Feedmcp

feedmcp is a Model Context Protocol (MCP) server designed to supercharge Large Language Models (LLMs) with your offline documentation using an Advanced RAG Pipeline natively entirely via local SQLite operations.

Advanced Local RAG Architecture

Feedmcp natively incorporates advanced context strategies without the overhead of external Vector DBs, optimized specifically to keep LLM context windows clean and relevant:

• Contextual Chunk Headers (CCH): Data is grounded intelligently; before chunks are embedded they are dynamically prepended with markdown hierarchies and document descriptors.
• Relevant Segment Extraction (RSE): Reconstructs contiguous data chunks logically. If multiple relevant matches occur sequentially in the same file, the server merges them into a single, seamless semantic document.
• Payload Truncation Safeguards: Automatically caps oversized text segments to ensure the LLM's context window isn't hijacked by a single massive file, prompting the agent to paginate using the read_doc tool instead.
• Agentic Query Support: Implements built-in tool prompts instructing connected LLMs directly to employ techniques like Hypothetical Document Embeddings (HyDE) and dynamic Query Reformulation workflows automatically.
• Natural Language Reliability: Best run natively with Grounded Language Models to ensure hallucinations are inherently reduced via verbatim context-streaming directly from .zstd offline files.

Features

• Direct ZSTD Support: Reads massively compressed .md.zstd archives, format used by feedai.
• Smart Semantic Chunking: Uses LangChainGo to intelligently split long markdown files by nested headers and size.
• Multi-Transport: Runs over stdio, sse, or streamable.
• CORS Enabled: Out-of-the-box support for web-based MCP inspectors.

Installation & Build

# Clone your repository
git clone https://github.com/akhenakh/feedmcp.git
cd feedmcp

# Build with the FTS5 tags enabled (REQUIRED)
go build -tags sqlite_fts5 -o feedmcp .

Quick Start Workflow

Ingest Documentation

feedmcp ingest ~/dev/feedai/docs/

Start the Server

For Claude Desktop or CLI Agents (Stdio):

feedmcp serve --db docs.sqlite --transport stdio

For Web Clients, Cursor, or Remote Agents (Streamable HTTP / SSE):

# Modern Streamable HTTP (available at http://localhost:8080/mcp)
feedmcp serve --db docs.sqlite --transport streamable --port 8080

Client Integration Examples

Opencode

Edit ~/.config/opencode/opencode.json:

  "mcp": {
    "feedmcp": {
      "type": "local",
      "command": ["feedmcp", "serve", "--db", "/home/youruser/dev/feedmcp/docs.sqlite", "--transport", "stdio"]
    }
  }

License

MIT