CodeGraph: Cut AI Agent Token Costs in Half with Local Code Knowledge Graphs

AI programming assistants spend massive tokens each new session rummaging through codebases — a hidden cost far higher than most realize. CodeGraph parses code structure with Tree-Sitter, generates a local SQLite knowledge graph, and exposes it via MCP so agents can query directly, cutting tool calls by 70%+ and token costs by nearly half, all while code never leaves the local machine.

Every new AI coding assistant session — whether Claude Code or Cursor — follows the same pattern: ask a question, then the agent loops through ls, grep, read_file, repeat to "explore" the codebase. You pay token costs for that entire retrieval process.

This "find the path first, then do the work" flow is the most overlooked cost. We're making high-capability models repeatedly do retrieval grunt work. CodeGraph's approach is direct: give the codebase a pre-built structured map via MCP (Model Context Protocol) so the agent skips the detour and jumps straight to problem solving. On average, tool calls drop over 70% and token costs fall by nearly half.

The Engineering Behind the Graph

CodeGraph's implementation is a stable combination of a few engineering layers.

1. Tree-Sitter Parsing Pipeline

CodeGraph relies on Tree-Sitter at its core. Unlike regex-driven search, it takes a syntactic parsing route, enabling more accurate identification of symbols and structural relationships in code.

2. MCP Infrastructure Bridge

CodeGraph communicates with AI agents through MCP, built on JSON-RPC, standardizing the "question-query-return" chain.

3. Optimized SQLite Graph Architecture

Extracted symbols and relationships are written into a local SQLite graph, providing fast queries and easy local iteration.

Core Advantages: Density and Privacy

Beyond speed, CodeGraph delivers two critical benefits in real projects: context density and privacy .

Context Density: Better Data, Fewer Tokens

Traditional agents often suffer "context bloat": to answer a small question, they stuff large source files into context. CodeGraph returns only relevant symbols and relationships, concentrating information and saving tokens.

Privacy: Code Stays Yours

In an era of cloud services, privacy is unavoidable. CodeGraph runs inside the local boundary : source code and the SQLite graph remain on your machine, and agents query via a local service — no code exposed to external indexing or training pipelines.

How CodeGraph Works

CodeGraph's efficiency comes from a four-stage pipeline: convert raw code into a queryable knowledge graph, then expose query capability to the agent.

Quick Start

If you want to try it immediately, the process is short — just a few steps to get running.

1. Installation

CodeGraph is a lightweight CLI tool, installable globally via npm:

npm install -g @colbymchenry/codegraph

2. Index the Codebase

Enter the project directory and run init. CodeGraph scans files, parses syntax structures, and generates the local SQLite graph.

codegraph init -i

3. Connect AI Agent

After indexing, bridge your favorite AI coding assistant (e.g., Claude Code or Cursor) via MCP. The agent can then query the graph directly, avoiding another round of inefficient exploration.

codegraph install

Collaboration Workflow: Developer, Agent, and Graph

CodeGraph's value shines in three-way collaboration: the developer sets the goal, the agent does reasoning, and CodeGraph supplies structured context. Solutions produced this way tend to align more closely with the codebase's actual state.

Agent Intelligence Landscape: CodeGraph and Other Tools

CodeGraph excels at structural mapping, but it doesn't exist in isolation. A complete toolchain includes lat.md [1] and graphify [2].

CodeGraph End-to-End Process

The end-to-end process emphasizes three points: local-first, privacy-controlled, and easy integration with existing coding assistants.

Through symbol and relationship mapping, CodeGraph lets agents grasp code structure faster. Combined with the MCP command set, agents can navigate complex repos and perform impact analysis, surfacing dependency risks early.

Key Takeaways

CodeGraph addresses a very practical problem: agents spend too long "finding information." It lowers that overhead with a pre-indexed map, letting agents enter the real execution phase sooner.

You can place CodeGraph into your own agent tool stack . Graphify leans toward the semantic layer, lat.md toward intent management, and CodeGraph handles the structural layer. Together, they improve agent stability and efficiency in complex projects.

The core idea is simple: the agent's most expensive action is often "searching everywhere"; CodeGraph builds the map first, so the agent can "follow the map."

References

[1] lat.md: https://www.lat.md/

[2] graphify: https://graphify.net/

[3] www.DeepNoMind.com: https://www.deepnomind.com