Building a Local Code Knowledge Graph with code-review-graph for Claude Code

Why AI needs a code map

Current AI coding assistants (e.g., Claude Code) can answer simple questions in small projects, but when a repository grows to thousands of files or becomes a monorepo the model must repeatedly read large portions of the codebase. This repeated scanning is slow, costly, and fragile because the AI lacks a stable, structured, reusable memory of the project.

Local knowledge graph + MCP

A local code knowledge graph abstracts the repository into nodes (functions, classes, imports, test functions, documentation, images) and edges (call relationships, inheritance, import links, test‑coverage links). The Model Context Protocol (MCP) exposes this graph to Claude Code, allowing the assistant to query the graph instead of raw files.

Local repository
 ↓
code-review-graph builds the graph
 ↓
Local graph data (SQLite)
 ↓
MCP server
 ↓
Claude Code queries nodes, relationships, impact range

How code-review-graph works

Running code-review-graph build first uses Tree-sitter to parse source code into an abstract syntax tree (AST). From the AST it extracts entities and relationships and stores them as graph nodes and edges.

Nodes : functions, classes, import statements, test functions, documentation files, images.

Edges : call relationships, inheritance, import dependencies, test‑coverage links.

For example, if function A calls function B, a call edge is created; if a test file covers a business function, a test‑coverage edge is added.

Local SQLite storage

The generated graph is stored in the project’s .code-review-graph/ directory using SQLite, so the code never leaves the developer’s machine. This design is important for private or sensitive codebases.

Blast‑radius analysis

The core capability of code-review-graph is blast‑radius analysis, which determines the minimal set of files affected by a change. The analysis proceeds in five steps:

Identify the changed files and functions.

Trace all callers of the changed functions.

Trace all dependencies of the changed functions.

Find related test files and test functions.

Compute the smallest affected file set.

The result is sent to Claude Code via MCP, allowing the assistant to work with a focused context instead of the entire repository.

Incremental updates

Full re‑indexing would be impractical, so the tool updates the graph incrementally. On each file save or Git commit it hashes the changed files (SHA‑256) and re‑parses only those files, updating related nodes locally. In practice, a project with 2,900 files can be re‑indexed in under 2 seconds , keeping the graph up‑to‑date without noticeable delay.

Quick‑start guide

Prerequisite : Python 3.10 or newer. python --version Install the tool with one of the following commands:

# Install with pip
pip install code-review-graph

# Or install in an isolated environment with pipx
pipx install code-review-graph

# Detect and configure supported AI platforms
code-review-graph install

# Build the graph for the current project
code-review-graph build

To configure Claude Code exclusively, run:

code-review-graph install --platform claude-code

This writes the MCP configuration to ~/.claude.json or .claude/settings.json. After restarting Claude Code you can ask questions such as:

If I modify the authentication module, which call paths are affected?

CLI workflow

# Build the graph
code-review-graph build

# Incremental update (only changed files)
code-review-graph update

# Show graph statistics
code-review-graph status

# Watch mode – automatically update on file changes
code-review-graph watch

# Visualize the graph in various formats
code-review-graph visualize --format graphml
code-review-graph visualize --format svg
code-review-graph visualize --format obsidian
code-review-graph visualize --format cypher

# Detect change impact and risk score
code-review-graph detect-changes

# Start the MCP server
code-review-graph serve

Excluding unwanted files

Generated code, third‑party libraries, or vendor directories can be excluded with a .code-review-graphignore file placed at the project root. Example content:

generated/**
*.generated.ts
vendor/**
node_modules/**

If the project is a Git repository, code-review-graph automatically respects .gitignore and only indexes files reported by git ls-files.

Windows troubleshooting

A common error on Windows is:

Invalid JSON: EOF while parsing
MCP error -32000: Connection closed

Four checks usually resolve the issue:

Ensure fastmcp version ≥ 3.2.4.

Avoid using a cmd /c wrapper in the MCP configuration.

Call the .exe directly.

Set the environment variable PYTHONUTF8=1.

Minimal MCP server configuration example:

{
  "mcpServers": {
    "code-review-graph": {
      "command": "C:\\path\\to\\venv\\Scripts\\code-review-graph.exe",
      "args": ["serve"],
      "env": {"PYTHONUTF8": "1"}
    }
  }
}

Multi‑repository and daemon mode

For teams maintaining several repositories, the daemon mode can manage all graphs centrally:

# Register repositories
crg-daemon add ~/project-a --alias proj-a
crg-daemon add ~/project-b

# Start the daemon
crg-daemon start

# Check status
crg-daemon status

# View logs for a specific repo
crg-daemon logs --repo proj-a -f

# Stop the daemon
crg-daemon stop

Slash commands in Claude Code

/code-review-graph:build-graph

– build or rebuild the code graph /code-review-graph:review-delta – review changes since the last commit /code-review-graph:review-pr – full PR review with blast‑radius analysis

MCP tools overview

get_minimal_context_tool

– fetch minimal context for a query get_impact_radius_tool – compute the blast radius of a change get_review_context_tool – generate token‑optimized review context query_graph_tool – query callers, callees, tests, imports, inheritance detect_changes_tool – risk scoring and impact analysis semantic_search_nodes_tool – name or semantic search of code entities

Conclusion

AI coding assistants struggle with large codebases because they repeatedly read files, which is slow and expensive. A local, SQLite‑backed code knowledge graph provides stable, structured context that dramatically reduces the amount of code the model must ingest. For small scripts the overhead may not be worth it, but once a project reaches multiple modules and complex call chains, the graph’s benefits become evident.

GitHub repository: https://github.com/tirth8205/code-review-graph