Reduce token usage and time for AI coding agents by indexing code structure locally
Instead of having agents guess call relationships by scanning files one by one, they retrieve only the necessary code from a pre-drawn graph. CodeGraph analyzes project symbols and dependencies, stores them locally, and delivers the source and paths needed for an answer in one go when an agent asks a question. This reduces token consumption and the number of tool calls.
Getting direct answers instead of exploring
Previously, AI agents had to slowly explore the codebase using grep or file-reading tools. CodeGraph skips this process, allowing agents to obtain relevant source, call paths, and change impact scope in a single tool call. A re-measurement in August 2026 confirmed an average 44% cost reduction and 62% token reduction across seven benchmark repositories. The savings were most significant for complex questions that previously required 28 to 43 tool calls for file exploration.
A Rust engine that adapts to machine specifications
CodeGraph runs on a native Rust kernel that supports more than 20 languages. This engine detects the system's actual core count and available memory to automatically adjust the worker pool and cache size. Workstations use a full parallel pipeline, while 2-core VPSs use a tuned pipeline that guarantees completion. When files are saved, only the changed parts are synchronized after 300ms, updating the index in about 0.3 seconds for a project with 4,400 files. All data is stored locally in a SQLite database and is not sent to external services.
Framework routing and cross-platform flow tracking
It recognizes routing files in web frameworks to connect URL patterns with handlers. It understands the route structure of 17 frameworks, including Django, Flask, FastAPI, Express, and NestJS. In iOS, React Native, and Expo projects, it tracks cross-language flows that static analysis misses, such as Swift and Objective-C bridging, TurboModules, and Fabric view components. It also includes invoke calls between the web frontend and Rust core in Tauri desktop apps in the graph. These features help agents understand the entire flow of an application.
What to know before adopting it
CodeGraph only delivers benefits when agents query it directly, so configurations that delegate exploration to file-reading sub-agents may actually incur overhead. Additionally, responses remain in the context window for a long time, occupying about 80% more residual context than file-reading agents when a session ends. This should be considered when running long sessions in small windows. Node.js is not required for installation; agent integration is performed with the codegraph install command, and project indexing with codegraph init. A hosted platform in the beta stage is also being prepared.
