Code Review Graph: AI Token Optimization for Code Reviews

With the growing adoption of AI tools for code generation and review, the issue of unnecessary token consumption has become increasingly evident in development environments. The frequent re-reading of entire codebases leads to inefficiencies in both cost and latency. To address this challenge, the open-source tool “code-review-graph,” available on GitHub, offers a potential solution.

The Problem

In AI-assisted code reviews, there is a tendency to reload the entire codebase to understand the context of changes. For large-scale projects comprising hundreds of files, this results in excessive token consumption relative to the actual number of modified lines, making the cost per review request significant.

The project’s motto, “Stop burning tokens,” succinctly captures this issue. While AI-powered code reviews have reached a practical level of utility, managing their running costs is critical for fostering a sustainable development workflow.

How Code-Review-Graph Works

Code-review-graph leverages the language parsing framework Tree-sitter to build a structural map of the codebase. When changes occur, it extracts only the affected areas as a diff and transmits them to AI assistants via the Model Context Protocol (MCP).

Traditional methods often send entire files as context. In contrast, code-review-graph creates a graph of the code’s structural information—such as function dependencies and class inheritance hierarchies—and identifies the minimal scope impacted by the changes. This approach significantly reduces the number of tokens sent to the AI.

The tool employs incremental change tracking. After the initial analysis of the entire codebase, only modified files are reprocessed. This allows even large-scale projects with 500+ files to build the initial structure within approximately 10 seconds, with subsequent updates automated via watch mode.

Installation and Supported Platforms

The tool is designed for straightforward installation. It requires Python 3.10 or later and can be installed using pip install code-review-graph or pipx install code-review-graph. Running the install command automatically detects available AI development tools and configures MCP settings for each platform.

Supported platforms include:

• GitHub Copilot (VS Code)
• GitHub Copilot CLI
• Cursor
• Claude Code
• Gemini CLI
• Codex
• Kiro
• CodeBuddy

If you want to target specific platforms, you can use the --platform option during installation. The tool prioritizes the use of uvx for faster package management but falls back to the code-review-graph command if unavailable.

Uninstallation is equally user-friendly, offering a --dry-run preview option before execution. Only files and entries owned by CRG (code-review-graph) are removed, ensuring other MCP servers and hooks remain unaffected.

Build Process and Workflow

The code-review-graph build command parses the codebase using Tree-sitter, storing structural elements like function definitions, class structures, and variable scopes in a graph database.

The typical workflow is as follows:

1. Run code-review-graph install in the repository root.
2. Execute code-review-graph build to generate the structural map.
3. Instruct your AI assistant: “Build the code review graph for this project.”
4. Use watch mode to continuously monitor and update changes.

By integrating with Git hooks or platform-specific hooks, the graph can automatically update during commits or pushes.

Measurable Token Savings

Although the project description offers a bold analogy likening its innovation to “the compact disc of Star Trek,” its impact can be assessed through quantitative benchmarks. These comparative studies with similar tools are part of the roadmap and expected to be released soon.

Typically, using structural code graphs for diff extraction can reduce token consumption by one or two orders of magnitude compared to sending entire files. The effect is particularly pronounced in large-scale monorepo projects.

The Significance of MCP Adoption

One noteworthy aspect of code-review-graph is its adoption of MCP, a protocol proposed by Anthropic for standardizing integration between AI assistants and external tools or data sources.

Previously, the method for passing codebase context to AI varied across tools, making standardization difficult. By using MCP, developers can share a single graph data structure across multiple AI development tools, reducing transition costs when switching AI assistants.

Competitive Differentiation

Although other tools, such as Sourcegraph Cody and Continue.dev, also leverage structural code analysis for AI reviews, code-review-graph sets itself apart with platform independence and automated setup.

The install command stands out by automatically detecting multiple AI tools and configuring MCP settings for each, simplifying the setup process. Developers can configure their entire environment with a single command.

Additionally, the project’s foundation in Python and its compatibility with pip/pipx make it particularly accessible to developers familiar with the Python ecosystem. It also supports uvx for faster package installation.

As noted in this related article on JavaScript AI code generation tools, AI-powered code generation and review tools are rapidly advancing. Code-review-graph targets a specific niche of “cost optimization” amidst this broader evolution.

Technical Constraints and Considerations

Since it relies on Tree-sitter, code-review-graph is limited to the languages supported by this parsing framework. While major programming languages (Python, JavaScript, TypeScript, Rust, Go, Java, C/C++) are covered, niche languages or in-house DSLs are not.

Additionally, as the code structure graph is based on static analysis, it cannot capture dynamic behaviors or runtime dependencies. For codebases heavily reliant on dependency injection or reflection, the graph may not fully represent actual dependencies.

Moreover, the tool is designed solely as a context provider and does not include features like automated code fixes or refactoring suggestions.

Licensing and Development

The project is hosted on GitHub under an ARR (All Rights Reserved) license. Developers should consult the repository for details on commercial use and other conditions.

The creator, tirth8205, plans to continue expanding the tool’s functionality, including publishing benchmark results and supporting additional platforms, as outlined in the project’s roadmap.

Context in the AI Code Review Market

As of 2026, AI code review tools are reaching maturity, with products like GitHub Copilot’s review functionality, Amazon CodeGuru Reviewer, and GitLab Code Suggestions competing in the space.

To stand out, tools must go beyond “code generation” to optimize the entire development process. Code-review-graph aims to complement existing AI tools by addressing the specific issue of token consumption.

Managing AI-related costs is becoming a critical concern for CTOs and VPoEs, especially in organizations with high code review frequency. The token savings offered by code-review-graph can directly impact ROI evaluations.

First Steps After Installation

After installation, developers can begin using the tool with the following steps:

1. Run the initial build command to analyze the entire codebase.
2. Use watch mode for incremental updates as changes occur.
3. Guide the AI assistant with a simple command: “Build the code review graph for this project.”

When requesting code reviews, the AI assistant uses the graph data provided by code-review-graph to accurately identify the impact of changes. There’s no need to instruct the AI to read through all related code, as was previously required.

Editorial Opinion

Optimizing token consumption in AI code reviews is not merely about reducing costs. Even with larger context windows, providing unnecessary information to the AI increases noise and degrades output quality. Code-review-graph’s structural graphs are commendable for automating the “selection of relevant information.”

In the short term, the tool is likely to be adopted by mid-sized companies and startups concerned with AI tool costs. In the long run, if context optimization tools like this become standardized, they could reshape the architecture of AI development tools. If IDEs and CLI tools begin to natively support MCP and graph data sharing becomes the norm, the quality of AI code reviews could reach a new level.

However, the ARR license could hinder widespread adoption, as developers may prefer open-source alternatives. Increased competition could further drive innovation in this space.

The question remains: is token optimization in AI code reviews merely a cost-saving measure, or is it a fundamental way to improve quality? The editorial team believes that providing structural understanding to AI has the potential to achieve the latter.

References

• “tirth8205 /

  code-review-graph", by **tirth8205** — GitHub Trending, 2026-07-21 (ARR)

• Source URL: https://github.com/tirth8205/code-review-graph