CodeGraph is a locally run code knowledge graph that coding agents query instead of searching a repository file by file. It is aimed at developers who already work inside Claude Code, Cursor, Codex, OpenCode, Gemini, Antigravity, Kiro, GitHub Copilot or Hermes Agent, and who have noticed how much of a session goes into the agent looking for the right file rather than changing it. The project is MIT-licensed, ships as a command-line tool, and keeps the whole index on your machine.
What it does
An agent without an index explores by guessing: grep for a name, open a file, read it, grep again. Every one of those steps costs tokens and a tool call, and the agent can still miss the definition it needed. CodeGraph builds the graph ahead of time — the symbols in a codebase and the relationships between them — and keeps it in step with the code as it changes, so the agent can ask once and get back the slice that matters. The project's own summary of the payoff is deliberately short: fewer tokens, fewer tool calls.
The release covered in the video, 1.6.1, extends the graph to application routes. Next.js, React Router, Expo, Vue and SvelteKit routes are mapped, which means a question like "where does this button go" is answered from the graph instead of from a chain of file reads. The same release follows a page's fetch through to the API endpoint behind it, and records when each call actually fires. The front end and the server work it triggers end up in one connected structure rather than two separate searches.
How it works
Indexing happens once up front, and from then on the index auto-syncs as code changes, so there is no build step to re-run after every edit. The README credits a Rust kernel for the indexing core, while GitHub reports C as the repository's main language — the two do not quite line up, which is worth knowing if you intend to read the source rather than just use the tool.
Everything runs locally. There is no account to create and no code leaving the machine, which is the detail that decides whether a tool like this is usable at all in many workplaces. Distribution is through npm as @colbymchenry/codegraph, with Node.js bundled so a separate Node installation is not required, and releases carry npm provenance plus signed build attestations. The README's badges list Windows, macOS and Linux as supported platforms.
Getting started
Install the npm package, then follow the per-agent setup on the documentation site at colbymchenry.github.io/codegraph, which is also where the supported-agent and supported-platform details live. If an older version is already on your machine, the first line of the README is the whole upgrade path: run codegraph upgrade. Release announcements go out on X as @getcodegraph.
When to use it / when not
This is worth installing when:
- the codebase is large or unfamiliar enough that your agent spends its first several tool calls orienting itself;
- you are working on a JavaScript or TypeScript front end in one of the mapped router frameworks, where the route-to-API link is exactly the thing an agent cannot infer from a single file;
- a local-only tool is the only kind your policies allow.
It is less compelling when the repository is small enough that the relevant files already fit in the agent's context — the graph then saves little, and you still pay for indexing. It is also worth checking the documentation before committing if your stack sits outside the frameworks named for route mapping, since the materials here name five JavaScript frameworks and say nothing about equivalent coverage elsewhere. And because the tool slots into a specific list of agents, a setup built around some other client needs confirming against the supported-agents list first.
The repository appeared in January 2026 and has gathered roughly 73,000 stars and 4,700 forks since, with pushes continuing through the autumn. That is a strong signal of interest, though stars say more about how widely a problem is felt than about how well a particular tool solves it in your repository.
If you drive an agent through a codebase you know well and keep watching it rediscover the layout from scratch, CodeGraph is targeting the right waste. The combination that makes it easy to try is the practical part: one npm install, no service, no code leaving the machine, and a permissive licence, so the test is cheap and reversible. Developers on big JavaScript and TypeScript front ends stand to gain the most from the 1.6.1 routing work, while anyone on a stack outside the mapped frameworks should read the documentation before expecting the same result. Treat it as an index you add to an existing agent workflow, not as a replacement for one.