API Docs

You run an MCP server that lets AI assistants search documentation semantically, retrieve precise chunks using recursive chunking, and answer with concise, relevant results. This server combines vector embeddings, a scalable store, and MCP endpoints to deliver fast, meaning-based access to your docs.

How to use

You connect an MCP client to the server to perform semantic searches across your documentation and retrieve targeted content. Use the HTTP endpoint to perform standard MCP requests for status, document listing, and query-based retrieval. You can also run a local, stdio-based client by launching the MCP command locally and piping requests through the standard input/output stream.

Typical usage patterns include: searching for specific topics to pull relevant chunks from multiple documents, requesting an overview of a particular document to understand its structure, and fetching a single section to present precise guidance. You can also request a paginated list of all chunks for a document to inspect how content is chunked and indexed.

How to install

Prerequisites: you need Node.js and npm, or a runtime suitable for your MCP client setup. You also need access to a server or hosting environment for the MCP endpoints and, if using vector storage, a Supabase project with the required SQL schema in place.

Step-by-step setup:

1. Create and configure a Supabase project with the provided SQL schema for metadata and chunks, including the pgvector extension and the similarity search function.

2. Set environment variables for your MCP server run: SUPABASE_URL, SUPABASE_KEY, and any required secrets such as APIFY_TOKEN for embedding and orchestration. Use placeholder values if you are preparing a template before production.

3. Prepare your documentation sources and indices by configuring your start URLs, max pages, and whether to force refresh. This enables the server to scrape content, generate embeddings, and populate the vector store.

4. Deploy or run the MCP server component configured for HTTP access and/or local stdio usage as described in the configuration section.

Configuration and examples

The server exposes an HTTP MCP endpoint for remote clients and an optional stdio workflow for local tooling. A typical HTTP endpoint URL looks like a MCP URL you will call from your client, for example: https://<YOUR_ACTOR_URL>/mcp. The local stdio workflow uses a command line invocation that runs the MCP process and communicates via standard input/output.

Environment variables you need to provide include your Supabase endpoint and key, plus any secrets you require for your embedding provider. You should store these securely in your hosting environment or your actor secrets management system.

Notes and troubleshooting

If you encounter issues with embedding generation or vector search, verify that the pgvector extension is installed in your database and that the embedding function is accessible. Check that your Supabase URL and anon/key are correctly configured in your environment. Ensure your start commands for stdio clients include the full runtime invocation with all necessary arguments.

For performance tuning, you can adjust the chunk size to balance token usage and retrieval precision. Target chunk sizes are between 500–800 tokens and a maximum of 1000 tokens per chunk. Smaller chunks improve retrieval precision, while larger chunks may reduce the number of queries needed.

Security and access control

Protect the MCP endpoints with appropriate access controls. Use API keys or tokens for remote access and restrict stdio usage to trusted environments. Store all sensitive credentials in secure secret management and rotate keys periodically.

Available tools

list_docs

List all docs with metadata (id, title, summary, sections) and basic information for quick browsing.

search

Perform semantic search across all docs to retrieve relevant chunks based on meaning.

get_doc_overview

Retrieve a document overview including summary and section structure.

get_section

Fetch content for a specific document section.

get_chunks

Provide paginated access to all content chunks within a document.

search_in_doc

Conduct semantic search within a single document to find precise sections.