MCP Codebase RAG Server
# MCP Codebase RAG Server
> Self-hosted MCP server that adds semantic vector search over your local codebases to any MCP-capable client (GitHub Copilot, Cline, Claude Desktop, etc.).
> **Goal**: Robust RAG for Copilot (or any MCP client) without paying for Cursor/Windsurf.
> **Zero cost. Zero limits. Full control.**
---
## ๐ Overview
### Problem Solved
- GitHub Copilot Pro has an excellent model but limited codebase RAG
- Cursor/Windsurf have good RAG but cost $15โ20/month
- Continue.dev has RAG but doesn't integrate natively with MCP-aware agents
### Solution
This MCP server provides:
1. **Indexing** of local codebases using vector embeddings
2. **Hybrid semantic search** via `search_codebase` tool โ combines dense (embeddings) + sparse (BM25) + RRF fusion
3. **Multi-project** support with isolated ChromaDB collections
4. **Universal integration** with any MCP client
### Tech Stack
| Component | Technology | Why |
|---|---|---|
| Embeddings | **sentence-transformers** (`all-MiniLM-L6-v2`) | Fast, lightweight, 384-dim |
| Code embeddings | **microsoft/unixcoder-base** (optional) | Code-specific model, activated via `EMBEDDING_MODEL` |
| Vector DB | **ChromaDB** | Simple, persistent, zero config |
| Code parsing | **Tree-sitter** + BM25 + RRF | Universal language-agnostic chunking and hybrid search |
| MCP SDK | **modelcontextprotocol/python-sdk** | Official standard |
| Runtime | Python 3.11+ | โ |
---
## ๐ Installation
### Prerequisites
- Python 3.11+
- `pip` or `uv`
### Install
```bash
git clone https://github.com/di5rupt0r/codebase-rag.git
cd codebase-rag
# Install as a package (adds the `codebase-rag` command to ~/.local/bin)
pip install -e .
```
### Health Check
```bash
python scripts/health_check.py
```
Expected output:
```
๐ MCP Codebase RAG Server Health Check
==================================================
Checking Embedding Provider... โ OK (3.03s)
Checking ChromaDB Connection... โ OK (0.14s)
Checking Search Functionality... โ OK (2.36s)
Checking Data Directory... โ OK (0.00s)
==================================================
Health Check Summary: 4/4 checks passed
๐ All systems operational!
```
---
## ๐ Quick Start
### 1. Index a Project
```bash
# Index the current directory
python scripts/index_project.py . --name my-project
# Index a specific path
python scripts/index_project.py ~/projects/api --name api-backend
# Force full reindex
python scripts/index_project.py . --name my-project --force
# Dry run to preview what will be indexed
python scripts/index_project.py . --name my-project --dry-run
```
### 2. Start the MCP Server
#### stdio (default โ for local clients)
```bash
codebase-rag
```
#### HTTP (for remote clients or always-on service)
```bash
MCP_TRANSPORT=streamable-http MCP_PORT=8080 codebase-rag
```
### 3. Configure Your MCP Client
#### VS Code (GitHub Copilot / Cline) โ stdio mode
Add to your VS Code `mcp.json`:
```json
{
"servers": {
"codebase-rag": {
"type": "stdio",
"command": "codebase-rag"
}
}
}
```
#### VS Code โ HTTP mode (when running as a service)
```json
{
"servers": {
"codebase-rag": {
"type": "http",
"url": "http://127.0.0.1:8080/mcp"
}
}
}
```
#### Claude Desktop
```json
{
"mcpServers": {
"codebase-rag": {
"command": "codebase-rag"
}
}
}
```
---
## ๏ฟฝ Search Capabilities
### Hybrid Search Architecture
The server implements a **hybrid search** system that combines:
1. **Dense Search** (Vector Embeddings)
- Semantic similarity using sentence-transformers
- Finds conceptually similar code
- Base: ChromaDB vector similarity
2. **Sparse Search** (BM25)
- Exact lexical term matching
- Finds precise identifiers and keywords
- Base: rank-bm25 with regex tokenization
3. **Reciprocal Rank Fusion** (RRF)
- Intelligent fusion of dense + sparse results
- k=60 (standard literature value)
- Improves both precision and recall
### Search Results
```json
{
"results": [
{
"path": "src/auth.py",
"content": "def authenticate_user(user, password): ...",
"score": 0.0325,
"type": "function",
"name": "authenticate_user",
"line_start": 15,
"line_end": 25
}
],
"total_indexed_chunks": 1247,
"query_time_ms": 23.4,
"search_type": "hybrid_rrf"
}
```
### Performance Characteristics
| Metric | Target | Description |
|---|---|---|
| Tree-sitter parsing | < 50ms/file | Universal language parsing |
| BM25 indexing | < 10ms/query | In-memory reconstruction |
| RRF fusion | < 1ms | In-memory score calculation |
| Total query time | < 100ms | End-to-end hybrid search |
| Memory overhead | < 50MB | For 5k chunks |
### Fallback Behavior
- **Tree-sitter unavailable** โ Line-based chunking
- **BM25 unavailable** โ Dense-only search
- **Both unavailable** โ Original dense search with keyword reranking
## ๏ฟฝ๏ธ MCP Tools
### `search_codebase`
Hybrid semantic search over an indexed project using vector embeddings + BM25 + RRF fusion.
**Input**:
```json
{
"query": "where is the authentication logic?",
"top_k": 5,
"project": "my-project",
"file_types": [".py", ".js"]
}
```
**Output**:
```json
{
"results": [
{
"path": "src/auth.py",
"content": "def authenticate_user(user, password):\n ...",
"score": 0.89
}
],
"total_indexed_chunks": 1247,
"query_time_ms": 23
}
```
### `reindex_project`
Re-index a project after large changes.
**Input**:
```json
{
"project_path": "/path/to/your/project",
"project_name": "my-project",
"force": false
}
```
### `list_indexed_projects`
List all indexed projects.
### `get_files`
List indexed files in a project.
**Input**: `{ "project": "my-project" }`
### `get_file_content`
Return the full content of an indexed file.
**Input**: `{ "path": "src/main.py" }`
---
## โ๏ธ Configuration
### Environment Variables
```bash
# ChromaDB path (default: ./data/chroma_db relative to install dir)
export CHROMA_DB_PATH="/custom/path/to/chroma"
# Embedding model (default: all-MiniLM-L6-v2)
# Use microsoft/unixcoder-base for better code-specific embeddings (~2GB, requires torch)
export EMBEDDING_MODEL="microsoft/unixcoder-base"
# HTTP transport settings (only needed in HTTP/service mode)
export MCP_TRANSPORT="streamable-http"
export MCP_HOST="127.0.0.1"
export MCP_PORT="8080"
# Set this when exposing via reverse proxy or Tailscale Funnel
export MCP_ALLOWED_HOST="your-hostname.example.com"
# Log level (default: INFO)
export LOG_LEVEL="DEBUG"
```
### Chunking (Advanced)
Edit `src/codebase_rag/config.py`:
```python
CHUNK_SIZE = 500 # characters per chunk
CHUNK_OVERLAP = 50 # overlap between chunks
DEFAULT_TOP_K = 5 # default results per search
```
### Supported File Types
Python, JavaScript, TypeScript, JSX, TSX, Java, C, C++, Go, Rust, Ruby, PHP, C#, Shell, YAML, JSON.
### Ignored Patterns
`*.pyc`, `__pycache__`, `.git`, `node_modules`, `.venv`, `venv`, `*.egg-info`, `.pytest_cache`
---
## ๐ Benchmarks
| Operation | Expected Time | Notes |
|---|---|---|
| Index 20 .py files (~5k LOC) | ~5โ8s | First run; incremental is much faster |
| Vector search (top_k=5) | ~20โ50ms | ChromaDB in-process |
| Query embedding | ~10โ20ms | sentence-transformers, CPU |
| Server cold start | ~2โ3s | Model loaded into memory |
---
## ๐ค Automation Scripts
### Auto-discovery
Scan a directory for Git repositories and index them all automatically:
```bash
python scripts/auto_index.py ~/projects
```
### Watch Mode
Watch a project for file changes and reindex incrementally (debounced, 5s):
```bash
python scripts/watch.py /path/to/project --name my-project
```
### Git Hook (post-commit reindex)
Install a post-commit hook so changed files are reindexed automatically after every commit:
```bash
python scripts/setup_git_hook.py /path/to/your/repo my-project
```
---
## ๐งช Tests
```bash
# All tests (116 passing)
pytest -v
# Specific modules
pytest tests/test_config.py -v
pytest tests/test_embeddings.py -v
pytest tests/test_indexer.py -v
pytest tests/test_server.py -v
# With coverage
pytest --cov=codebase_rag --cov-report=html
```
---
## ๐ง Deploy as a systemd Service (Linux)
A template service file is provided at `systemd/codebase-rag-server.service`.
Replace `YOUR_USERNAME` with your actual Linux username before installing:
```bash
# Substitute your username in-place
sed -i "s/YOUR_USERNAME/$USER/g" systemd/codebase-rag-server.service
# Install and start
sudo cp systemd/codebase-rag-server.service /etc/systemd/system/
sudo systemctl daemon-reload
sudo systemctl enable codebase-rag-server
sudo systemctl start codebase-rag-server
# Check
sudo systemctl status codebase-rag-server
sudo journalctl -u codebase-rag-server -f
```
### Exposing Remotely via Tailscale Funnel (optional)
To use the server from a remote machine (Codespaces, company laptop, etc.):
```bash
# Expose port 8080 via Tailscale Funnel
tailscale funnel 8080
# Add to your service file:
# Environment="MCP_ALLOWED_HOST=your-machine.your-tailnet.ts.net"
# Then in your remote mcp.json:
# "url": "https://your-machine.your-tailnet.ts.net/mcp"
```
---
## ๐ Troubleshooting
**Slow first start**: The embedding model (~100MB) is downloaded on first use. Run `health_check.py` to pre-load it.
**High memory usage**: The default model uses ~500MB RAM. If needed, use an even smaller model via `EMBEDDING_MODEL`.
**Permission errors**: Ensure the running user has write access to `data/chroma_db/`.
**Debug mode**:
```bash
LOG_LEVEL=DEBUG codebase-rag
```
---
## ๐ Contributing
1. Fork the project
2. Create a feature branch: `git checkout -b feature/your-feature`
3. Follow strict TDD: RED โ GREEN โ REFACTOR
4. Atomic, descriptive commits
5. Open a pull request with tests
```bash
# Dev setup
pip install -e ".[dev]"
pytest -v --cov=codebase_rag
```
---
## ๐ License
MIT License โ see [LICENSE](LICENSE).
---
## ๐ References
- [MCP Protocol](https://modelcontextprotocol.io)
- [ChromaDB Docs](https://docs.trychroma.com)
- [sentence-transformers](https://www.sbert.net)
- [FastMCP](https://github.com/jlowin/fastmcp)
---
TDQS
Scored across 5 tools
Each tool has a clearly distinct purpose: search, reindex, list projects, list files, and get file content. There is no meaningful overlap; even the two 'get' tools are unambiguous because one returns metadata for files and the other returns the actual content.
All tool names follow a consistent verb_noun pattern: search_codebase, reindex_project, list_indexed_projects, get_files, get_file_content. The verbs are distinct and descriptive, and the pattern is uniformly applied without mixing styles.
With 5 tools, the server is well-scoped for a codebase RAG system. Each tool addresses a core need (indexing, search, and file retrieval) without unnecessary bloat, fitting comfortably in the ideal range and making the surface easy to navigate.
The tool set covers the primary lifecycle: index (reindex), list indexed projects, search, list files, and fetch file content. Minor gaps exist, such as no explicit delete project or incremental update mechanism, but these are not critical for core RAG workflows and can be worked around.