Code graph

The environment every thread runs in has CodeGraph installed: a local index of the Kotlin, Swift and TypeScript code (symbols, calls, who depends on what). It answers "how does X work" or "what breaks if I change Y" in one call instead of many greps and file reads.

Use it

• Main thread: call the codegraph_explore MCP tool with a question or a symbol or file name. It returns the relevant source, the call paths and the blast radius (callers and tests).
• Subagents, or if the MCP tool is missing: run codegraph explore "<question>" or codegraph node <symbol|file> in Bash. If the repo has no index yet, run codegraph init once (about 6 s here).
• Use it to find code and see impact before a change. Still read the files you edit, and still follow CONTEXT.md for naming. The graph knows code, not decisions: ADRs and docs stay the source of truth.
• Never commit .codegraph/; it ignores itself.

Where it is set up

• Install and index: the environment's setup script installs codegraph and runs codegraph init in the repo folder, so every new thread starts with an index.
• MCP server: the environment's setup script registers it at user level with claude mcp add --scope user codegraph -- codegraph serve --mcp. That gives the main thread the codegraph_explore tool and keeps the index in sync as files change. claude mcp list should show codegraph as connected. If the tool is missing in a thread, say so and use the CLI.
• Why not a repo .mcp.json: a project-scoped server stays "Pending approval" until someone approves it, and a cloud thread cannot answer that prompt, even with enabledMcpjsonServers set. A user-level server needs no approval.
• Copilot cloud agent: .github/workflows/copilot-setup-steps.yml installs codegraph and runs codegraph init. It has no MCP server, so it uses the CLI.
• To change the version or the install, edit the environment, and the version in copilot-setup-steps.yml.