateway runs CodeGraph through npx. You write codegraph ..., and

the gateway executes npx -y @colbymchenry/codegraph ... directly without a shell wrapper.

1. This is NOT the CodeGraph MCP server integration. Do not run

codegraph serve --mcp. This built-in tool only supports selected CLI commands. Upstream MCP-only tools such as codegraph_trace, codegraph_explore, and codegraph_node are not available through this built-in tool.

Prerequisites

Node.js/npm must be installed so npx is available. The upstream npm package currently requires a supported Node runtime, and the first call may download or refresh the @colbymchenry/codegraph npm package cache.

Always set cwd to the project root or a directory inside the configured allowed directories. CodeGraph stores its local SQLite index under .codegraph/.

Allowed Commands

Use these commands through the codegraph tool:

codegraph --version
codegraph help [subcommand]
codegraph init [path] -i
codegraph index [path]
codegraph sync [path]
codegraph status [path]
codegraph query <search> [--kind <kind>] [--limit <n>] [--json]
codegraph files [path] [--format <format>] [--filter <glob>] [--max-depth <n>] [--json]
codegraph context <task> [--max-nodes <n>] [--max-code <n>] [--no-code] [--format markdown|json]
codegraph callers <symbol> [--limit <n>] [--json]
codegraph callees <symbol> [--limit <n>] [--json]
codegraph impact <symbol> [--depth <n>] [--json]
codegraph affected <files...> [--depth <n>] [--filter <glob>] [--json] [--quiet]

Blocked commands:

codegraph serve --mcp
codegraph install
codegraph uninstall
codegraph uninit
codegraph unlock

Do not use the no-argument codegraph installer through this tool.

Typical Workflow

Check whether the project is initialized:

codegraph({"exec": "codegraph status", "cwd": "D:/project", "skillToken": "<token>"})

Initialize and index a project when needed:

codegraph({"exec": "codegraph init -i", "cwd": "D:/project", "skillToken": "<token>"})

Search for a symbol or inspect a broad area:

codegraph({"exec": "codegraph query AuthService", "cwd": "D:/project", "skillToken": "<token>"})
codegraph({"exec": "codegraph context \"how authentication middleware works\" --max-nodes 30", "cwd": "D:/project", "skillToken": "<token>"})

Inspect call relationships:

codegraph({"exec": "codegraph callers AuthService --json", "cwd": "D:/project", "skillToken": "<token>"})
codegraph({"exec": "codegraph callees AuthService --limit 30", "cwd": "D:/project", "skillToken": "<token>"})
codegraph({"exec": "codegraph impact AuthService --depth 2", "cwd": "D:/project", "skillToken": "<token>"})

Find affected tests or files:

codegraph({"exec": "codegraph affected src/auth.ts", "cwd": "D:/project", "skillToken": "<token>"})

Notes

• Prefer codegraph status before assuming the project is indexed.
• Use codegraph sync after large file changes if the index looks stale.
• Use codegraph context first for architecture or "how does this work"

questions; use query for exact symbol lookup; use callers, callees, or impact for change planning.

• Prefer --json when you need structured output for follow-up reasoning.
• Do not use shell redirection, pipes, or affected --stdin with this tool. If

you need affected tests, pass file paths as arguments. If you need to persist output, use multi_edit_file for files and read_file for inspection.