seco-labor-mcp
> π¨π **Part of the [Swiss Public Data MCP Portfolio](https://github.com/malkreide)**
# SECO Labor Market MCP Server

[](https://github.com/malkreide/seco-labor-mcp/actions/workflows/ci.yml)
[](https://pypi.org/project/seco-labor-mcp/)
[](https://www.python.org/)
[](https://modelcontextprotocol.io/)
[](https://github.com/malkreide/seco-labor-mcp)
[](LICENSE)
π **English** | **[Deutsch](README.de.md)**
An MCP (Model Context Protocol) server for Swiss labor market data from **SECO** (Staatssekretariat fΓΌr Wirtschaft) and **AMSTAT** via opendata.swiss.
<p align="center">
<img src="assets/demo.png" alt="Demo: Claude queries youth unemployment via seco-labor-mcp tool call" width="720">
</p>
---
## Overview
This server connects AI models to Swiss labor market statistics β unemployment rates, job seekers, open positions, youth unemployment, and occupational breakdowns β all without requiring an API key.
**Primary audiences:**
- π« **Schulamt / Education planning** β youth unemployment, vocational guidance data
- π **Research & analysis** β labor market trends, cantonal comparisons
- π€ **AI agents** β automated labor market monitoring and reporting
**Anchor query:**
*"Welche Berufsgruppen haben im Kanton ZΓΌrich die hΓΆchste Jugendarbeitslosigkeit, und welche Lehrberufe unterliegen der Stellenmeldepflicht?"*
[β More use cases by audience β](EXAMPLES.md)
---
## Data Sources (Phase 1 β No Auth Required)
| Source | Description | Status |
|--------|-------------|--------|
| [opendata.swiss](https://opendata.swiss/de/dataset) | CKAN catalogue; the pinned BFS table `T3.3.0.1` carries the SECO annual series | β
Live |
| [arbeit.swiss](https://www.arbeit.swiss) | Monthly press reports (PDF, structured URL pattern) | β
Live |
| [amstat.ch](https://www.amstat.ch) | AMSTAT reference portal | β οΈ JavaScript SPA, no public REST API |
| [unfallstatistik.ch](https://www.unfallstatistik.ch) | Unfallstatistik UVG (SSUV/KSUV c/o Suva) β occupational accidents and diseases | β οΈ PDF only, no API (see below) |
---
## Architecture
```
βββββββββββββββββββββββββββββββββββββββββββββββββββββββ
β seco-labor-mcp β
β β
β βββββββββββββββ ββββββββββββββββββββββββββββ β
β β FastMCP β β 9 MCP Tools β β
β β Server βββββΊβ seco_search_datasets β β
β β (stdio / β β seco_get_dataset β β
β β HTTP) β β seco_get_unemployment_* β β
β βββββββββββββββ β seco_get_youth_* β β
β β β seco_get_job_seekers β β
β βΌ β seco_get_open_positions β β
β βββββββββββββββ β seco_get_monthly_url β β
β β httpx β β seco_list_cantons β β
β β async β ββββββββββββββββββββββββββββ β
β ββββββββ¬βββββββ β
βββββββββββΌββββββββββββββββββββββββββββββββββββββββββββ
β
βΌ
βββββββββββββββββββββββββββββββββββββ
β opendata.swiss CKAN API β
β https://opendata.swiss/api/3/ β
β action/package_search β
β action/package_show β
βββββββββββββ¬ββββββββββββββββββββββββ
β
βΌ
βββββββββββββββββββββββββββββββββββββ
β SECO Data Resources β
β CSV / XLSX / PDF Downloads β
β (monthly labor market data) β
βββββββββββββββββββββββββββββββββββββ
```
---
## Where the figures come from β and what is missing
**SECO is no longer a publisher on opendata.swiss.** Verified 2026-08-14:
`organization_show` returns 404, and none of the 176 entries in
`organization_list` is SECO. Until then the server filtered every search on
that organisation and therefore returned **nothing** β a name lookup that
misses looks exactly like an empty search.
The registered unemployed and job seekers are still SECO's figures: the **BFS
publishes them** in table `T3.3.0.1` and names SECO in the footer. The server
reads that table through a **pinned dataset id** (`sources.py`), checked
against the live source by a live test.
| Series | 2000 | 2025 |
|---|---|---|
| Registered job seekers (SECO) | 124.6 | 214.1 |
| Registered unemployed (SECO) | 72.0 | 133.7 |
| ILO unemployed (BFS) | 126.5 | 248.5 |
*thousands, annual average*
The three series do **not** measure the same thing: in 2000 the ILO figure is
1.76Γ the registered one. The server reports them separately and labelled, and
never converts one into the other.
### The cantonal layer: four cantons, four schemas
There is no national monthly series β but **four cantons publish their own
RAV figures**, each in its own portal with its own column names. For those,
`seco_get_unemployment_overview(canton=β¦)` returns real values:
| Canton | Granularity | from | Level | Note |
|---|---|---|---|---|
| **TG** | monthly | 2016-01 | canton | only series **by age class** β youth unemployment as a count |
| **FR** | monthly | 2004-01 | canton **and Switzerland** | carries the national monthly figure as a comparison row |
| **ZG** | monthly | 1993-01 | canton | youth unemployment only as a **rate**, not a count |
| **ZH** | **annual** | 1991 | **municipality** | no monthly values; districts and regions sit in the same column as municipalities and are separated out |
**The other 22 cantons get a named refusal** β no figure from another canton
and no national aggregate. Partial coverage that feels complete is worse than
none.
The four series are **not comparable with each other** and do not add up to a
Swiss figure: different time axes, different geographic levels, and in ZG's
case a rate rather than a count.
**Still not available:** unemployment by occupational group, open positions as
a national series, and youth unemployment for Switzerland or for 24 of the 26
cantons. The affected tools say so and return **no** substitute figure. These
values exist interactively on [amstat.ch](https://www.amstat.ch/v2/amstat_de.html),
which offers no interface a server could call.
---
## Tools
| Tool | Description | Key Use Case |
|------|-------------|--------------|
| `seco_search_datasets` | Search labour-market datasets on opendata.swiss (publisher shown per hit) | Discovery |
| `seco_get_dataset` | Full metadata + download links for a dataset | Data access |
| `seco_get_unemployment_overview` | Registered unemployed: national annual, cantonal for TG/FR/ZG/ZH | Labor market overview |
| `seco_get_youth_unemployment` | Youth unemployment (15β24) β **TG** (count) and **ZG** (rate) only | π Berufswahlberatung |
| `seco_get_job_seekers` | Registered job seekers, national, annual series from 2000 | Training demand |
| `seco_get_open_positions` | Open positions β **no national series available** | Sector analysis |
| `seco_get_unemployment_by_occupation` | Breakdown by Berufshauptgruppe β **no machine-readable source** | π Vocational guidance |
| `seco_get_monthly_report_url` | Generate/verify PDF report URL | Source access |
| `seco_list_cantons` | All 26 canton codes and names | Utility |
| `seco_get_uvg_overview` | UVG key figures on occupational accidents and diseases | Risk overview |
| `seco_get_uvg_by_branch` | Results per NOGA 2008 economic branch | π Vocational guidance |
| `seco_get_uvg_trends` | Ten-year accident time series per branch | Trend analysis |
12 of a maximum of 15 tools.
---
## Unfallstatistik UVG (SSUV)
The three `seco_get_uvg_*` tools cover the risk side of the same labour market
the unemployment tools describe: how many occupational accidents and diseases
occur per branch, and how that develops over ten years.
**The publisher is not SECO.** The Unfallstatistik UVG is issued by the
Koordinationsgruppe KSUV and the Sammelstelle SSUV c/o Suva, Lucerne. The
`seco_` prefix addresses this server, not the source; every response names the
actual publisher in its `source` field.
### Architecture decision: C (dump-first)
Verified live on 2026-08-05, full write-up in
[`PROBE_REPORT_UVG.md`](PROBE_REPORT_UVG.md).
The source has **no API**. A link scan across every data page returned 165 PDFs
and zero files with `.csv`, `.xlsx` or `.json`. opendata.swiss does not list the
source at all (`count=0` for six of seven search terms), and the BFS dam-api
silently ignores its filter parameters. What remains is machine-readable in
practice but not by design:
| Access | Format | Refresh |
|---|---|---|
| `schluesselzahlen_d.htm` | HTML table, 5 years, Switzerland-wide | annually |
| `Ts{YY}.pdf` | annual edition, tables 1.2 and 2.4 by NOGA | annually, June |
| `WirtKl_{BUV\|NBUV}_{NN}.pdf` | ten-year series per NOGA division | annually, January |
PDFs are cached for 24 h and fetched with 2s/4s/8s backoff.
### What every response tells you
- `source_freshness.data_year` β the **data** year, not the edition year. The
2026 edition reports 2024; that two-year lag is stated, not buried.
- `totals_check` β parsed rows are summed and compared against the total
printed in the same publication. A broken layout shows up here instead of
becoming a plausible wrong number.
- `significant` β the source marks statistically significant year-on-year
changes with an asterisk. That flag is preserved per data point, so a change
is only reported as significant where the source says so.
---
## Installation
### Claude Desktop (stdio)
Add to `claude_desktop_config.json`:
```json
{
"mcpServers": {
"seco-labor": {
"command": "uvx",
"args": ["seco-labor-mcp"]
}
}
}
```
### Cloud / HTTP
```bash
pip install seco-labor-mcp
MCP_TRANSPORT=http PORT=8000 seco-labor-mcp
```
`http` is the transport to use. It is the only one that can carry the modern
protocol era: measured against this server object, `http` negotiates
**`2026-07-28`** while `sse` caps every client at **`2025-11-25`**, even a
client that offers the modern era. `sse` and `streamable-http` remain accepted
for existing deployments β `tests/test_transport_aera.py` runs both and records
which era each one actually yields.
The HTTP server binds to **`127.0.0.1` (loopback) by default** to prevent
NeighborJack on shared networks. For container deployments where you actually
need to accept traffic from outside the container, set `HOST=0.0.0.0`
explicitly β ideally in your Dockerfile / orchestrator config, and only behind
an upstream proxy or firewall:
```bash
HOST=0.0.0.0 MCP_TRANSPORT=http PORT=8000 seco-labor-mcp # container only
```
An unknown `MCP_TRANSPORT` value now exits with an error. It used to fall back
to stdio silently, so a typo produced a server that simply never appeared on
the expected port.
### Development
```bash
git clone https://github.com/malkreide/seco-labor-mcp.git
cd seco-labor-mcp
pip install -e ".[dev]"
pytest tests/ -m "not live" -v
```
---
## Usage Examples
### Search for youth unemployment data
```
Tool: seco_search_datasets
Input: { "query": "Jugendarbeitslosigkeit Alter", "limit": 5 }
```
### Get cantonal unemployment for ZΓΌrich
```
Tool: seco_get_unemployment_overview
Input: { "canton": "ZH", "response_format": "markdown" }
```
### Get monthly report URL
```
Tool: seco_get_monthly_report_url
Input: { "year": 2026, "month": 2, "language": "de" }
```
---
## Key Concepts
### Arbeitslose vs. Stellensuchende
> **EselsbrΓΌcke**: Arbeitslose β Stellensuchende β Arbeitslose sind eine Teilmenge.
| Term | Definition | Dec 2025 |
|------|-----------|----------|
| Arbeitslose | RAV-registered, immediately available | ~149'000 (3.2%) |
| Stellensuchende | All RAV-registered (incl. training programs) | ~233'900 |
### Youth Unemployment Seasonality
- **July/August**: Sharp increase (school leavers without placements)
- **September/October**: Decline (apprenticeship starts)
- The residual that remains after the autumn decline signals structural need for bridge programs (BrΓΌckenangebote)
### Stellenmeldepflicht (since 2020)
Occupations with β₯5% unemployment rate must be reported to the RAV before posting publicly. The list changes annually. This is directly relevant for vocational counseling β these professions have highest availability for Swiss job seekers.
---
## Portfolio Synergies
| Server | Synergy |
|--------|---------|
| `swiss-statistics-mcp` | BFS population/employment data for deeper context |
| `zurich-opendata-mcp` | City of Zurich-level education and social data |
| `swiss-snb-mcp` | Economic context (GDP, wages) for labor market interpretation |
| `fedlex-mcp` | ALV (Arbeitslosenversicherung) legislative framework |
---
## Known Limitations
- `amstat.arbeit.swiss` has no public REST API (JavaScript SPA) β workaround via CKAN
- Occupational/sectoral detail requires CSV download from SECO resources
- Monthly press report URL patterns may vary for older reports
- Cantonal sub-municipal data not available at this level
- UVG figures come from PDF parsing β the layout was stable across the 2025 and
2026 editions, but a redesign can break it. The `totals_check` in every
response is what makes such a break visible rather than silent.
- UVG data lags roughly two years (the 2026 edition reports 2024)
- UVG branch detail follows NOGA 2008 and groups some divisions (`41 β 42`,
`77, 79 β 82`); there is no cantonal breakdown at this level
- Detailed UVG data beyond the publications sits behind the SSUV closed user
group and is out of scope for this no-auth server
**Phase 2 roadmap:**
- Automatic CSV caching with 24h TTL
- Direct XLSX parsing for cantonal breakdowns
- Integration with `zh-education-mcp` for Schulamt-specific correlations
---
## Data License
Two different licences apply β the code of this server is MIT either way, but the
data is not covered by it.
**SECO / AMSTAT data** published on opendata.swiss is under **Creative Commons
CCZero** (public domain).
Source: Staatssekretariat fΓΌr Wirtschaft (SECO) β [seco.admin.ch](https://www.seco.admin.ch)
**Unfallstatistik UVG data** is **not** openly licensed. The publication states:
> Β«Abdruck β ausser fΓΌr kommerzielle Nutzung β mit Quellenangabe gestattet.Β»
> (Reproduction permitted, except for commercial use, with attribution.)
That is a non-commercial restriction with an attribution requirement. It belongs
to KSUV/SSUV and cannot be lifted by this repository's MIT licence: the MIT terms
cover the code, not the figures the code retrieves. **If you use this server
commercially, the UVG tools are not covered** β clarify directly with the
Sammelstelle (`unfallstatistik@suva.ch`). Every UVG response repeats this
restriction in its `source` field, because a README is not passed to the model.
---
## Safety & Limits
| Aspect | Details |
|--------|---------|
| **Access** | Read-only (`readOnlyHint: true`) β the server cannot modify or delete any data |
| **Personal data** | No personal data β all sources are aggregated, anonymous public statistics |
| **Rate limits** | No enforced external limits; server caps queries at 20 results by default; 30 s HTTP timeout |
| **Authentication** | No API keys required β opendata.swiss and arbeit.swiss are publicly accessible |
| **Licenses** | SECO data under [Creative Commons CCZero](https://creativecommons.org/publicdomain/zero/1.0/) (public domain) |
| **Terms of Service** | Subject to ToS of: [opendata.swiss](https://opendata.swiss/de/terms-of-use), [SECO](https://www.seco.admin.ch), [arbeit.swiss](https://www.arbeit.swiss) |
| **GDPR / DSG** | Fully compliant β no personal data transmitted or stored; all data is official public statistics |
---
## MCP Protocol Version
This server serves **two protocol eras** over the same server object, on
fastmcp 4.x / `mcp` 2.x:
| Era | Revision | Shape |
|-----|----------|-------|
| Modern | **`2026-07-28`** | no handshake β `server/discover`, one self-contained envelope per request |
| Handshake | **`2025-11-25`** | `initialize`, then a stateful session |
A client that offers the modern era gets it; a client that only knows the
handshake era still gets served. Both are pinned separately in
`tests/test_protokoll_aeren.py`, and both are *measured* β the test negotiates
a real connection against this server object rather than comparing two
constants.
Pinning only `LATEST_PROTOCOL_VERSION` would not be enough: in `mcp` 2.x that
name is an alias for the *modern* era, so it would leave the handshake ceiling
free to move β and that ceiling is what most clients in the field actually
speak.
Until 0.4.0 this server ran fastmcp 3.x, which pins `mcp` 1.x. There
`2025-11-25` is the highest revision the SDK knows at all, so `2026-07-28` was
not partially supported β it was absent. The test that used to guard the
one-era state now guards its opposite: it fails if a downgrade takes the
modern era away again.
Note for anything reading server metadata: on a modern connection there is no
`InitializeResult`. Use the era-neutral `protocol_version` / `server_info`
instead of `initialize_result`.
---
## Contributing
See [CONTRIBUTING.md](CONTRIBUTING.md) for development guidelines.
---
## Security
See [SECURITY.md](SECURITY.md) for the security posture and how to report a
vulnerability.
---
## License
Released under the [MIT License](LICENSE) β Copyright Β© 2026 Hayal Oezkan.
---
## Author
**Hayal Oezkan** Β· [github.com/malkreide](https://github.com/malkreide)
<!-- mcp-name: io.github.malkreide/seco-labor-mcp -->
<!-- BEGIN GENERATED: install -->
## Installation
Run via [`uv`](https://docs.astral.sh/uv/)'s `uvx` β 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`):
```json
{
"mcpServers": {
"seco-labor-mcp": {
"command": "uvx",
"args": [
"seco-labor-mcp"
]
}
}
}
```
<!-- END GENERATED: install -->
TDQS
Scored across 12 tools
Each tool targets a distinguishable data source or operation, and the descriptions carefully contrast overlapping concepts like registered unemployed vs. job seekers. However, several unemployment-related getters (overview, youth, by occupation, job seekers) cover closely related ground, so an agent must read descriptions carefully to avoid selecting the wrong one.
All tools follow a consistent seco_<verb>_<noun> snake_case pattern, with get, search, and list used predictably. This makes the tool surface easy to scan and reason about.
Twelve tools is well within the ideal range for a domain-specific data server. Each tool appears justified, from core unemployment statistics to metadata lookup and canton utilities, without unnecessary bloat.
The server covers the main labor market data needs: unemployment, job seekers, open positions, occupational breakdowns, youth unemployment, metadata search, and even occupational accident data. Minor gaps like monthly or cantonal breakdowns are explicitly documented as unavailable rather than left as dead ends.