LODA

LODA MCP Server provides a token-efficient, section-aware document search capability for Claude Desktop and Claude Code using the Model Context Protocol. It helps you search large documents quickly while staying within token budgets and without relying on heavy external vector stores.

How to use

You integrate the LODA MCP Server with your Claude client so Claude can query your documents and retrieve the most relevant sections that fit within a token budget. Start searches by pointing Claude at your local or remote MCP server, and then issue queries like requesting the most pertinent sections related to a topic or term. The server will return complete, structured sections rather than arbitrary text chunks, preserving document coherence.

To use it effectively, configure Claude to communicate with the MCP server using the provided settings, then issue searches or prompts that specify a query and a maximum number of tokens you are willing to allocate for context. The MCP server uses Bloom filters for fast elimination, section-aware parsing to respect document structure, and a token budget to select the best-fitting sections.

How to install

Prerequisites you need installed on your system before setting up the MCP server are Node.js and a Git client.

Install and run the MCP server locally by cloning the project, installing dependencies, and starting the server.

# Clone the MCP server repository
git clone https://github.com/patrickkarle/loda-mcp-server.git

# Change into the project directory
cd loda-mcp-server

# Install dependencies
npm install

# Start the server in stdio mode (local development)
# This is the command you would use for a typical local run
node document_access_mcp_server.js --mode=stdio

Configuration and running modes

You configure Claude to connect to the MCP server by adding the MCP server entry to Claude’s configuration. The following examples show how to configure a local stdio MCP server named loda.

{
  "mcpServers": {
    "loda": {
      "command": "node",
      "args": ["document_access_mcp_server.js", "--mode=stdio"]
    }
  }
}

Additional HTTP mode for testing

For development and testing without Claude, you can run the MCP server in HTTP mode and access endpoints via HTTP requests.

node document_access_mcp_server.js --mode=http --port=49400

# Health check
curl http://localhost:49400/health

# List tools
curl http://localhost:49400/tools

# Example search
curl -X POST http://localhost:49400/tools/loda_search \
  -H "Content-Type: application/json" \
  -d '{"documentPath": "api-docs.md", "query": "authentication"}'

What you can search and how results look

The server focuses on returning the most relevant sections that fit within your token budget, preserving section structure and readability. You’ll receive structured results that include headers, section IDs, and token estimates, making it easy to decide which portions to load into your LLM prompt.

Notes on structure and performance

LODA uses Bloom filters to quickly prune non-matching sections, respects document structure, and scores candidates based on header matches, query term presence, and multi-term coverage. Cached document structures and Bloom filters improve response times for repeated searches.

Available tools

loda_search

Main search tool that returns the best document sections within a token budget.

list_document_sections

Get the hierarchical structure of a document.

read_section

Read a specific section by ID with context.

read_lines

Read specific line ranges within a document.

search_content

Basic regex search (non-LODA optimized).