Back to list

CodeGraph: A Searchable Map of Your Codebase for AI Coding Agents

A local-first map of your whole codebase, so your coding agent answers in one question instead of digging through files. More than 77,000 stars.

Emmett Webb
Author
@Emmett Webb
Project ShowcaseDeveloper ToolsOnlineclaude-code
0
0

Project intro

When a coding agent needs to understand your code, it usually does it the slow way: grep, open a file, grep again, open another file — dozens of tool calls before it even starts the real work. CodeGraph builds a complete map of the codebase first: every symbol, every call between files, every dependency. The agent asks one question and gets the exact code it needs, plus the paths between the pieces and what a change would touch. More than 77,000 people have starred it on GitHub.

It takes three steps to start. Install the CLI with one command, run codegraph install to wire it into the agents you already use — Claude Code, Cursor, Codex, OpenCode, Gemini, Kiro, Copilot and a dozen more — then codegraph init builds the map for a project. A Rust kernel does the parsing across more than 35 languages, from TypeScript and Python to Go, Java, Rust and Swift, and the map stays on your machine: everything runs locally, and it updates itself every time a file changes, so the index is never out of date.

The difference is measurable, and the author measured it in public. Across seven real open-source codebases — VS Code, Excalidraw, Django, Tokio and three more — an agent answering one architecture question with CodeGraph used 88% fewer tool calls, 62% fewer tokens and 44% less cost than the same agent digging through files, and it never opened a single file. It can also follow jumps a text search cannot, like dynamic dispatch, and show the blast radius of a change before you make it.

Two honest notes. This is a mature tool, not a fresh launch: the latest release is v1.6.2 from 3 October 2026, and the newest code landed on 1 November, so do not expect a new feature every week. And the author is upfront about a trade-off: CodeGraph's answers are dense, so in long sessions they leave about 80% more text sitting in the agent's context window than a file-reading agent leaves behind. A hosted version that reviews pull requests is on a waitlist, but the local tool is free, MIT licensed, and ready to install.

Actions

Visit project URL

Tech stack

typescriptrust

Project info

Project category
Developer Tools
Project status
Online
Collaboration intent
None

Comments (0)

ME
0/1000
Loading comments...