Fuel Network & Sway Language

You run a local MCP server that lets IDEs search and interact with Fuel and Sway documentation. It indexes docs into a local vector store and exposes a fast, semantic search interface you can query from your editor or toolchain.

How to use

You can connect an MCP client to either STDIO (local) or HTTP (remote) transport to search Fuel and Sway documentation. Start the server in STDIO mode for a quick in-editor experience, or use HTTP when you want to query from other tools or processes over the network. The server also provides a health endpoint when using HTTP.

How to install

Prerequisites you need before installation are a modern runtime and a package manager. You will use Bun to install dependencies, index documentation, test searches, and start the MCP server.

# Install Bun (if you don’t have it yet)
curl -fsSL https://bun.sh/install | bash

# Clone the project repository
git clone https://github.com/FuelLabs/fuel-mcp-server
cd fuel-mcp-server

# Install dependencies
bun install

# Index documentation into the local vector store
bun run src/indexer.ts ./docs

# Test a search query locally
bun run src/query.ts --run "What is FuelVM?"

# Start the MCP server in STDIO mode
bun run src/cli.ts

# Start the MCP server with HTTP transport on port 3500
bun run src/cli.ts --transport http --port 3500

Configuration and server endpoints

The server exposes two primary MCP transport configurations you can use from your editor or tooling. The first is STDIO, which runs locally and communicates over the standard input/output channel. The second is HTTP, which exposes an MCP endpoint at a local URL.

{
  "mcpServers": {
    "fuel_mcp": {
      "type": "stdio",
      "command": "bun",
      "args": ["run", "src/cli.ts", "--transport", "stdio"]
    },
    "fuel_http": {
      "type": "http",
      "url": "http://127.0.0.1:3500/mcp",
      "args": []
    }
  }
}

Environment variables

Configure embedding, indexing, and query behavior using these environment variables. They control where the index is stored, which embedding model to use, chunk sizing, and how many results to return by default.

# Vector index location
VECTRA_INDEX_PATH=./vectra_index

# Embedding model used to generate vector representations
EMBEDDING_MODEL=Xenova/all-MiniLM-L6-v2

# Target tokens per content chunk for indexing
CHUNK_SIZE=2000

# Default number of search results to return
NUM_RESULTS=5

# Optional: enable verbose debugging output
LOG_LEVEL=debug

Tools and endpoints

The MCP server exposes a semantic search tool for Fuel docs. You can perform queries and retrieve relevant results along with optional relevance scores.

// Example tool endpoints described in the server features
// searchFuelDocs(query: string, nResults?: number, includeScore?: boolean): Result[]
// provideStdContext(): { sways: string[]; types: string[] }

Available tools

searchFuelDocs

Semantic search endpoint that queries Fuel docs using embeddings and returns relevant results, with optional scores.

provideStdContext

Tool that returns standard library paths and types for Sway