MCP Doc Search

MCP Document Search Server

Overview

This project was built as a hands-on exercise to understand the Model Context Protocol (MCP) by implementing all three core MCP primitives:

• Prompts

• Tools

• Resources

The final outcome is a simple documentation search MCP server that exposes:

• A search_docs tool for searching documentation

• Dynamic resources representing documents stored in a local docs/ folder

The server can be tested and explored using the MCP Inspector.

Related MCP server: MCP Docs Server

MCP Concepts Learned

1. Prompts

Prompts are reusable instruction templates exposed by an MCP server.

Pattern:

• prompts/list

• prompts/get

Example:

simple

A client can discover available prompts and retrieve the prompt content.

2. Tools

Tools expose executable functionality.

Pattern:

• tools/list

• tools/call

Examples implemented during learning:

fetch
add
multiply
search_docs

A tool receives arguments, performs some action, and returns structured results.

3. Resources

Resources expose data that can be read by a client.

Pattern:

• resources/list

• resources/read

Examples:

```text
docs://mcp
docs://architecture
docs://retrieval

Resources represent existing data and are analogous to files or documents.

MCP Mental Model

Tool     = Function
Resource = File
Prompt   = Template

Examples:

search_docs("MCP")      -> Tool
docs://mcp             -> Resource
code-review-template   -> Prompt

Current Architecture

MCP Client
    │
    ├── search_docs(query)      [Tool]
    │
    └── docs://*                [Resources]
              │
              └── read_resource()

Search Flow

User Query
    ↓
search_docs("MCP")
    ↓
Returns matching resource URIs
    ↓
docs://mcp
    ↓
read_resource("docs://mcp")
    ↓
Returns document content

This demonstrates the common MCP pattern:

Tool → Resource Chain

Project Structure

doc-search/
│
├── README.md
├── pyproject.toml
│
└── mcp_doc_search/
    │
    ├── __init__.py
    ├── __main__.py
    ├── server.py
    │
    └── docs/
        ├── architecture.txt
        ├── retrieval.txt
        └── mcp.txt

Features

Tool: search_docs

Input:

{
  "query": "MCP"
}

Behavior:

• Searches all .txt files in the docs directory

• Performs a case-insensitive search

• Returns matching resource URIs

Example output:

docs://mcp

Resources

Resources are generated dynamically from the docs directory.

Examples:

docs://mcp
docs://architecture
docs://retrieval

Reading a resource returns the document content.

Testing

The server can be tested using MCP Inspector.

Example configuration:

Command:

uv

Arguments:

run python -m mcp_doc_search

The Inspector can then:

• List tools

• Call search_docs

• List resources

• Read resources

Key Takeaways

• MCP follows a consistent discovery and execution pattern.

• Tools are used for actions and discovery.

• Resources are used for retrieving known content.

• Prompts are reusable instruction templates.

• Many knowledge and retrieval servers follow a Tool → Resource architecture.

• The MCP layer remains stable even when the retrieval implementation evolves from simple file search to BM25, vector search, hybrid search, or databases.

Future Improvements

• BM25 search

• Vector search

• Hybrid search

• SingleStore integration

• Result ranking and scoring

• Resource metadata

• Claude Desktop / Cursor integration

• Retrieval-Augmented Generation (RAG)

create venv: uv venv Go to correct toml file and run: uv sync uv run python -m mcp_simple_prompt --help

uv run python -m mcp_doc_search --help

Install inspector: npx @modelcontextprotocol/inspector and runs it port:6274 (For stdio transport, Inspector itself launches the server, so need not run uv run mcp-simple-prompt in another terminal) or run python -m mcp_doc_search in the arguments

Enter the proxy token or launch the Proxy configured url and enter-> Command: uv and arguments: run mcp-simple-prompt