🇨🇭 Part of the Swiss Public Data MCP Portfolio

swiss-food-safety-mcp

🌐 English | Deutsch

MCP server connecting AI models to Swiss Federal Food Safety and Veterinary Office (BLV) open data — food recalls, animal disease surveillance, food control results, antibiotic usage, children's nutrition surveys and the pesticide register. No authentication required.

Overview

swiss-food-safety-mcp gives AI assistants like Claude direct access to official Swiss food safety and veterinary data from the Federal Food Safety and Veterinary Office (BLV / Bundesamt für Lebensmittelsicherheit und Veterinärwesen). It provides 11 tools covering food recalls, animal disease surveillance, food control results, antibiotic usage in veterinary medicine, nutrition surveys for children, and the pesticide register.

All data comes from official Swiss federal sources (opendata.swiss, lindas.admin.ch, news.admin.ch). No API keys or authentication are required.

This server follows the No-Auth-First philosophy and is part of a Swiss public sector MCP portfolio.

Anchor demo query: "Are there any current BLV food warnings relevant to Zurich school canteens — and which notifiable animal diseases are currently reported in the canton?"

Demo

→ More use cases by audience →

Features

• 🚨 Public warnings & recalls — Live RSS feed of BLV product recalls and health warnings
• 🐄 Animal disease surveillance — Notifiable animal diseases since 1991 (InfoSM) via the LINDAS SPARQL cube
• 🐦 Avian influenza monitoring — Wild bird surveillance data with geodata
• 🥩 Food control results — Cantonal food inspection results and violation rates
• 💊 Antibiotic usage veterinary — ISABV data on antibiotic use in animal medicine
• 🧒 Children's nutrition survey — menuCH-Kids questionnaire tallies (answer counts, not nutrient intake)
• 🌿 Pesticide register — Swiss approved pesticide products and active ingredients
• 📊 Dataset discovery — Browse all 28 BLV datasets on opendata.swiss via CKAN API
• 🔗 Dual transport — stdio (Claude Desktop) + Streamable HTTP (cloud/Render.com)
• 🗣️ Bilingual — English-first documentation, German secondary

Prerequisites

• Python 3.11+
• uv or uvx (recommended) — install uv

Installation

Using uvx (recommended — no install needed)

uvx swiss-food-safety-mcp

Using uv

uv tool install swiss-food-safety-mcp
swiss-food-safety-mcp

From source

git clone https://github.com/malkreide/swiss-food-safety-mcp
cd swiss-food-safety-mcp
uv sync
uv run swiss-food-safety-mcp

Quickstart

Add to claude_desktop_config.json:

macOS: ~/Library/Application Support/Claude/claude_desktop_config.json
Windows: %APPDATA%\Claude\claude_desktop_config.json

{
  "mcpServers": {
    "swiss-food-safety": {
      "command": "uvx",
      "args": ["swiss-food-safety-mcp"]
    }
  }
}

Try it immediately in Claude Desktop:

"Which BLV food warnings are currently active?"
"Are there any notifiable animal diseases reported in Zurich canton this year?"

Other MCP Clients (Cursor, Windsurf, VS Code + Continue)

{
  "mcpServers": {
    "swiss-food-safety": {
      "command": "uvx",
      "args": ["swiss-food-safety-mcp"]
    }
  }
}

Cloud Deployment (Streamable HTTP)

For use via claude.ai in the browser (e.g. on managed workstations without local software):

# Loopback only (default) — safe for local testing:
swiss-food-safety-mcp --http
# Server runs on 127.0.0.1:8002

# External exposure (e.g. behind the Render TLS proxy):
swiss-food-safety-mcp --http --host 0.0.0.0

⚠️ The HTTP transport binds to 127.0.0.1 by default. Pass --host 0.0.0.0 only when external exposure is intended. Set BLV_MCP_ALLOWED_ORIGINS (comma-separated, no wildcard) to permit browser clients; it defaults to https://claude.ai.

Render.com (recommended):

1. Push/fork the repository to GitHub
2. On render.com: New Web Service → connect GitHub repo
3. Set the start command to: swiss-food-safety-mcp --http --host 0.0.0.0
4. In claude.ai under Settings → MCP Servers, add: https://your-app.onrender.com/mcp

Docker:

docker build -t swiss-food-safety-mcp .
docker run -p 8002:8002 swiss-food-safety-mcp
# or, with explicit CPU/memory limits:
docker compose up

The image is a non-root, multi-stage build; the container already binds 0.0.0.0 and includes a healthcheck. docker-compose.yml additionally caps CPU and memory.

💡 "stdio for the developer laptop, Streamable HTTP for the browser."

🔧 Configuration — every runtime setting is overridable via BLV_MCP_* environment variables (BLV_MCP_HTTP_HOST, BLV_MCP_HTTP_PORT, BLV_MCP_ALLOWED_ORIGINS, BLV_MCP_TIMEOUT, BLV_MCP_OTEL_ENDPOINT, …). Outbound requests are restricted to Swiss federal hosts (*.admin.ch, opendata.swiss). Optional OpenTelemetry tracing: install with pip install swiss-food-safety-mcp[otel] and set BLV_MCP_OTEL_ENDPOINT.

Available Tools

Example Queries

Architecture

┌─────────────────┐     ┌─────────────────────────────┐     ┌──────────────────────────────┐
│   Claude / AI   │────▶│   Swiss Food Safety MCP     │────▶│  Swiss Federal Open Data     │
│   (MCP Host)    │◀────│   (MCP Server)              │◀────│                              │
└─────────────────┘     │                             │     │  opendata.swiss (CKAN/CSV)   │
                        │  11 Tools · No Auth         │     │  lindas.admin.ch (SPARQL)    │
                        │  Stdio | Streamable HTTP    │     │  news.admin.ch (RSS/XML)     │
                        └─────────────────────────────┘     └──────────────────────────────┘

Synergies with Related MCP Servers

Project Structure

swiss-food-safety-mcp/
├── src/
│   └── swiss_food_safety_mcp/
│       ├── __init__.py        # Package metadata
│       └── server.py          # All tools, resources, prompts
├── tests/
│   ├── __init__.py
│   └── test_server.py         # Unit tests (no live API calls)
├── .github/
│   └── workflows/
│       └── ci.yml             # Python 3.11–3.13 matrix
├── pyproject.toml             # hatchling build, uv-compatible
├── CHANGELOG.md
├── CONTRIBUTING.md            # Contribution guide (English)
├── CONTRIBUTING.de.md        # Contribution guide (German)
├── SECURITY.md               # Security policy (English)
├── SECURITY.de.md            # Security policy (German)
├── LICENSE                    # MIT
├── README.md                  # This file (English)
└── README.de.md               # German version

Data Sources

All data is open government data (OGD) under Creative Commons with attribution requirement.

Known Limitations

• RSS feed: Limited to the most recent BLV publications; no historical archive
• Pesticide register: XML parsing may be slow for queries returning large result sets
• CKAN datasets: Opendata.swiss rate limits apply under heavy usage
• Animal disease data: Canton-level filtering depends on data completeness in the source
• Datasets are pinned, not searched: each data tool names its dataset slug and the resource that carries the data (see DATENQUELLEN in server.py). A keyword search takes the first hit and therefore falls back silently onto something plausible — that is how blv_get_animal_health_stats came to return antibiotics data, and how blv_get_food_control_results came to return a code list out of a dataset whose 26 resources include 18 of them. scripts/record_fixtures.py re-measures the pinned pairs on every run; a renamed dataset now fails loudly.
• Children's nutrition is questionnaire tallies, not nutrient intake: the only menuCH-Kids dataset published on opendata.swiss carries answer counts (Geschlecht, Sprachregion, Altersgruppe, Frage, Antwort, Anzahl). The docstring previously promised nutrient intake against dietary recommendations and offered "Energie", "Zucker", "Eisen" as filter examples — those matched nothing and returned an empty list. Adult food-consumption data exists as a separate dataset that this server does not cover.
• The SPARQL-to-CSV fallback is gone. It could never work: the one CSV resource of the fallback dataset is a ZIP file declared as format: CSV. With the endpoint corrected the fallback is also unnecessary — and a fallback that hides a broken query is worse than none.

Safety & Limits

• Read-only: All tools perform HTTP GET requests only — no data is written, modified, or deleted.
• No personal data: The APIs return aggregated public health and food safety statistics. No personally identifiable information is processed or stored by this server.
• Rate limits: opendata.swiss CKAN and lindas.admin.ch SPARQL are public APIs; use limit and filtering parameters conservatively. The server enforces a 30-second timeout per request.
• Data freshness: RSS warnings reflect the latest BLV publications at query time. Statistical datasets (animal diseases, food control, antibiotics) are updated periodically by the BLV. No caching is performed by this server.
• Terms of service: Data is subject to the ToS of each source — opendata.swiss, lindas.admin.ch, news.admin.ch. BLV data is published under Creative Commons with attribution.
• No guarantees: This server is a community project, not affiliated with the BLV or the Swiss federal administration. Availability depends on upstream APIs.

Deployment & Scaling

This server is Phase 1 — read-only (see ROADMAP.md): all 11 tools are read-only queries with no write surface.

Run it as a single instance. The Streamable HTTP transport keeps per-session state, so horizontal scaling would require Mcp-Session-Id sticky routing at the load balancer plus a shared session store — neither is implemented, by design, for a server of this scope. A single Render instance (or one container) is the supported deployment; docker-compose.yml sets explicit CPU/memory limits for self-hosting.

Testing

# Unit + contract tests (no network) — this is what CI runs
PYTHONPATH=src pytest tests/ -m "not live"

# All tests including live API checks
PYTHONPATH=src pytest tests/

# Re-measure which dataset and resource each tool hits
PYTHONPATH=src python scripts/record_fixtures.py

54 tests — 53 offline, 1 live.

Why the fixtures are recorded rather than written

A hand-written mock encodes its author's assumption and therefore cannot refute it: production code and fixture come from the same head, the same hour, the same reading of the docs. Where both are wrong, both are wrong together — and the suite stays green.

This repo had it in pure form. Every mocked CKAN resource was named "name": "CSV". On opendata.swiss the same field reads Food establishments 2025 or Food establishments codelist administrative measures — and that difference alone decided whether a tool returned inspection results or a code legend. The mocks could not express the distinction, so no test could fail on it.

What is recorded is therefore the selection: for each tool, the pinned dataset slug, the resource that was hit, and that file's header line. The header is the object of the exercise — it separates data from a legend, and it shows whether the BOM and the delimiter were handled. PROVENANCE.md names the source, the date, the selection rule and the SHA-256 for each file.

Two of the recorded measurements are controls: an invented path under lindas.admin.ch (POST 404, so the 404 on /sparql is real) and an invented class in the fsvo namespace (0 instances, so the previously queried foag class genuinely does not exist). Without them each measurement would only show what we received. The recorder aborts if a control stops discriminating, if a pinned resource disappears, if a header line is empty or starts with a BOM, or if one of the findings is superseded.

MCP Protocol Version

This server speaks spec 2026-07-28 natively. It runs fastmcp 4.x, which pins mcp 2.x, and that SDK serves two protocol eras over the same server object:

A current client gets 2026-07-28; one that has not moved yet falls back to the handshake and still gets every tool. Neither revision is chosen by this server — the SDK negotiates, and the pin records which pair that negotiation is allowed to produce.

The modern era is not just a higher number. It has no initialize handshake and no InitializeResult: Client.initialize() raises there, and the server metadata comes from server/discover instead. A check that still measured initialize_result.protocolVersion would therefore be testing only the legacy half while reporting a pass for both.

tests/test_protocol_version.py measures each era over its own path rather than comparing constants to one another: it pins both revisions against LATEST_MODERN_VERSION / LATEST_HANDSHAKE_VERSION, connects a real client in each mode, asserts the structural absence of InitializeResult in the modern one, and checks that both eras list the same tools. test_das_sdk_fuehrt_weiterhin_zwei_aeren guards the other direction — a downgrade back to mcp 1.x would otherwise reduce half of those assertions to an import error nobody reads.

The transport allow-list moved with the SDK: Mcp-Method, Mcp-Name and MCP-Protocol-Version are the routing headers spec 2026-07-28 mirrors into HTTP, and mcp.shared.inbound now reads them, so CORS lists them. Mcp-Param-* is deliberately still absent — the spec mints those only for parameters a tool marks with x-mcp-header, and no tool here does.

Changelog

See CHANGELOG.md

Contributing

See CONTRIBUTING.md

Security

See SECURITY.md for the security policy and how to report a vulnerability.

License

MIT License — see LICENSE

Author

Hayal Oezkan · github.com/malkreide

Credits & Related Projects

• Data: opendata.swiss / BLV – Federal Food Safety and Veterinary Office (BLV)
• Protocol: Model Context Protocol – Anthropic / Linux Foundation
• Related: zurich-opendata-mcp – MCP server for Zurich city open data
• Portfolio: Swiss Public Data MCP Portfolio

Installation

Run via uv's uvx — no clone or manual install needed. Add to your MCP client config (mcpServers for Claude Desktop, Cursor and Windsurf; use a top-level servers key for VS Code in .vscode/mcp.json):

{
  "mcpServers": {
    "swiss-food-safety-mcp": {
      "command": "uvx",
      "args": [
        "swiss-food-safety-mcp"
      ]
    }
  }
}