MCP Context Graph
A self-contained, in-memory graph database for AI Agents. Provides semantic code understanding through the Model Context Protocol (MCP).
README
MCP Context Graph
A self-contained, in-memory graph database for AI Agents. Provides semantic code understanding through the Model Context Protocol (MCP).
Why This Tool?
AI coding assistants often struggle with large codebases. They either:
- Read entire files (expensive, hits context limits)
- Grep for text (misses semantic relationships)
- Lose track of where functions are called from
MCP Context Graph solves this by building a semantic graph of your codebase that the AI can query efficiently.
Key Differentiators
| Feature | Benefit |
|---|---|
| Token-Level Source Maps | Stores minified signatures but can expand to exact original source with character-accurate mapping |
| Semantic Call Graph | "Who calls this function?" answered in milliseconds, not by reading every file |
| Polyglot Engine | Python, TypeScript, JavaScript parsed with tree-sitter grammars |
| Zero Install | Run instantly with uvx - no pip install, no dependencies to manage |
| MCP Native | Built for AI agents - exposes tools through Model Context Protocol |
| Smart Exclusions | Respects .gitignore, skips node_modules, .venv, __pycache__ automatically |
| Lazy Ingestion | Files indexed on-demand, auto-refreshed when modified |
How It Works
Your Codebase MCP Context Graph AI Agent
| | |
| ──── tree-sitter parse ────> | |
| | |
| <── minified signatures ──── | |
| + source maps | |
| | |
| | <── find_callers("fn") ── |
| | ──> [caller1, caller2] ───> |
| | |
| | <── expand_source(id) ──── |
| | ──> exact original code ─> |
The AI gets fast semantic queries without loading entire files. When it needs the full source, it can expand specific symbols using source maps.
Features
- Polyglot Support: Parses Python, TypeScript, and JavaScript using tree-sitter
- Source Maps: Token-level provenance for precise context extraction
- Lazy Ingestion: Files are indexed on-demand and refreshed automatically
- Zero Configuration: Works instantly via
uvxwith sensible defaults
Installation
Primary Method (Recommended)
No installation required. Run directly with uvx:
uvx mcp-context-graph /path/to/your/project
Alternative: Install via pip/uv
# Using uv
uv pip install mcp-context-graph
# Using pip
pip install mcp-context-graph
How to Use
MCP Context Graph runs as an MCP server that AI assistants can connect to. Configure it in your MCP client:
A. In Cline (VS Code)
Add to your Cline MCP settings (cline_mcp_settings.json):
{
"mcpServers": {
"context-graph": {
"command": "uvx",
"args": ["mcp-context-graph", "."],
"autoApprove": []
}
}
}
The . argument uses the current workspace directory as the project root.
B. In Claude Desktop
Add to your Claude Desktop configuration (claude_desktop_config.json):
{
"mcpServers": {
"context-graph": {
"command": "uvx",
"args": ["mcp-context-graph", "/absolute/path/to/your/project"]
}
}
}
Note: Claude Desktop requires an absolute path. Relative paths like . will not work correctly.
Available Tools
Once connected, the following tools are available to the AI agent:
| Tool | Description |
|---|---|
index_project |
Full scan of project directory. Builds the code graph. |
find_symbol |
Find function/class definitions by name. |
find_callers |
Find all locations that call a specific function. |
get_context |
Get a context window around a symbol (callers, callees). |
expand_source |
De-minify a node using source maps for full original code. |
debug_dump_graph |
Export the graph as Mermaid, JSON, or DOT format. |
Example Workflow
-
AI indexes the project:
index_project() -
AI finds a function definition:
find_symbol(name="calculate_tax", include_calls=true) -
AI explores what calls that function:
find_callers(name="calculate_tax") -
AI gets broader context (2 levels of connections):
get_context(name="calculate_tax", depth=2, format="markdown") -
AI expands a specific symbol to see full source:
expand_source(symbol_id="abc123")
Source Maps: The Secret Sauce
Most code indexers store either:
- Full source code (expensive)
- Just symbol names (loses context)
MCP Context Graph stores minified signatures with character-accurate source maps:
# Original (45 bytes)
def calculate_tax(amount: float, rate: float) -> float:
"""Calculate tax for the given amount."""
return amount * rate
# Stored signature (minified)
def calculate_tax(amount: float, rate: float) -> float: ...
# Source map
Segment(minified: 0-52, original: 0-52) # signature preserved exactly
When the AI needs the full implementation, expand_source maps the minified offsets back to the original file and returns exact source code.
Benchmarks
Benchmark Results
| Metric | Value |
|---|---|
| Files processed | 46 |
| Nodes created | 624 |
| Raw source size | 341.2 KB |
| Minified size | 24.6 KB |
| Compression ratio | 13.9x |
| Ingest time | 46 ms |
| find_definition | 5 μs |
| find_callers | 9 μs |
| get_context(depth=2) | 5 μs |
Cost Efficiency (USD per 1,000 calls)
Token counts via OpenRouter API (exact)
| Model | Full File | Graph | Savings | % |
|---|---|---|---|---|
| gpt-5.2 | $184.05 | $15.40 | $168.65 | 91.6% |
| gpt-4o-mini | $11.04 | $0.92 | $10.12 | 91.6% |
| claude-sonnet-4.5 | $281.17 | $25.41 | $255.76 | 91.0% |
| gemini-2.5-pro | $320.24 | $28.78 | $291.46 | 91.0% |
Key takeaways:
- The AI can work with a 14x smaller representation of your codebase
- Queries complete in microseconds, not seconds
- 91-93% cost savings on context token usage across all major LLM providers
- Full source is always available on-demand via source maps
Run benchmarks on your own project:
OPENROUTER_API_KEY=your-key uv run python benchmarks/run_benchmarks.py /path/to/project
CLI Usage
# Index current directory
uvx mcp-context-graph .
# Index a specific project
uvx mcp-context-graph /path/to/project
# Show version
uvx mcp-context-graph --version
Development
Prerequisites
- Python 3.12+
- uv package manager
Setup
# Clone the repository
git clone https://github.com/padobrik/mcp-context-graph.git
cd mcp-context-graph
# Install dependencies
uv sync
# Run tests
uv run pytest tests/
# Run linting
uv run ruff check .
# Run type checking
uv run mypy src/
Project Structure
src/mcp_context_graph/
core/ # Graph data structures (Node, Edge, Graph)
ingest/ # File parsing and graph construction
languages/ # Language-specific configurations (Python, TypeScript)
mcp/ # MCP server and tool handlers
provenance/ # Source map implementation
License
MIT License. See LICENSE for details.
Recommended Servers
playwright-mcp
A Model Context Protocol server that enables LLMs to interact with web pages through structured accessibility snapshots without requiring vision models or screenshots.
Magic Component Platform (MCP)
An AI-powered tool that generates modern UI components from natural language descriptions, integrating with popular IDEs to streamline UI development workflow.
Audiense Insights MCP Server
Enables interaction with Audiense Insights accounts via the Model Context Protocol, facilitating the extraction and analysis of marketing insights and audience data including demographics, behavior, and influencer engagement.
VeyraX MCP
Single MCP tool to connect all your favorite tools: Gmail, Calendar and 40 more.
graphlit-mcp-server
The Model Context Protocol (MCP) Server enables integration between MCP clients and the Graphlit service. Ingest anything from Slack to Gmail to podcast feeds, in addition to web crawling, into a Graphlit project - and then retrieve relevant contents from the MCP client.
Kagi MCP Server
An MCP server that integrates Kagi search capabilities with Claude AI, enabling Claude to perform real-time web searches when answering questions that require up-to-date information.
E2B
Using MCP to run code via e2b.
Neon Database
MCP server for interacting with Neon Management API and databases
Exa Search
A Model Context Protocol (MCP) server lets AI assistants like Claude use the Exa AI Search API for web searches. This setup allows AI models to get real-time web information in a safe and controlled way.
Qdrant Server
This repository is an example of how to create a MCP server for Qdrant, a vector search engine.