CodeGraph

CodeGraph is a local code knowledge graph for AI coding agents. It pre-indexes a repository into a SQLite-backed graph of symbols, files, functions, classes, imports, call relationships, inheritance links, framework routes, and searchable metadata. Agents such as Claude Code, Codex, Cursor, OpenCode, and similar developer assistants can query that graph instead of repeatedly scanning the project with grep, glob, and file reads. The core benefit is lower discovery cost. Large codebases force agents to spend many tool calls just figuring out where logic lives. CodeGraph turns that discovery phase into indexed lookups. The README reports benchmark medians across seven real open-source repositories, including VS Code, Excalidraw, Django, Tokio, OkHttp, Gin, and Alamofire. In those tests, runs with CodeGraph were reported as 35% cheaper on average, with 59% fewer tokens, 49% faster answers, and 70% fewer tool calls. CodeGraph is designed to run locally. It ships as a self-contained bundle with its own runtime, so users do not need to compile native dependencies before trying it. Installation can be done through the project install script or with npm/npx. After running codegraph init in a repository, the tool creates a .codegraph directory and indexes the project for supported agent integrations. For builders, CodeGraph is most useful on repositories where repeated code exploration is expensive: monorepos, mature frameworks, and unfamiliar client codebases. It does not replace tests or source reading, and indexed data can become stale if the project changes without re-indexing. But as an agent acceleration layer, it gives coding assistants a faster map of the system and can reduce unnecessary context spending while keeping code data local. A typical workflow is to initialize CodeGraph inside a repository before handing a task to an AI coding agent. The agent can then ask structural questions such as where a symbol is defined, which functions call it, how a route reaches a handler, or which files participate in a feature. That kind of map is valuable when the model needs to make a change without reading hundreds of unrelated files. The local-first design is important for companies that cannot send full repository context to external indexing services. CodeGraph still works with AI agents that call models, but the code index itself remains on the developer machine. Teams should re-index after large refactors and treat the graph as a navigation aid rather than a source of truth. If tests, compilers, or source files disagree with the index, the source wins. Used correctly, CodeGraph is a sharp productivity layer for agent-driven maintenance, code review, onboarding, and architecture questions.