KnowledgeMCP

Give your docs an MCP endpoint. Every AI agent can use them.

https://github.com/user-attachments/assets/304fa31a-5138-4d02-b407-4c518f6a4145

KnowledgeMCP turns any documentation source (websites, PDFs, Confluence, Notion, S3, GitHub) into a standards-compliant Model Context Protocol (MCP) endpoint. Claude, GitHub Copilot, Cursor, and any other MCP-compatible agent can search and read those docs instantly — with no LLM calls at query time (we use a tiny local embedding model + hybrid BM25/kNN search in OpenSearch).

• 🔌 MCP-native — three tools (docs_search, code_sample_search, docs_fetch) any agent can plug into
• 💰 Zero-cost query path — local embeddings + OpenSearch hybrid search. No OpenAI/Bedrock fees per query.
• 🐳 docker compose up works — runs fully local, no AWS account, no credit card
• ☁️ Production-ready AWS path when you want it — Lambda + DynamoDB + SQS + S3 + managed OpenSearch via the bundled SAM template

Quick start

git clone https://github.com/hashwnath/KMCP.git
cd KMCP
make up                # docker compose up -d --build

Then:

• Dashboard → http://localhost:3000 (signup → add a source → search)
• Admin REST API → http://localhost:8081
• MCP endpoint → http://localhost:8000/mcp/{your-tenant-slug}

First-time start downloads the fastembed model (~30 MB) and OpenSearch (~700 MB image).

How agents use it

Point any MCP client at your tenant URL:

{
  "mcpServers": {
    "MyDocs": {
      "url": "http://localhost:8000/mcp/your-tenant-slug",
      "type": "http"
    }
  }
}

The agent gets three tools:

Architecture

┌────────────────────────────────────────────────────────────┐
│   AI Agents  (Claude, Cursor, Copilot, Continue, ...)      │
└──────────────────────────┬─────────────────────────────────┘
                           │ POST /mcp/{tenant_slug}
┌──────────────────────────▼─────────────────────────────────┐
│   MCP Server (FastMCP)  — docs_search / code_search / fetch │
└──────────────────────────┬─────────────────────────────────┘
            ┌──────────────┼──────────────┐
            ▼              ▼              ▼
    ┌───────────────┐ ┌──────────┐ ┌──────────────┐
    │   OpenSearch  │ │  SQLite  │ │  Filesystem  │
    │ (BM25 + kNN)  │ │ tenants  │ │  blobs       │
    │  ~768 token   │ │ sources  │ │  uploads     │
    │  chunks       │ │ jobs     │ │              │
    └───────────────┘ └──────────┘ └──────────────┘
                           ▲
┌──────────────────────────┴─────────────────────────────────┐
│ Admin API (Starlette)  +  Background Worker                 │
│   signup/login (JWT)        crawl → markdown → chunk →      │
│   sources CRUD              embed → OpenSearch              │
│   analytics                                                 │
└────────────────────────────────────────────────────────────┘

(In AWS mode, swap SQLite → DynamoDB, Filesystem → S3, the worker queue → SQS, and run each service as its own Lambda. The application code is unchanged because every AWS call routes through src/common/backends/.)

Supported source types

Configuration

Defaults work for local docker-compose. To customise, copy .env.example to .env and edit. The most useful knobs:

AWS production deployment

See docs/AWS_DEPLOYMENT.md for the SAM template (Lambda + DynamoDB + SQS + S3 + OpenSearch + SES), cost estimate, and operational runbook.

Contributing

PRs welcome. See CONTRIBUTING.md for the codebase tour and local dev setup.

make test        # full pytest suite (BACKEND=local)
make test-aws    # AWS-mocked suite
make up          # docker compose up -d --build

License

• Backend (src/, infra/, top-level configs) — AGPL-3.0
• Frontend (frontend/) — MIT

The AGPL-3.0 license means hosted/SaaS use must publish modifications under the same license. If that's a problem for your use case, please open an issue so we can discuss commercial licensing.

Why KnowledgeMCP?

Community

• GitHub Discussions — questions, ideas, show-and-tell
• Issues — bug reports, feature requests

Acknowledgements

• FastMCP — the MCP server framework
• fastembed — ONNX-runtime embedding library
• OpenSearch — hybrid BM25 + kNN search
• Microsoft Learn MCP server team — for documenting hard-earned lessons that shaped the tenant-context-via-middleware design