🇨🇭 Part of the Swiss Public Data MCP Portfolio

wsl-envidat-mcp 🌲❄️⛰️

MCP server connecting AI models to Swiss environmental research data from WSL via EnviDat — forest, snow, avalanches, natural hazards and biodiversity, no API key required.

🇩🇪 Deutsche Version

Phase

This server is in Phase 1: Read-only Wrapper.

Phase-2 ideas (caching layer, semantic aggregation tool combining forest + snow + hazard data into a "Lage-Übersicht"): tracked under docs/.

Overview

The WSL (Eidgenössische Forschungsanstalt für Wald, Schnee und Landschaft / Swiss Federal Research Institute for Forest, Snow and Landscape) is one of Europe's leading environmental research institutes. Its open data platform EnviDat provides access to 1,000+ research datasets, time series of up to 130 years, and data from 6,000+ monitoring stations.

This MCP server exposes the EnviDat CKAN API as 10 tools and 2 resources, enabling AI assistants to search, filter and retrieve WSL research data by keyword, domain, or geographic bounding box — all without an API key.

Anchor demo query: "How was air quality and forest health around Schulhaus Leutschenbach in Zurich — and what does the WSL say about the current forest condition in the canton?"

Demo

Features

• 10 tools covering full-text search, domain-specific queries, spatial search, and curated thematic tools (avalanche, forest, natural hazards)
• 2 MCP resources for organizations and research domains
• 5 research domains: Forest · Biodiversity · Natural Hazards · Snow & Ice · Landscape
• 815+ datasets, time series since 1890, data from the SLF avalanche research institute
• No API key required — all data publicly accessible via open licenses
• Dual transport: stdio (Claude Desktop / local) + Streamable HTTP (cloud deployment)
• Model-agnostic: works with Claude, GPT-4, and any MCP-compatible client

Prerequisites

• Python 3.11+
• pip or uv / uvx
• Internet connection (live API calls to envidat.ch)

Installation

# Recommended: uvx (no installation needed)
uvx wsl-envidat-mcp

# Or with pip
pip install wsl-envidat-mcp

# Development
git clone https://github.com/malkreide/wsl-envidat-mcp.git
cd wsl-envidat-mcp
pip install -e ".[dev]"

Quickstart

Claude Desktop

Edit ~/Library/Application Support/Claude/claude_desktop_config.json (macOS) or %APPDATA%\Claude\claude_desktop_config.json (Windows):

{
  "mcpServers": {
    "wsl-envidat": {
      "command": "uvx",
      "args": ["wsl-envidat-mcp"]
    }
  }
}

Restart Claude Desktop, then ask:

• "What WSL datasets exist on fatal avalanche accidents in Switzerland?"
• "Show me forest inventory data from the LFI for the canton of Zurich."
• "Which natural hazard research data does the SLF publish on EnviDat?"
• "Are there WSL datasets on drought conditions in summer 2022?"
• "What biodiversity data is available for alpine ecosystems?"

Configuration

No API key required. Optional environment variables:

Cloud Deployment (Streamable HTTP)

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

# Local: keep MCP_HOST at its default 127.0.0.1
MCP_TRANSPORT=streamable-http PORT=8000 python -m wsl_envidat_mcp.server

# Container: bind to all interfaces inside the container only
MCP_TRANSPORT=streamable-http MCP_HOST=0.0.0.0 PORT=8000 python -m wsl_envidat_mcp.server

💡 "stdio for the developer laptop, streamable-http for the browser."

⚠️ Multi-Replica Cloud Deployments: Session state lives in the server. Run a single replica or enable sticky sessions (Railway/Render setting, or sessionAffinity: ClientIP on Kubernetes Services).

⚠️ Multi-Tenant / Unauthenticated Streamable HTTP: This server has no auth layer (auth_model: none). Streamable HTTP without a reverse-proxy + OAuth/API-Gateway is intended only for single-user deployments (e.g. one user's claude.ai browser session). For multi-tenant use, front the server with an authenticating gateway.

Container image (recommended for cloud)

A hardened multi-stage image is published to GitHub Container Registry on every main push and semver tag. Runs as non-root (uid=1000), no build tools in the runtime layer, multi-arch (linux/amd64 + linux/arm64).

The publishing workflow is .github/workflows/container.yml — not publish.yml, which only uploads the Python package to PyPI on a release. container.yml is delivery, not a gate: it does not gate a pull request, and a red run there means the image is missing, not that the code is broken.

docker run --rm -p 8000:8000 \
  --read-only --tmpfs /tmp \
  --cap-drop=ALL --security-opt=no-new-privileges \
  ghcr.io/malkreide/wsl-envidat-mcp:latest

Kubernetes hardening (excerpt):

securityContext:
  runAsNonRoot: true
  runAsUser: 1000
  readOnlyRootFilesystem: true
  allowPrivilegeEscalation: false
  capabilities: { drop: ["ALL"] }

Available Tools

Example Use Cases

Resources

Valid domain values: wald, biodiversitaet, naturgefahren, schnee_eis, landschaft

Architecture

┌─────────────────┐     ┌───────────────────────────┐     ┌──────────────────────────┐
│   Claude / AI   │────▶│    WSL EnviDat MCP        │────▶│       envidat.ch          │
│   (MCP Host)    │◀────│    (MCP Server)           │◀────│                          │
└─────────────────┘     │                           │     │  CKAN API  (REST/JSON)   │
                        │  10 Tools · 2 Resources   │     │  Solr full-text search   │
                        │  Stdio | Streamable HTTP  │     │  1,000+ research datasets│
                        │                           │     │  815+ open datasets      │
                        │  server.py                │     │  Time series since 1890  │
                        │  api_client.py            │     └──────────────────────────┘
                        └───────────────────────────┘

Infrastructure Components

Project Structure

wsl-envidat-mcp/
├── src/wsl_envidat_mcp/
│   ├── __init__.py         # Package
│   ├── server.py           # MCP server — 10 tools, 2 resources
│   └── api_client.py       # HTTP client for EnviDat CKAN API
├── tests/
│   └── test_integration.py # 11 live API integration tests
├── .github/workflows/
│   ├── ci.yml              # Gates: lint, format, syntax, import, tests, version sync
│   ├── live.yml            # Scheduled live tests against envidat.ch (cron 05:47 UTC)
│   ├── container.yml       # Builds & pushes the ghcr.io image (delivery, not a gate)
│   └── publish.yml         # Uploads the Python package to PyPI on a release
├── pyproject.toml          # Project config (hatchling build backend)
├── CHANGELOG.md
├── CONTRIBUTING.md         # Contribution guide (English)
├── CONTRIBUTING.de.md      # Contribution guide (German)
├── SECURITY.md             # Security policy & posture (English)
├── SECURITY.de.md          # Security policy & posture (German)
├── LICENSE                 # MIT
├── README.md               # This file (English)
└── README.de.md            # German version

Combination with Other MCP Servers

This server is part of the Swiss Open Data MCP Portfolio and integrates well with:

Known Limitations

• Solr search: OR is treated as a stopword — use single, specific search terms per query
• Domain search: Results depend on WSL's internal keyword tagging — not all datasets are tagged consistently
• Spatial search: Bounding box filtering is approximate; verify coordinates with individual dataset metadata
• Live API: All tools make live calls to envidat.ch — results depend on availability of the public API
• Languages: Dataset metadata is primarily in English and German; some older entries may be in German only

Safety & Limits

• Read-only: All tools perform HTTP GET requests only — no data is written, modified, or deleted on EnviDat.
• No personal data: The API returns research metadata, dataset descriptions, and download URLs. No personally identifiable information (PII) is processed or stored by this server.
• Rate limits: The EnviDat CKAN API is public without documented rate limits. Use limit and rows parameters conservatively. The server enforces a 30-second timeout per request.
• Data freshness: All tools make live API calls — results reflect the current state of the EnviDat catalog at query time. No caching is performed by this server.
• Terms of service: Data is subject to the EnviDat Terms of Use. Individual datasets are published under various open licenses (Creative Commons, CC0) — see dataset metadata.
• No guarantees: This is a community project, not affiliated with WSL or EnviDat. Availability depends on the upstream EnviDat API.

For the full security posture (egress allow-list, redirect handling, accepted risks) see SECURITY.md.

MCP Protocol Version

This server speaks two protocol eras over the same endpoint. The client's first request on a connection decides which one applies; a later claim from the other era is refused.

Both revisions are pinned in tests/test_protocol_version.py and asserted against the installed SDK, so a Dependabot bump of mcp cannot move either one silently. This server builds no ASGI app to send an initialize through, so the gate asserts the SDK constants rather than a measured response — the weaker form, named rather than left unsaid.

SUPPORTED_MCP_PROTOCOL_VERSION in server.py names the modern era; a mismatch against the SDK logs a warning at startup. A warning is not a gate — that is what the test file is for.

Note that the SDK's LATEST_PROTOCOL_VERSION is an alias for the modern era, not for the handshake era — pinning against it alone would leave the era that current clients actually negotiate free to drift.

Update policy. When the gate fails, do not edit the constant blindly: read the spec changelog between the two revisions, verify the server still behaves, then move the constant, this section, README.de.md and CHANGELOG.md together.

What the server uses from 2026-07-28

Pinning a revision is not the same as speaking it. Until now this server pinned 2026-07-28 and used the shape of the era before it; the table above was true and the wire was older than it sounded.

Two channels per tool result. The revision gives a tool result content for the reader and structuredContent for the application, with outputSchema as the contract between them. Every tool now fills both: the same human-readable Markdown as before in content, and the data behind it — typed, schema-backed — in structuredContent.

This is not a new interface. The fields are exactly the payload that response_format="json" has always produced. That option is the predecessor of structuredContent: it exists only because, before 2026-07-28, there was one channel, so the machine-readable version had to travel as a JSON string inside the reader's channel. response_format still works and still controls only the text block.

What it replaces is worse than nothing. A signature of -> str makes the SDK publish an outputSchema of {"result": string} and repeat the Markdown block underneath it. Measured on wsl_search against the recorded fixture: 2058 characters of text, 2058 characters of structuredContent.result, character-identical. An outputSchema is a promise; that one promised structure and delivered prose under a key called result.

serverInfo carries its six fields. It carried two, one of them empty. version was the expensive one: the number is kept in step across pyproject.toml, server.json and both README badges by scripts/check_version_sync.py — and reached every place except the one a client reads. It now comes from the package metadata, so no literal enters src/.

Tool titles moved to title. They sat in annotations.title, the slot from before this revision. A client that follows the schema and reads tool.title got None and displayed the slug.

Deliberately not used: icons on server, tools and resources, and the task-augmented call flow (execution). Both are optional, and neither has anything real to point at here — this repository ships no icon asset, and every call is a short CKAN query that finishes well inside a request. Declaring either would be a claim without a thing behind it.

tests/test_structured_output.py holds all of the above against a live ClientSession, with a negative control that shows what the SDK does when a server does not do this — so the day the default changes, the assertions say so instead of quietly passing.

Testing

# Unit tests — offline, no network access, all CKAN responses mocked via respx
PYTHONPATH=src pytest -m "not live"

# Live integration tests — actual HTTP calls to envidat.ch
PYTHONPATH=src pytest -m live

# Linting
ruff check src/
ruff format --check src/

CI runs the offline suite on every PR. The live suite runs only on main pushes and manual workflow_dispatch triggers, so build status is not coupled to upstream availability.

Changelog

See CHANGELOG.md

Contributing

See CONTRIBUTING.md

Security

See SECURITY.md (Deutsch) for the security posture and how to report a vulnerability.

License

MIT License — see LICENSE

Data on EnviDat is published under various open licenses (Creative Commons, CC0) — see individual dataset metadata.

Author

Hayal Oezkan · malkreide

Credits & Related Projects

• Data: EnviDat – WSL Swiss Federal Research Institute for Forest, Snow and Landscape
• Protocol: Model Context Protocol – Anthropic / Linux Foundation
• 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": {
    "wsl-envidat-mcp": {
      "command": "uvx",
      "args": [
        "wsl-envidat-mcp"
      ]
    }
  }
}