🇩🇪 Deutsche Version

🏛️ parlament-mcp

Part of the Swiss Public Data MCP Portfolio – connecting AI models to Swiss public data sources.

Note: This server covers the federal level (Curia Vista). This repo additionally hosts a self-contained subproject under openparldata-mcp/ — the subnational counterpart for the 26 cantons and ~70 municipal parliaments (OpenParlData.ch). The two are independent servers.

An MCP server that connects AI models to the Swiss Federal Parliament via the Curia Vista OData API (ws.parlament.ch). Access motions, interpellations, votes, members, sessions, and the verbatim debate transcripts of the Amtliches Bulletin – with no API key required (Phase 1 – No-Auth-First).

🎯 Anchor Demo Queries

Metadata layer:

"Welche Vorstösse zu KI in der Schule sind hängig?" → parlament_search_business(keyword="KI", keyword2="Schule", status="Eingereicht")

Verbatim transcripts (Amtliches Bulletin):

"What did National Councillor Munz say in the 2024 spring session about the Volksschule? Give me the exact wording with a correct AB citation." → parlament_search_transcripts(speaker_name="Munz", session_id=5202, keyword="Volksschule") → then parlament_get_transcript(transcript_id=…) for the full wording.

Returns short, citable excerpts (AB 2024 N, 2024-03-13, Munz Martina) with a stable source URL; the verbatim text is fetched on demand, never in bulk.

→ More use cases by audience →

Perfect for the KI-Fachgruppe Stadtverwaltung Zürich: find pending motions on AI in education, or quote what was actually said in the chamber – instantly.

🔧 Tools

🏗️ Architecture

┌──────────────────────────────────┐
│     MCP Host (Claude Desktop /    │
│     Claude API / IDE)             │
└─────────────┬─────────────────────┘
              │ MCP Protocol (JSON-RPC 2.0)
              │ Transport: stdio (local) / SSE (cloud)
┌─────────────▼─────────────────────┐
│          parlament-mcp            │
│   FastMCP · Python · Pydantic v2  │
│                                   │
│  ┌── metadata layer (server.py) ──┐
│  │  search_business · get_business │
│  │  search_members · get_votes     │
│  │  get_sessions                   │
│  └─────────────────────────────────┘
│  ┌── transcript layer ────────────┐  ← separate module (transcripts.py)
│  │  search_transcripts (excerpts) │    · Language='DE' dedups editions
│  │  get_transcript (verbatim)     │    · Type=1 = real speeches only
│  └─────────────────────────────────┘    · retry + 45s read timeout
└─────────────┬─────────────────────┘
              │ HTTPS / OData v3
┌─────────────▼─────────────────────┐
│  ws.parlament.ch / odata.svc      │
│  Curia Vista – No Auth Required   │
│                                   │
│  Business · Vote · MemberCouncil  │   metadata path
│  Session ─< Meeting ─< Subject ─< Transcript   transcript path
└───────────────────────────────────┘

🚀 Installation

Claude Desktop (stdio)

Add to ~/Library/Application Support/Claude/claude_desktop_config.json:

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

Local development

git clone https://github.com/malkreide/parlament-mcp
cd parlament-mcp
pip install -e .
python -m parlament_mcp.server

Cloud / Railway (SSE)

MCP_TRANSPORT=sse MCP_HOST=0.0.0.0 PORT=8080 python -m parlament_mcp.server
# SSE endpoint: http://your-host:8080/sse

Network binding

By default the server binds to 127.0.0.1 (localhost only). Set MCP_HOST=0.0.0.0 only inside a container/cloud context (Docker, Railway, Render, Kubernetes). Never bind to 0.0.0.0 on a local dev machine – it exposes the server to your local network (NeighborJack risk); the server logs a warning if you do so outside a detected container.

Transport is selected via MCP_TRANSPORT (stdio default, or sse / streamable-http); --http is kept as an alias for streamable-http.

Authentication (optional)

The HTTP transport is open by default (public read-only data). To require a bearer token, serve via the CORS/auth app factory and set MCP_BEARER_TOKENS:

MCP_BEARER_TOKENS="alice:tok_abc,bob:tok_def" MCP_ALLOWED_ORIGINS="https://claude.ai" \
  uvicorn parlament_mcp.server:create_http_app --factory --host 0.0.0.0 --port 8080

Each request then needs Authorization: Bearer <token>; identity comes from the validated token, not a session header (see docs/security.md).

Docker

docker compose up --build          # binds 127.0.0.1:8080 only
# or build the hardened image directly (non-root, read-only FS):
docker build -t parlament-mcp .

Kubernetes manifests (hardened securityContext, resource limits, egress NetworkPolicy, Mcp-Session-Id sticky routing) live in deploy/k8s/; an HAProxy stick-table example is in deploy/haproxy.cfg.

🔗 Synergies

Power query example:

"Zeige mir alle Zürcher Motionen zu KI in der Bildung
 und verlinke die relevanten Bundesgesetze aus fedlex-mcp."

📊 Data Source

• API: ws.parlament.ch/odata.svc
• Authentication: None (Phase 1 – No-Auth-First)
• Protocol: OData v3 / JSON
• Coverage: All parliamentary businesses since 1978; votes since ~2000. Structured verbatim transcripts (Amtliches Bulletin) from 1999-12-06 onward (earlier years 1891–1999 exist only as archive scans — see Known Limitations).
• Update cycle: Real-time (official government data)

Copyright — verbatim quotation is allowed

Official proceedings of Swiss authorities are excluded from copyright under Art. 5 para. 1 lit. a URG (Swiss Copyright Act). The verbatim wording of parliamentary debates in the Amtliches Bulletin may therefore be reproduced and quoted freely — which is exactly what the transcript tools return: the wording itself, never a summary standing in for it.

📜 Data sources & licenses

Every tool returns a typed structured response (FastMCP exposes the output schema) carrying source, license, provenance, match_type and count alongside typed results. Data is passed through unmodified.

🧭 Phase

This server is in Phase 1 — Read-only Wrapper (all tools readOnlyHint: true, no writes). The full phase model and transition criteria are in docs/roadmap.md.

🔖 MCP Protocol Version

Tested/targeted against MCP spec 2025-06-18 (pinned as PROTOCOL_VERSION in src/parlament_mcp/config.py). SDK updates are proposed monthly via Dependabot; spec-version bumps are recorded in CHANGELOG.md.

🧱 MCP primitives

Phase 1 uses Tools only — rationale and the Phase-2 Resources plan are in docs/adr/ADR-003-mcp-primitives.md.

🏷️ Tool annotations

All tools declare explicit annotations consistent with their behaviour:

📈 Observability

Structured JSON logs go to stderr (stdout stays reserved for the stdio protocol). OpenTelemetry tracing wraps each tool call and auto-instruments outgoing HTTP; set OTEL_EXPORTER_OTLP_ENDPOINT (with the otel-export extra) to ship spans. See docs/security.md for the full security posture (Lethal-Trifecta assessment, egress allow-list, gateway hardening).

🛡️ Safety & Limits

Known Limitations

General

• OData substringof() filter is case-sensitive for some fields.
• Session names may be null in the API for very recent sessions – use session ID.

Transcripts (Amtliches Bulletin) — verified live 2026-07-19:

• Temporal coverage: from 1999-12-06 only. The structured Transcript entity reaches back to Dec 1999. Debates from 1891–1999 exist only as scanned archive documents (Bundesarchiv / Amtsdruckschriften) and are not connected here (no OCR in scope). A query whose whole date window predates coverage returns an explanatory error, not an empty result.
• No page number in the source. The API carries no page/column field, so the classic AB <year> N <page> form cannot be built. We emit a stable, verifiable substitute — AB <year> <N|S>, <date>, <speaker> — plus the authoritative source_url (SubjectId) and the transcript_id. This is a documented, honest trade-off, not an omission.
• Language behaviour (important). Language is the edition, not the spoken language. The tools filter Language eq 'DE' purely to deduplicate the three byte-identical editions (DE/FR/IT) down to one copy. Every speech is returned in its original wording; a French or Italian speech is not hidden — its real language is reported in the language field (de/fr/it). The response states this via language_note.
• Truncation is explicit, never silent. Search returns short excerpts (snippet, ~320 chars) with is_excerpt, total_length_chars and a hint to fetch the full text. parlament_get_transcript caps output at max_chars; when it truncates it sets is_excerpt=True and returns a next_offset to continue.
• Verbatim only, never summarised. The wording is the product; the tools never substitute a summary for the actual text.
• Latency: a free-text keyword combined with a speaker_name is the slowest path (~40 s). Add a session_id, business_number or date window to keep reads around 1–2 s.

🧪 Testing

pip install -e ".[dev]"

# Unit + mocked integration tests (no network), as run in CI:
PYTHONPATH=src pytest tests/ -m "not live"

# Include live tests against the real ws.parlament.ch API:
PYTHONPATH=src pytest tests/ -m live

HTTP is mocked with respx; network-dependent tests are marked @pytest.mark.live and excluded from CI via -m "not live". Tool definitions are pinned in tool-hashes.json (python -m parlament_mcp.tool_hashes --check).

Contributing

See CONTRIBUTING.md.

Security

See SECURITY.md for the security policy and posture (vulnerability reporting, Lethal-Trifecta assessment, accepted risks).

License

MIT © Hayal Oezkan — see LICENSE

Author

Hayal Oezkan · github.com/malkreide

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": {
    "parlament-mcp": {
      "command": "uvx",
      "args": [
        "parlament-mcp"
      ]
    }
  }
}