OCDS German Procurement Search

vergabe-mcp

Local MCP server for German public procurement search. Connects your AI assistant (Claude, GPT, etc.) to the Vergabe Dashboard API for semantic search, tender matching, and company-profile management.

Your queries and company profiles never leave your machine. They are embedded locally with a multilingual ONNX model; only the resulting embedding vectors, OCIDs, and filter values are sent to the API. Data minimisation by design.

Hosted alternative: if you just want search from your AI assistant, there is a hosted MCP connector (OAuth sign-in, no installation) — see the setup guide. This local server is for running the client in your own environment: queries and profiles are embedded locally, only vectors and filter values leave the machine, and it unlocks API-key features.

New to German public procurement? The Vergabe Dashboard knowledge base explains eForms, EU thresholds, and the tender lifecycle, and KI für Vergabe covers the hosted AI side of this server.

Quick Start

1. Get an API key

Sign up at vergabe-dashboard.qune.de and create an API key. API keys require an active Pro plan (€99/month; the local server itself is free and open source — the API gate rides key issuance).

2. Install

Via npx (easiest — downloads the correct binary automatically):

npx @qune-tech/vergabe-mcp --api-key sk_live_YOUR_KEY_HERE

Or download a pre-built binary from GitHub Releases:

Linux / macOS:

# Example for Linux x86_64 — adjust the filename for your platform

tar xzf vergabe-mcp-linux-x86_64.tar.gz sudo mv vergabe-mcp /usr/local/bin/vergabe-mcp

Windows: Extract the zip and move vergabe-mcp.exe somewhere on your PATH.

Or build from source:

git clone https://github.com/qune-tech/vergabe-mcp.git

cd vergabe-mcp cargo build --release

Binary at target/release/vergabe-mcp

3. Configure your AI client

Claude Desktop — edit claude_desktop_config.json:

Using npx:

{

"mcpServers": { "vergabe": { "command": "npx", "args": ["-y", "@qune-tech/vergabe-mcp", "--api-key", "sk_live_YOUR_KEY_HERE"] } } }

Using the binary directly:

{

"mcpServers": { "vergabe": { "command": "vergabe-mcp", "args": ["--api-key", "sk_live_YOUR_KEY_HERE"] } } }

Claude Code — add .mcp.json to your project root:

Using npx:

{

"mcpServers": { "vergabe": { "command": "npx", "args": ["-y", "@qune-tech/vergabe-mcp", "--api-key", "sk_live_YOUR_KEY_HERE"] } } }

Using the binary directly:

{

"mcpServers": { "vergabe": { "command": "vergabe-mcp", "args": ["--api-key", "sk_live_YOUR_KEY_HERE"] } } }

Cursor — Settings → MCP Servers → Add:

Using npx:

• Command: npx
• Args: -y @qune-tech/vergabe-mcp --api-key sk_live_YOUR_KEY_HERE

Using the binary directly:

• Command: vergabe-mcp
• Args: --api-key sk_live_YOUR_KEY_HERE

LM Studio — Settings → MCP → Add Server:

1. Click + Add Server and choose STDIO
2. Fill in:

Using npx:

• Name: vergabe
• Command: npx
• Arguments: -y @qune-tech/vergabe-mcp --api-key sk_live_YOUR_KEY_HERE

Using the binary directly:

• Name: vergabe
• Command: full path to the binary, e.g. /usr/local/bin/vergabe-mcp
• Arguments: --api-key sk_live_YOUR_KEY_HERE

1. Click Save
2. In the chat, select a model that supports tool use and enable the vergabe server

LM Studio requires models with tool-calling support (e.g. Qwen 2.5, Mistral, Llama 3.1+).

Smaller models may not use all 11 tools reliably — 7B+ recommended.

Replace sk_live_YOUR_KEY_HERE with your actual API key. You can also pass the key via the VERGABE_API_KEY environment variable.

Available Tools (11)

CLI Options

Usage: vergabe-mcp [OPTIONS]

Options: --db Local profiles database [default: profiles.db] --data-dir Data directory [default: data] --api-url Vergabe Dashboard API [default: https://vergabe-dashboard.qune.de] --api-key API key [env: VERGABE_API_KEY] -h, --help Print help

How It Works

LLM ←stdio→ vergabe-mcp (local)

│ Local: company profiles (SQLite) + sentence embedder (ONNX) │ HTTPS: vectors, OCIDs, filter values, API key └──HTTPS──→ Vergabe Dashboard API (/api/v1)

The MCP server runs locally on your machine:

• Company profiles (name, description, CPV interests, location) are stored in a local SQLite database — they never leave your network.
• Text embeddings are computed locally with a multilingual ONNX model (multilingual-e5-small, 384-dim). Search queries use the query: prefix and go to POST /api/v1/search/vector; profile descriptions use the passage: prefix and go to POST /api/v1/match/vector.
• Only embedding vectors (384 floats), the OCIDs you fetch, filter values, and your API key are sent to the API. Your query and profile text stay local.

Privacy & data flow

What stays on your machine: profile text, search-query text, CPV interests, location.

What the API sees: embedding vectors, the OCIDs you read, the filter values you use, and your API key. These reveal *commercial interest* (which sectors, buyers, value bands, tenders you look at) but not the underlying text. Embedding vectors are a derived, pseudonymous representation — minimisation, not elimination.

First-run model download: on first use the server downloads the embedding model from huggingface.co:

• What: model.onnx + tokenizer.json, ~118 MB total.
• When: the first time the embedder runs; cached afterwards at ~/.cache/vergabe/models/multilingual-e5-small.
• From: huggingface.co (a US-operated third party). No user data is sent in this fetch — it is a plain model download.

For air-gapped / enterprise installs, place model.onnx and tokenizer.json in the cache directory above and the runtime download is skipped.

Requirements

• An API key from vergabe-dashboard.qune.de on the Pro plan (€99/month)
• ~120 MB disk space for the ONNX model (downloaded automatically on first run)
• Internet connection to reach the API

License

MIT — see LICENSE.