Fuel Network & Sway Language MCP Server

This project provides a Multi-Component Protocol (MCP) server specifically designed for the Fuel Network and Sway Language ecosystem. It allows IDEs (like VS Code with the appropriate extension) to connect and seamlessly interact with Fuel documentation, enabling easier searching, understanding, and development within Fuel projects.

This server indexes Fuel and Sway documentation (including markdown files) into a Qdrant vector database using open-source embeddings (via Transformers.js). This allows for powerful semantic search capabilities directly within the development environment.

Features

• Makes the entire docs.fuel.network content locally searchable to agents
• Hybrid search (RAG + keyword via qdrant)
• qdrant db can be hosted for remote LLMs
• Contains the scripts to index new docs

Quick Install

# Git clone the repo
git clone --depth 1 https://github.com/FuelLabs/fuel-mcp-server

# Docker compose
docker compose -f fuel-mcp-server/docker-compose.yml up -d

# Copy this
realpath fuel-mcp-server

Edit your Cursor mcp.json

{
  "mcpServers": {
    "fuel-sever": {
      "command": "docker",
      "args": [
        "compose",
        "-f { replace w/ real path to fuel-mcp-server }/docker-compose.yml",
        "exec",
        "-T mcp-server bun run mcp-server"
      ]
    }
  }
}

Project Structure

.
├── docs/                     # Directory containing sample markdown files
│   └── fuel-docs.md          # Example doc
├── src/
│   ├── chunker.ts            # Logic for splitting markdown into chunks
│   ├── chunker.test.ts       # Tests for the chunker
│   ├── indexer.ts            # Main script to index docs into QdrantDB
│   ├── indexer.test.ts       # Tests for the indexer
│   ├── query.ts              # Script to query the QdrantDB collection
│   ├── query.vest.ts         # Tests for querying
│   └── mcp-server.ts         # MCP server implementation
├── node_modules/             # Project dependencies
├── qdrant_storage/           # Local Qdrant data persistence (if using Docker volume)
├── Xenova/                   # Cached embedding models
├── .env.example              # Example environment variables
├── .gitignore
├── bun.lockb                 # Bun lockfile
├── package.json
├── tsconfig.json
├── vitest.config.ts          # Vitest configuration
└── README.md

Development Prerequisites

• Bun: Install from https://bun.sh/
• QdrantDB: A running instance is required. The easiest way is using Docker:

  # Pull the Qdrant image
  docker pull qdrant/qdrant
  
  # Run Qdrant with persistent storage (creates ./qdrant_storage)
  docker run -p 6333:6333 -p 6334:6334 \\
      -v \"$(pwd)/qdrant_storage:/qdrant/storage:z\" \\
      qdrant/qdrant

  The scripts assume QdrantDB is accessible at http://localhost:6333. You can configure this using the QDRANT_URL environment variable. If your Qdrant instance requires an API key (e.g., Qdrant Cloud), set the QDRANT_API_KEY environment variable.

Running with Docker (Recommended)

This project includes a docker-compose.yml file to easily run both the Qdrant database and the MCP server in containers.

Prerequisites:

• Docker: Install from https://www.docker.com/
• Docker Compose: Usually included with Docker Desktop.

Steps:

1. Clone the repository (if you haven't already).
2. (Optional) Create a .env file: Copy .env.example to .env and configure environment variables if needed (e.g., QDRANT_API_KEY for Qdrant Cloud). Note: QDRANT_URL is automatically handled by Docker Compose for communication between the server and Qdrant containers. You can add other variables needed by the mcp-server here (like EMBEDDING_MODEL, QDRANT_COLLECTION).
3. Build and Start Containers: Open a terminal in the project root directory and run:

   docker compose up --build -d

   • --build: Builds the mcp-server image based on the Dockerfile.
   • -d: Runs the containers in detached mode (in the background). This command will:
   • Pull the qdrant/qdrant image if not present.
   • Build the mcp-server image.
   • Start containers for both Qdrant and the MCP server.
   • Set up a network for the containers to communicate.
   • Mount ./qdrant_storage for persistent Qdrant data.
4. Index Documents: To run the indexer script inside the running mcp-server container:

   # Index files in ./docs using default settings defined in the container
   docker compose exec mcp-server-app bun run src/indexer.ts
   
   # Index files specifying arguments (run inside the container)
   docker compose exec mcp-server-app bun run src/indexer.ts /app/docs my_collection Xenova/bge-small-en-v1.5

   • Remember that file paths (like /app/docs) are relative to the container's filesystem (/app is the WORKDIR defined in the Dockerfile). If you need to index files from your host machine, you might need to mount additional volumes in docker-compose.yml.
   • Environment variables from your .env file should be automatically picked up by the mcp-server container if defined under its environment section in docker-compose.yml.
5. The MCP Server is Running: The docker compose up command already started the MCP server as defined in the Dockerfile (CMD ["bun", "run", "mcp-server"]). It's accessible via docker compose exec for stdio communication.
6. Connect with Cursor:
   • Follow the previous instructions for connecting Cursor, but use the following stdio command:

   {
     "mcpServers": {
       "fuel-sever": {
         "command": "docker",
         "args": [
           "compose",
           "-f { replace w/ real path to fuel-mcp-server }/docker-compose.yml",
           "exec",
           "-T mcp-server bun run mcp-server"
         ]
       }
     }
   }

   • Replace { replace w/ real path to fuel-mcp-server } with the actual absolute path to your project directory where the docker-compose.yml file resides.
7. Stop Containers: To stop and remove the containers, network, and volumes defined in docker-compose.yml:

   docker compose down

   To stop without removing:

   docker compose stop

Using Taskfile (Alternative to Docker)

For a simplified setup, you can use Taskfile which provides easy commands for common operations.

Installation

macOS:

brew install go-task

Other platforms: See taskfile.dev/installation for installation instructions.

Usage

Once installed, you can use these simple commands:

# Complete setup (build, start, index)
task setup

# Start services without indexing  
task start

# Check service status
task status

# View logs
task logs

For all available commands, run:

task help

Installation

1. Clone the repository.
2. Install dependencies:

   bun install

3. (Optional) Create a .env file: Copy .env.example to .env and configure QDRANT_URL and QDRANT_API_KEY if needed.

Usage

1. Add Documents: Place your markdown files (.md) inside the docs/ directory (or specify a different directory when running the indexer).

2. Run Tests (Optional):

   bun test

3. Index Documents: Run the indexer script. This will read files from the specified directory (or ./docs by default), chunk them, generate embeddings using the configured model, and add them to the Qdrant collection.

   # Delete the qdrant_storage db
   rm -rf qdrant_storage
   
   # Run qdrant locally
   docker run -p 6333:6333 -p 6334:6334 -v "$(pwd)/qdrant_storage:/qdrant/storage" qdrant/qdrant
   
   # Index files in ./docs using default settings
   bun run index

   Script Arguments for Indexer:

   • docsDir (optional, positional): Path to the directory containing markdown files (default: ./docs).
   • collectionName (optional, positional): Name of the Qdrant collection to use (default: bun_qdrant_docs).
   • modelName (optional, positional): Sentence Transformer model from Hugging Face (default: Xenova/all-MiniLM-L6-v2).
   • targetChunkSize (optional, positional): Target token size for chunks (default: 2000).

   Environment Variables for Indexer:

   • QDRANT_URL: URL of your Qdrant instance (default: http://localhost:6333).
   • QDRANT_API_KEY: API key for Qdrant (if required).

4. Query Documents: Run the query script with your question as a command-line argument. You must include the --run flag before your query.

   bun run src/query.ts --run \"What is the FuelVM?\"

   Environment Variables for Query:

   • QDRANT_URL: URL of your Qdrant instance (default: http://localhost:6333).
   • QDRANT_API_KEY: API key for Qdrant (if required).
   • QDRANT_COLLECTION: Specify the collection to query (default: bun_qdrant_docs). Should match the one used for indexing.
   • EMBEDDING_MODEL: Specify the embedding model (default: Xenova/all-MiniLM-L6-v2). Should match the one used for indexing.
   • NUM_RESULTS: Number of results to retrieve (default: 5).

   Example with custom collection and number of results:

   QDRANT_COLLECTION=my_qdrant_collection NUM_RESULTS=3 bun run src/query.ts --run \"How do predicates work?\"

MCP Server (for IDE Integration)

This project includes an MCP (Model Context Protocol) server (src/mcp-server.ts) that exposes the Fuel documentation search functionality as a tool. This allows compatible clients, like Cursor, to connect and use the search capabilities directly within the IDE.

Running the MCP Server

Ensure QdrantDB is running and you have indexed your documents (see steps above).

To start the MCP server, run the following command. Configure environment variables as needed (especially QDRANT_URL, QDRANT_API_KEY, QDRANT_COLLECTION, EMBEDDING_MODEL if you used non-default values during indexing/querying).

# Example using default settings
bun run mcp-server

# Example with custom settings
QDRANT_URL=http://your-qdrant-host:6333 QDRANT_COLLECTION=my_docs bun run mcp-server

The server will connect via standard input/output (stdio) and wait for a client to connect.

Implementation Details

• Chunking (src/chunker.ts): Splits markdown by code blocks (\`\`\`) first. Text sections are then further split by paragraphs (\\n\\n) aiming for the target token size.
• Indexing (src/indexer.ts): Reads markdown, chunks content, generates embeddings using Transformers.js, and upserts points (vector + payload) into a specified Qdrant collection. Uses batching for efficiency.
• Querying (src/query.ts): Takes a text query, generates its embedding, and performs a similarity search against the Qdrant collection to retrieve the most relevant document chunks.
• MCP Server (src/mcp-server.ts): Implements the MCP protocol, listening on stdio. Exposes the queryDocs functionality as an MCP tool, handling request/response cycles with the client (e.g., Cursor).
• Embeddings: Uses Sentence Transformer models (e.g., Xenova/all-MiniLM-L6-v2) via the Transformers.js library to create vector representations of text chunks.