Labsco
MCP SERVER

Embedding MCP Server

by Geeksfino ↗

Build a portable knowledge base with txtai and serve it over MCP — semantic search, a knowledge graph and text pipelines, all running locally with no external services.

Vector Stores & RAG Retrieval
Summary
A knowledge base you can carry.

The portability is the selling point: build once and the whole base — embeddings, graph and content — is a folder or a single `.tar.gz` you can move to another machine or hand to someone else. Because it is built on txtai, you are not locked into the bundled builder; any txtai base loads. Beyond plain semantic search it offers the graph side — path finding and community detection — and a causal boost that specifically helps "why" and "how" questions. It runs entirely locally, so nothing about your documents leaves the machine.

What it is

Two pieces: a command-line builder that turns your documents into a txtai knowledge base, and an MCP server that serves it. The knowledge base can be a folder or a single compressed `.tar.gz` archive, and any txtai-built base loads — you do not have to use the bundled builder.

What you get
  • Semantic search over the knowledge base — finding by meaning, not just keywords
  • Knowledge-graph querying: traversal, path finding between concepts, and community detection over automatically built graphs
  • Text-processing pipelines such as summarisation and extraction
  • A portable base built once and moved as a folder or a `.tar.gz` archive, shareable and reloadable
  • A causal-boost mechanism that lifts content explaining relationships, improving "why" and "how" answers
  • A builder CLI to process documents, create embeddings and export bases — `kb-build`, `kb-search`, `kb-mcp-server`
Requirements

Python 3.10 or newer, with `uv` recommended. Install with `uv pip install kb-mcp-server` (or pip/conda), or run without installing via `uvx --from kb-mcp-server@0.3.0 kb-mcp-server --embeddings /path/to/knowledge_base`. The server takes the knowledge base with `--embeddings` and can bind `--host` and `--port`, or read `TXTAI_EMBEDDINGS`; transport is stdio or SSE. Everything runs locally with no data sent to external services. The pyproject names the package `kb-mcp-server` at 0.3.0. Building a base uses a txtai YAML config, and the repository ships templates for use cases like technical docs, research papers and code repositories.

Setup effort

One command — pip install kb-mcp-server