A Model Context Protocol (MCP) server that provides search capabilities for Cloudflare AutoRAG instances. This server enables AI assistants like Claude to directly search and query your AutoRAG knowledge base using three distinct search methods.
Features
🔍 - Vector similarity search without query rewriting or answer generation
A Model Context Protocol (MCP) server that provides search capabilities for Cloudflare AutoRAG instances.
星标
12
工具
—
可用率
—
Basic Search
✏️ Rewrite Search - Vector search with AI query rewriting but no answer generation (returns document chunks only)
🤖 AI Search - Full AI-powered search with optional AI response and configurable query rewriting
⚙️ Configurable Parameters - Support for score_threshold (default: 0.5) and max_num_results (1-50, default: 10)
📄 Pagination Support - AI search supports cursor-based pagination for large result sets (v1.2.0+)
🏢 Multi-AutoRAG Support - Manage and search across multiple AutoRAG instances (v2.0.0+)
🌐 Remote Deployment - Runs on Cloudflare Workers for scalability
🔗 MCP Compatible - Works with Claude Desktop and other MCP clients
Tools
autorag_basic_search
Performs a basic vector similarity search in your Cloudflare AutoRAG index without AI query rewriting or answer generation. Returns raw document chunks only.
Parameters:
query (string, required) - The search query text (max 10,000 characters)
max_num_results (number, optional) - Maximum number of results to return (1-50, default: 10)
autorag_name (string, optional) - Name of the AutoRAG instance to use (defaults to configured default)
autorag_rewrite_search
Performs a vector search with AI query rewriting but no answer generation. Uses Cloudflare's search() method with configurable rewrite_query for better semantic matching and returns only document chunks.
Parameters:
query (string, required) - The search query text (max 10,000 characters)
max_num_results (number, optional) - Maximum number of results to return (1-50, default: 10)
rewrite_query (boolean, optional) - Whether to rewrite query for better matching (default: true)
autorag_name (string, optional) - Name of the AutoRAG instance to use (defaults to configured default)
autorag_ai_search
Performs AI-powered search using Cloudflare's aiSearch() method with optional AI-generated response. Returns document chunks and optionally an AI answer based on the include_ai_response parameter. Supports pagination for large result sets.
Parameters:
query (string, required) - The search query text (max 10,000 characters)
Replace https://your-worker-url.workers.dev/ with your actual deployed Worker URL.
After updating the configuration:
Restart Claude Desktop
You should see the AutoRAG search tools available in your conversation
Configuration
Environment Variables
The server uses the following Cloudflare Worker bindings:
AI - Cloudflare AI binding for AutoRAG access (handles all AutoRAG operations)
AUTORAG_NAME - Your AutoRAG instance name (for single instance configuration)
AUTORAG_INSTANCES - Comma-separated list of AutoRAG instances (for multi-instance configuration)
AUTORAG_DESCRIPTIONS - Comma-separated list of descriptions for each instance
Wrangler Configuration
The wrangler.toml file includes:
name = "cf-autorag-mcp"
main = "src/server.ts"
compatibility_date = "2024-09-23"
compatibility_flags = ["nodejs_compat"]
[vars]
# For single AutoRAG instance:
AUTORAG_NAME = "your-autorag-instance-name"
# For multiple AutoRAG instances (v2.0.0+):
# AUTORAG_INSTANCES = "default-autorag,secondary-autorag,specialized-autorag"
# AUTORAG_DESCRIPTIONS = "Main knowledge base,Secondary knowledge base,Specialized documents"
[ai]
binding = "AI"
Note: The VECTORIZE binding is not required. AutoRAG manages its own vector index access internally through the AI binding.
Usage Examples
Once configured with Claude Desktop, you can use the tools like this:
Basic Search (no query rewriting, no AI response):
Search for documents about "machine learning" in my AutoRAG with a minimum score threshold of 0.7
Rewrite Search (AI query rewriting, no AI response):
Use rewrite search to find information about "deployment strategies" with query rewriting enabled
AI Search with Document Chunks Only (default behavior):
Use AI search to find information about "deployment strategies" with max 5 results
AI Search with AI-Generated Response:
Use AI search to find information about "deployment strategies" and include the AI-generated response
Multi-AutoRAG Usage (v2.0.0+):
List all available AutoRAG instances
Search for "security policies" in the secondary-autorag instance
Use AI search in specialized-autorag to find "compliance requirements" with AI response
Important Notes:
autorag_basic_search performs pure vector search without any AI enhancements
autorag_rewrite_search uses AI query rewriting but returns document chunks only
autorag_ai_search by default returns document chunks only (letting the client LLM generate responses), but can optionally include Cloudflare's AI-generated response
All tools use a default score threshold of 0.5 if not specified
All tools support the same parameter structure for consistent usage
Metadata filtering is not supported in Workers bindings - use the REST API if you need filtered queries
Development
Local Development
# Start local development server
npm run dev
# Build for production
npm run build
Project Structure
cf-autorag-mcp/
├── src/
│ └── server.ts # Main MCP server implementation
├── wrangler.toml # Cloudflare Workers configuration
├── package.json # Dependencies and scripts
└── README.md # This file
Technical Details
Protocol: JSON-RPC 2.0 over HTTP
Runtime: Cloudflare Workers with Node.js compatibility
MCP Version: 2024-11-05
Transport: HTTP-based (no streaming)
Default Score Threshold: 0.5 for all search tools
Parameter Validation: Comprehensive validation with clear error messages
Troubleshooting
Common Issues
"AutoRAG instance not found"
Verify your AUTORAG_NAME in wrangler.toml
Ensure your AutoRAG instance is properly created and indexed
"MCP server disconnected"
Check that your Worker URL is correct in the Claude Desktop config
Verify the Worker is deployed and accessible
"Tool not found" errors
Restart Claude Desktop after configuration changes
Check the Worker logs: npx wrangler tail
Empty search results
Try lowering the score_threshold parameter (default is 0.5)
Verify your AutoRAG index has been populated with documents
Check that your query terms exist in the indexed content