rust-docs-mcp-server

Rust Docs MCP Server


⭐ Like this project? Please star the repository on GitHub to show your support and stay updated! ⭐

Motivation

Modern AI-powered coding assistants (like Cursor, Cline, Roo Code, etc.) excel at understanding code structure and syntax but often struggle with the specifics of rapidly evolving libraries and frameworks, especially in ecosystems like Rust where crates are updated frequently. Their training data cutoff means they may lack knowledge of the latest APIs, leading to incorrect or outdated code suggestions.

This MCP server addresses this challenge by providing a focused, up-to-date knowledge source for a specific Rust crate. By running an instance of this server for a crate (e.g., serde, tokio, reqwest), you give your LLM coding assistant a tool (query_rust_docs) it can use _before_ writing code related to that crate.

When instructed to use this tool, the LLM can ask specific questions about the crate's API or usage and receive answers derived directly from the _current_ documentation. This significantly improves the accuracy and relevance of the generated code, reducing the need for manual correction and speeding up development.

Multiple instances of this server can be run concurrently, allowing the LLM assistant to access documentation for several different crates during a coding session.

This server fetches the documentation for a specified Rust crate, generates embeddings for the content, and provides an MCP tool to answer questions about the crate based on the documentation context.

Features

• Targeted Documentation: Focuses on a single Rust crate per server

instance.

• Feature Support: Allows specifying required crate features for

documentation generation.

• Semantic Search: Uses OpenAI's text-embedding-3-small model to find the

most relevant documentation sections for a given question.

• LLM Summarization: Leverages OpenAI's gpt-4o-mini-2024-07-18 model to

generate concise answers based _only_ on the retrieved documentation context.

• Caching: Caches generated documentation content and embeddings in the

user's XDG data directory (~/.local/share/rustdocs-mcp-server/ or similar) based on crate, version, _and_ requested features to speed up subsequent launches.

• MCP Integration: Runs as a standard MCP server over stdio, exposing tools

and resources.

Prerequisites

• OpenAI API Key: Needed for generating embeddings and summarizing answers.

The server expects this key to be available in the OPENAI_API_KEY environment variable. (The server also requires network access to download crate dependencies and interact with the OpenAI API).

Installation

The recommended way to install is to download the pre-compiled binary for your operating system from the GitHub Releases page.

1. Go to the

Releases page.

1. Download the appropriate archive (.zip for Windows, .tar.gz for

Linux/macOS) for your system.

1. Extract the rustdocs_mcp_server (or rustdocs_mcp_server.exe) binary.
2. Place the binary in a directory included in your system's PATH environment

variable (e.g., /usr/local/bin, ~/bin).

Building from Source (Alternative)

If you prefer to build from source, you will need the Rust Toolchain installed.

1. Clone the repository:

   git clone https://github.com/Govcraft/rust-docs-mcp-server.git
   cd rust-docs-mcp-server

1. Build the server:

   cargo build --release

Usage

Important Note for New Crates:

When using the server with a crate for the first time (or with a new version/feature set), it needs to download the documentation and generate embeddings. This process can take some time, especially for crates with extensive documentation, and requires an active internet connection and OpenAI API key.

It is recommended to run the server once directly from your command line for any new crate configuration before adding it to your AI coding assistant (like Roo Code, Cursor, etc.). This allows the initial embedding generation and caching to complete. Once you see the server startup messages indicating it's ready (e.g., "MCP Server listening on stdio"), you can shut it down (Ctrl+C). Subsequent launches, including those initiated by your coding assistant, will use the cached data and start much faster.

Running the Server

The server is launched from the command line and requires the Package ID Specification for the target crate. This specification follows the format used by Cargo (e.g., crate_name, crate_name@version_req). For the full specification details, see man cargo-pkgid or the Cargo documentation.

Optionally, you can specify required crate features using the -F or --features flag, followed by a comma-separated list of features. This is necessary for crates that require specific features to be enabled for cargo doc to succeed (e.g., crates requiring a runtime feature like async-stripe).

# Set the API key (replace with your actual key)
export OPENAI_API_KEY="sk-..."

# Example: Run server for the latest 1.x version of serde
rustdocs_mcp_server "serde@^1.0"

# Example: Run server for a specific version of reqwest
rustdocs_mcp_server "reqwest@0.12.0"

# Example: Run server for the latest version of tokio
rustdocs_mcp_server tokio

# Example: Run server for async-stripe, enabling a required runtime feature
rustdocs_mcp_server "async-stripe@0.40" -F runtime-tokio-hyper-rustls

# Example: Run server for another crate with multiple features
rustdocs_mcp_server "some-crate@1.2" --features feat1,feat2

On the first run for a specific crate version _and feature set_, the server will:

1. Download the crate documentation using cargo doc (with specified features).
2. Parse the HTML documentation.
3. Generate embeddings for the documentation content using the OpenAI API (this

may take some time and incur costs, though typically only fractions of a US penny for most crates; even a large crate like async-stripe with over 5000 documentation pages cost only $0.18 USD for embedding generation during testing).

1. Cache the documentation content and embeddings so that the cost isn't

incurred again.

1. Start the MCP server.

Subsequent runs for the same crate version _and feature set_ will load the data from the cache, making startup much faster.

MCP Interaction

The server communicates using the Model Context Protocol over standard input/output (stdio). It exposes the following:

• Tool: query_rust_docs
• Description: Query documentation for the specific Rust crate the server

was started for, using semantic search and LLM summarization.

• Input Schema:

    {
      "type": "object",
      "properties": {
        "question": {
          "type": "string",
          "description": "The specific question about the crate's API or usage."
        }
      },
      "required": ["question"]
    }

• Output: A text response containing the answer generated by the LLM based

on the relevant documentation context, prefixed with From <crate_name> docs:.

• Example MCP Call:

    {
      "jsonrpc": "2.0",
      "method": "callTool",
      "params": {
        "tool_name": "query_rust_docs",
        "arguments": {
          "question": "How do I make a simple GET request with reqwest?"
        }
      },
      "id": 1
    }

• Resource: crate://<crate_name>
• Description: Provides the name of the Rust crate this server instance is

configured for.

• URI: crate://<crate_name> (e.g., crate://serde, crate://reqwest)
• Content: Plain text containing the crate name.

• Logging: The server sends informational logs (startup messages, query

processing steps) back to the MCP client via logging/message notifications.

Example Client Configuration (Roo Code)

You can configure MCP clients like Roo Code to run multiple instances of this server, each targeting a different crate. Here's an example snippet for Roo Code's mcp_settings.json file, configuring servers for reqwest and async-stripe (note the added features argument for async-stripe):

{
  "mcpServers": {
    "rust-docs-reqwest": {
      "command": "/path/to/your/rustdocs_mcp_server",
      "args": [
        "reqwest@0.12"
      ],
      "env": {
        "OPENAI_API_KEY": "YOUR_OPENAI_API_KEY_HERE"
      },
      "disabled": false,
      "alwaysAllow": []
    },
    "rust-docs-async-stripe": {
      "command": "rustdocs_mcp_server",
      "args": [
        "async-stripe@0.40",
        "-F",
        " runtime-tokio-hyper-rustls"
      ],
      "env": {
        "OPENAI_API_KEY": "YOUR_OPENAI_API_KEY_HERE"
      },
      "disabled": false,
      "alwaysAllow": []
    }
  }
}

Note:

• Replace /path/to/your/rustdocs_mcp_server with the actual path to the

compiled binary on your system if it isn't in your PATH.

• Replace YOUR_OPENAI_API_KEY_HERE with your actual OpenAI API key.
• The keys (rust-docs-reqwest, rust-docs-async-stripe) are arbitrary names

you choose to identify the server instances within Roo Code.

Example Client Configuration (Claude Desktop)

For Claude Desktop users, you can configure the server in the MCP settings. Here's an example configuring servers for serde and async-stripe:

{
  "mcpServers": {
    "rust-docs-serde": {
      "command": "/path/to/your/rustdocs_mcp_server",
      "args": [
        "serde@^1.0"
      ]
    },
    "rust-docs-async-stripe-rt": {
      "command": "rustdocs_mcp_server",
      "args": [
        "async-stripe@0.40",
        "-F",
        "runtime-tokio-hyper-rustls"
      ]
    }
  }
}

Note:

• Ensure rustdocs_mcp_server is in your system's PATH or provide the full path

(e.g., /path/to/your/rustdocs_mcp_server).

• The keys (rust-docs-serde, rust-docs-async-stripe-rt) are arbitrary names

you choose to identify the server instances.

• Remember to set the OPENAI_API_KEY environment variable where Claude Desktop

can access it (this might be system-wide or via how you launch Claude Desktop). Claude Desktop's MCP configuration might not directly support setting environment variables per-server like Roo Code.

• The example shows how to add the -F argument for crates like async-stripe

that require specific features.

Caching

• Location: Cached documentation and embeddings are stored in the XDG data

directory, typically under ~/.local/share/rustdocs-mcp-server/<crate_name>/<sanitized_version_req>/<features_hash>/embeddings.bin. The sanitized_version_req is derived from the version requirement, and features_hash is a hash representing the specific combination of features requested at startup. This ensures different feature sets are cached separately.

• Format: Data is cached using bincode serialization.
• Regeneration: If the cache file is missing, corrupted, or cannot be

decoded, the server will automatically regenerate the documentation and embeddings.

How it Works

1. Initialization: Parses the crate specification and optional features from

the command line using clap.

1. Cache Check: Looks for a pre-existing cache file for the specific crate,

version requirement, and feature set.

1. Documentation Generation (if cache miss):

• Creates a temporary Rust project depending only on the target crate,

enabling the specified features in its Cargo.toml.

• Runs cargo doc using the cargo library API to generate HTML

documentation in the temporary directory.

• Dynamically locates the correct output directory within target/doc by

searching for the subdirectory containing index.html.

1. Content Extraction (if cache miss):

• Walks the generated HTML files within the located documentation directory.
• Uses the scraper crate to parse each HTML file and extract text content

from the main content area (<section id="main-content">).

1. Embedding Generation (if cache miss):

• Uses the async-openai crate and tiktoken-rs to generate embeddings for

each extracted document chunk using the text-embedding-3-small model.

• Calculates the estimated cost based on the number of tokens processed.

1. Caching (if cache miss): Saves the extracted document content and their

corresponding embeddings to the cache file (path includes features hash) using bincode.

1. Server Startup: Initializes the RustDocsServer with the

loaded/generated documents and embeddings.

1. MCP Serving: Starts the MCP server using rmcp over stdio.
2. Query Handling (query_rust_docs tool):

• Generates an embedding for the user's question.
• Calculates the cosine similarity between the question embedding and all

cached document embeddings.

• Identifies the document chunk with the highest similarity.
• Sends the user's question and the content of the best-matching document

chunk to the gpt-4o-mini-2024-07-18 model via the OpenAI API.

• The LLM is prompted to answer the question based _only_ on the provided

context.

• Returns the LLM's response to the MCP client.

License

This project is licensed under the MIT License.

Copyright (c) 2025 Govcraft

Sponsor

Govcraft is a one-person shop—no corporate backing, no investors, just me building useful tools. If this project helps you, sponsoring keeps the work going.