This project provides a Model Context Protocol (MCP) server for the Fuel Network and Sway Language ecosystem. It allows IDEs (like Cursor) to search and interact with Fuel documentation directly within the development environment.
The server indexes Fuel and Sway documentation into a local Vectra vector database using open-source embeddings (via Transformers.js) for powerful semantic search capabilities.
- Local semantic search of docs.fuel.network content
- No Docker dependency - runs with just Bun
- Fast file-based vector storage with Vectra
- Enhanced result filtering and formatting
- Hybrid search with keyword fallback
# Clone the repo
git clone https://github.com/FuelLabs/fuel-mcp-server
cd fuel-mcp-server
# Install dependencies
bun install
# Index documentation
bun run src/indexer.ts ./docs
# Test search
bun run src/query.ts --run "What is FuelVM?"
# Start MCP server
bun run src/cli.tsbun run src/cli.ts
# or explicitly
bun run src/cli.ts --transport stdiobun run src/cli.ts --transport http --port 3500
# Server runs at http://127.0.0.1:3500/mcp
# Health check: http://127.0.0.1:3500/healthbun run src/cli.ts --help
bun run src/cli.ts --transport http --port 3500
bun run src/cli.ts --transport stdioAdd to your MCP config file:
{
"mcpServers": {
"fuel-server": {
"command": "bun",
"args": ["run", "/absolute/path/to/fuel-mcp-server/src/cli.ts", "--transport", "stdio"]
}
}
}.
├── docs/ # Markdown documentation files
├── src/
│ ├── cli.ts # Main CLI entry point
│ ├── server.ts # MCP server factory
│ ├── transports/
│ │ ├── stdio.ts # STDIO transport
│ │ └── http.ts # HTTP transport
│ ├── chunker.ts # Markdown chunking logic
│ ├── indexer.ts # Document indexing script
│ ├── query.ts # Search query script
│ └── *.test.ts # Test files
├── vectra_index/ # Local vector database (created after indexing)
├── package.json
└── README.md
- Bun: Install from bun.sh
Place markdown files in ./docs or specify a different directory:
# Index docs in ./docs (default)
bun run src/indexer.ts
# Index custom directory
bun run src/indexer.ts /path/to/your/docs
# With custom settings
EMBEDDING_MODEL=Xenova/bge-small-en-v1.5 bun run src/indexer.ts ./docs# Basic search
bun run src/query.ts --run "What is the FuelVM?"
# Custom number of results
NUM_RESULTS=10 bun run src/query.ts --run "smart contracts"# Start MCP server (stdio-mode)
bun run src/cli.ts
# With HTTP transport
bun run src/cli.ts --transport http --port 3500bun test| Variable | Default | Description |
|---|---|---|
VECTRA_INDEX_PATH |
./vectra_index |
Vector database location |
EMBEDDING_MODEL |
Xenova/all-MiniLM-L6-v2 |
Hugging Face model |
CHUNK_SIZE |
2000 |
Target tokens per chunk |
NUM_RESULTS |
5 |
Search results count |
LOG_LEVEL |
Set to debug for verbose output |
- Chunking: Preserves code blocks, splits by paragraphs with context awareness
- Indexing: Generates embeddings with enhanced metadata for better search
- Querying: Semantic search with quality filtering and keyword fallback
- MCP Server: Exposes search as tool via stdio communication
- Storage: File-based Vectra index (no external database required)
searchFuelDocs
query(string): Search querynResults(number, optional): Number of results (default: 5)includeScore(boolean, optional): Include relevance scores
provideStdContext
- Returns Sway standard library paths and types
# Install dependencies
bun install
# Run tests
bun test
# Index sample docs
bun run src/indexer.ts ./docs
# Test search functionality
bun run src/query.ts --run "test query"
# Start MCP server for development (STDIO)
bun run src/cli.ts
# Start MCP server for development (HTTP)
bun run src/cli.ts --transport http --port 3500