Haystack MCP — Connect AI Pipelines to MCP Clients

What it is

Haystack MCP bridges the Haystack AI framework with the Model Context Protocol (MCP). It lets you expose any Haystack pipeline, whether for document search, question answering, or retrieval-augmented generation, as an MCP server. AI tools like Claude Code, Cursor, and other MCP-compatible clients can then query your pipelines directly without custom integration code.

Developers who have built Haystack RAG pipelines and want to make them accessible to AI coding agents benefit from this integration. It turns your document search infrastructure into a tool that any MCP client can call.

How it saves time or tokens

Without Haystack MCP, connecting a Haystack pipeline to an AI agent requires building custom API endpoints and tool definitions. Haystack MCP handles the protocol translation automatically, saving the development time of building and maintaining a custom integration layer. The token_estimate for this workflow is approximately 3,800 tokens.

How to use

1. Install haystack-ai and the MCP server package
2. Build a Haystack pipeline (retrieval, QA, or RAG)
3. Wrap it with HaystackMCPServer and run

Example

from haystack_mcp import HaystackMCPServer
from haystack import Pipeline
from haystack.components.retrievers import InMemoryBM25Retriever
from haystack.document_stores.in_memory import InMemoryDocumentStore

# Build a retrieval pipeline
store = InMemoryDocumentStore()
# ... add documents to store ...
pipeline = Pipeline()
pipeline.add_component('retriever', InMemoryBM25Retriever(store))

# Expose as MCP server
server = HaystackMCPServer(pipeline)
server.run()  # Now accessible via MCP protocol

Related on TokRepo

• RAG tools — Browse retrieval-augmented generation frameworks
• MCP integrations — Explore MCP server configurations

Common pitfalls

• The MCP server must be running for AI tools to access the pipeline; configure it as a persistent service
• Large document stores increase retrieval latency; optimize your Haystack pipeline before exposing it via MCP
• MCP protocol compatibility varies between clients; test with your specific AI tool before relying on it