feedmcp: An MCP Server That Empowers LLMs with Offline Document Capabilities

feedmcp is an MCP (Model Context Protocol)-based server that enables large language models (LLMs) to efficiently utilize offline document resources via local SQLite and an advanced RAG pipeline, without relying on external vector databases. Its core advantages include easy deployment, privacy security, low cost, and full offline availability, making it an ideal solution for individual developers and small teams to empower LLMs with private knowledge bases.

The capabilities of large language models (LLMs) are often limited by the timeliness and coverage of their training data. When developers need AI assistants to understand private codebases, internal documents, or the latest technical materials, traditional solutions either rely on expensive fine-tuning or require building complex vector database infrastructure, which has a high barrier to entry. As an open standard launched by Anthropic, the Model Context Protocol (MCP) establishes a unified communication protocol between LLMs and external data sources, and feedmcp is a lightweight, fully localized solution based on this protocol.

The core design philosophy of feedmcp is no need for external vector databases—all operations are done in local SQLite. Unlike many RAG solutions that rely on cloud services like Pinecone or Weaviate, it embeds embedding, retrieval, and context management all into a single SQLite file, bringing significant advantages such as easy deployment, privacy security, low cost, and full offline availability.

feedmcp achieves efficient document retrieval through the following key mechanisms:

1. Context Chunk Header (CCH)：Dynamically add Markdown hierarchical structure and descriptors to each document chunk to avoid taking content out of context;
2. Relevant Segment Extraction (RSE)：Merge consecutive relevant matches in the source file into a coherent semantic document;
3. Payload Truncation Protection：Automatically limit the length of oversized text and prompt to use the read_doc tool for paginated reading;
4. Proxy Query Support：Guide LLMs to use HyDE and dynamic query reconstruction to optimize retrieval results.

feedmcp is developed in Go language, implements intelligent chunking based on LangChainGo, and supports semantic segmentation of long Markdown files with nested headings and sizes. It provides multiple transmission methods: stdio (suitable for Claude Desktop or CLI proxies), streamable HTTP, SSE; natively supports the .zstd compression format (compatible with the feedai project), allowing direct reading of large-scale compressed offline document archives.

Application Scenarios: For example, when a development team maintains hundreds of thousands of lines of internal code documents, they can ingest the documents into local SQLite via feedmcp, query them in an MCP-supported client, and get coherent relevant code snippets, file paths, function signatures, etc. Privacy Security: The pure local architecture ensures that sensitive documents never leave the local machine, and there is no need to send embedding vectors to third-party vector databases, making it suitable for scenarios such as handling confidential code, medical records, or legal documents.

feedmcp represents an important direction for the evolution of RAG technology towards lightweight and localization, proving that advanced retrieval capabilities do not require complex infrastructure—production-level document enhancement effects can be achieved through sophisticated algorithms and the flexibility of SQLite. For individual developers and small teams who want to empower LLMs with private knowledge bases, it is an ideal starting point (easy installation, zero cost, controllable privacy). As the MCP ecosystem matures, such tools will become standard components in LLM application development.