Codegraph and other tools to turn a codebase into a graph

What was covered

• Why a graph. Code is stored as text, but development questions are about structure, flow and data. Julian walked through a five-phase process: analyse the codebase, extract definitions, resolve links, store, and serve the result to agents or people.
• Kinds of tool. Security analysis, navigation indexes, and code graphs designed to make coding agents work better.
• An experiment with Codegraph. A small Python app with a logging service, two backends and three callers, plus a hard case with deliberate aliases to defeat matching by name. Running codegraph init analyses the code, resolves references and creates a local SQLite database. Dependencies and the impact of a change can then be queried through MCP or the CLI, with confidence up to 0.9 on method-level resolution.
• Claims against results. The creators report 80% fewer tool calls, faster work, fewer tokens and lower cost. In the toy project it used fewer tool calls, opened fewer files, finished a turn sooner and ran faster, but it used more tokens.

What came up

• Why the extra tokens. The tool makes several sequential calls without reading files directly, so the saving would only show on large projects.
• The deck. The presentation was drafted and generated with Opus 5.5 and edited afterwards. Felipe asked for it to be shared in the AI Guild chat to make the community’s work visible.

Next steps

Felipe will install Codegraph and test it on his current project, and the group will pick a tool to debate in the new quarterly format.