Part of the Swiss Public Data MCP Portfolio — a collection of open-source MCP servers connecting AI agents to Swiss public and open data.
This is a private project. It is not affiliated with, endorsed by, or operated on behalf of any employer or public authority.
lindas-mcp
MCP server for LINDAS — the Swiss administration's linked-data SPARQL knowledge graph (cube.link)
2026-11-09
ดาว
—
เครื่องมือ
—
อัปไทม์
—
MCP server for LINDAS — the linked-data knowledge graph of the Swiss administration.
LINDAS (Linked Data Service) is the Swiss Confederation's SPARQL knowledge
graph, run by the Federal Archives. Instead of tables, it publishes data as RDF
triples: around 2000 statistical data cubes (cube.link) from federal
offices, plus the geo-linked data that powers visualize.admin.ch.
Mnemonic: «I14Y is the library catalogue, LINDAS is the library itself.»i14y-mcp tells you a dataset exists.
LINDAS holds the data and lets you query across all of it at once.
This server wraps LINDAS in guarded tools rather than exposing raw SPARQL,
because the store rewards precise queries and times out on broad ones.
🎯 Anchor Demo Query
«Which forest-fire danger level currently applies, who publishes it, and
under which licence?»
The codes come back as labels — «grosse Gefahr», not 4. And the licence is a
Fedlex URI you can resolve with fedlex-mcp.
Demo
The two-phase access pattern
LINDAS cubes are self-describing but coded, and the server does two steps to
read them:
Structure — the cube's SHACL shape: its dimensions (filterable axes),
its measures (the numbers), and which dimensions carry code lists.
Data — the observations, with coded values resolved to human labels
using that structure.
Mnemonic: «LINDAS speaks in postcodes, not place names.» An observation
says region 1805; the server turns that into «Alpennordhang» for you.
Both steps happen inside query_cube_observations. It fetches the
structure itself, so a caller does not have to call get_cube_structure
first to get readable rows — doing so only repeats the cube-metadata and
dimension queries. get_cube_structure is the tool for finding out what a
cube contains (dimension names, key vs. measure, licence) and for writing a
run_sparql query against it. It reports only whether a dimension has a code
list, not the entries, so it is not a decoding aid either.
The lindas/ package is deliberately layered so it can be lifted into other
LINDAS-backed servers unchanged. client.py knows only HTTP and SPARQL;
cube.py knows the cube.link vocabulary; the tools know only cube.py. Raw
SPARQL never reaches the agent except through the guarded run_sparql escape
hatch.
Architecture decision
Architecture A (live SPARQL only), with a strict vocabulary guardrail.
Verified live on 2026-07-21:
The endpoint is stable, needs no authentication, and returns a clean HTTP 400
with a diagnostic on malformed queries.
Blind scans (SELECT *, COUNT(*) over the whole store) time out at
60–90 s; the same question anchored on ?x a cube:Cube answers in ~2 s.
Consequences, baked into the tools:
Every query template is anchored on a known class. No unbounded scans.
Two-phase access is enforced; the agent never sees raw codes.
run_sparql is capped at 500 rows and 30 s and marked as advanced.
The client timeout sits at 45 s, in front of the store's own 60–90 s abort.
Find cubes by topic. Entry point. Deduplicates versions.
get_cube_structure
Phase 1: dimensions, measures, licence.
query_cube_observations
Phase 2: data points with codes resolved to labels.
list_publishers
Federal bodies publishing cubes, with counts.
resolve_municipality
Name ↔ URI ↔ BFS number — the portfolio join key.
run_sparql
Advanced escape hatch. Capped, guarded.
api_status
Reachability check with cube count.
All tools are annotated readOnlyHint: true.
Reading a search_cubes result
returned is a count of what came back, not a statement about what exists. Two
fields say how complete the answer is:
Field
Meaning
truncated
false — every match is in cubes. true — there is more, or there may be more and the server could not rule it out.
total_matched
The exact total when one is available on the same unit as returned, null when no comparable number exists.
Check truncated before concluding anything. On true, do not report the
result as complete and do not answer "there are N cubes about X" from
returned. Widen in this order: raise limit (up to 100); if it is still
truncated at 100, narrow with creator_uri from list_publishers and ask one
federal body at a time; use latest_only=False only when you actually want the
version history, because it widens the result rather than escaping the cut.
total_matched is null more often than you might expect, and on purpose.
With latest_only=true (the default) the count LINDAS can answer cheaply counts
cube versions, while the tool returns version-collapsed cubes. Measured on
2026-09-20 with German labels: «wald» matches 127 published versions that
collapse to 35 logical cubes, «energie» 33 to 13. A total_matched of 127 next
to a returned of 20 would claim 107 missing cubes where at most 15 exist to
find — so the field says null, which is true, instead of a number that is not.
truncated stays reliable there and is the field to act on.
Where the server has seen every matching row — which is the common case, because
it fetches one row beyond what it needs — total_matched is exact in both
branches, and truncated is derived from it rather than from the fetch limit.
LINDAS_MCP_TRANSPORT accepts stdio (default), streamable-http or sse.
The transport decides the path: streamable-http serves /mcp, sse
serves /sse. Anything else falls through to stdio, which opens no port at
all — in a container that surfaces only as a failing health check.
Both HTTP transports bind to HOST, default 127.0.0.1;
set HOST=0.0.0.0 explicitly to expose it (only behind a reverse proxy).
LOG_LEVEL tunes the JSON stderr logs.
Hosting it as a remote connector
The connector URL is https://<host>/mcp. The path is not configurable —
it comes from the transport, and streamable-http is the one a current client
expects. sse and its /sse path are the specification's superseded transport;
nothing was removed and they still serve, but a new connector should not be
pointed at them.
A hosted deployment — Railway, Fly, a container behind any reverse proxy — needs
three variables, and the third is the one a deployment inherits wrongly because
nothing fails loudly without it:
Variable
Hosted value
If unset
LINDAS_MCP_TRANSPORT
streamable-http
Falls through to stdio: the process starts, opens no port, and surfaces only as a failing health check. The container image already sets it.
HOST
0.0.0.0
Binds loopback only, so the published port reaches nothing. The image sets it deliberately (SEC-016).
LINDAS_MCP_ALLOWED_HOSTS
the public hostname
Host validation is switched off entirely — see below.
LINDAS_MCP_ALLOWED_HOSTS is a comma-separated list of hostnames, without
scheme and without port: lindas-mcp.example.ch,alias.example.ch, not
https://lindas-mcp.example.ch:443. The value is matched literally against the
incoming Host header, and behind TLS on port 443 that header carries no port.
Loopback forms are added automatically, so the container health check keeps
working.
It fails in two opposite directions:
Unset on a non-loopback bind: the protection is off altogether. No
hostname is derivable in that situation — the server is reached under a
service or public DNS name this process does not know, and a guessed list
would answer every real request with 421. So no allow-list is installed at
all and the Host header is never checked, which is the SDK's own default.
The only sign is a startup warning, dns_rebinding_protection_off.
Set to the wrong name: every real request gets HTTP 421. The match is
exact and port-precise — mcp.example.ch does not cover
Host: mcp.example.ch:8443. If the proxy forwards a non-default port, name
both forms.
ALLOWED_ORIGINS is a separate question and concerns browser clients only.
It is the CORS origin list, comma-separated, and unset means no browser client
is permitted at all — that is the default. A client that is not a browser
sends no Origin and is unaffected. * is still accepted and logs a warning.
One detail worth knowing if you do serve browsers: the origins derived from
LINDAS_MCP_ALLOWED_HOSTS are the http:// ones, so an https:// browser
origin has to be named in ALLOWED_ORIGINS yourself.
Docker
docker compose up --build # binds 0.0.0.0 inside the container, publishes :8000
The image runs as a non-root user, read-only, with resource limits and a
TCP health check (see Dockerfile and compose.yaml).
Join keys
LINDAS is a connector layer, and two of its identifiers make it composable with
the rest of the portfolio:
The Fedlex link is the quiet surprise: many cubes declare their licence as a
legal-basis URI (fedlex.data.admin.ch/eli/cc/...), so you can go from a data
point straight to the law that governs it.
Known limitations
Verified live on 2026-07-21.
Broad SPARQL times out. The store aborts unanchored scans at 60–90 s.
The guarded tools avoid this; run_sparql warns about it and caps runtime.
Observations are coded. Dimension values are URIs, not labels. The server
resolves them via each dimension's code list, but resolution costs one extra
query per coded dimension. Set resolve_labels=False to skip it.
No server-side observation filtering by arbitrary value. LINDAS has no
cheap way to filter observations by a dimension value inside a cube, so
query_cube_observations reads the first N observations. Analytical slicing
belongs in run_sparql.
Licences vary per cube and are declared as dcterms:license, frequently
a Fedlex URI rather than a plain name. Always surface the licence field.
Version handling is heuristic.search_cubes deduplicates by stripping
the version suffix from the cube URI and keeping the highest schema:version
among published cubes. Unusual URI shapes may not collapse cleanly; use
latest_only=False to inspect every version. The same heuristic is why
total_matched can be null: it is a URI-shape guess in Python, and no cheap
SPARQL count expresses it, so restating it in a query would put the same guess
in a second place where it can drift.
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.
Era
Revision
Who reaches it
initialize handshake
2024-11-05 … 2025-11-25
What today's clients speak. The server answers with the revision asked for, or with the 2025-11-25 ceiling when the request asks for something newer.
Per-request envelope
2026-07-28
A request carrying the 2026-07-28_meta envelope opens a modern connection.
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. The handshake ceiling is measured against a live initialize through
the assembled ASGI stack, not read off a constant name.
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.
What 2026-07-28 changes here
The modern era has no initialize handshake, and therefore no handshake result
in which a client learns who it is talking to. Three consequences are served
explicitly rather than left at the SDK's defaults:
Surface
Behaviour
serverInfo
Stamped into _meta on every response and into server/discover. Name, title, version, description and website URL are resolved from the installed distribution's metadata, never written by hand. MCPServer defaults version to "" and substitutes nothing, so an unset version is a required field that says nothing.
instructions
Returned by server/discover — the only orientation channel a modern client has. It names the entry point, says that query_cube_observations resolves labels on its own, and says what get_cube_structure is actually for.
Log delivery
logging/setLevel is gone (SEP-2577); a client opts in per request via the reserved _meta key io.modelcontextprotocol/logLevel. Without it the server sends nothing; with debug it sends one notifications/message per tool call on that request's stream.
tools/list, server/discover
Carry ttlMs 300000 and cacheScopepublic (SEP-2549).
All of it is measured through the assembled stack in
tests/test_spec_2026_07_28.py — both eras,
both transports, and both branches of the log opt-in. The tool contract itself
is guarded independently by tool-definitions.lock.json (SEC-022), and the
modern era is asserted to serve exactly the tools listed there, so the two eras
cannot drift into two different servers behind one address.
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. SDK upgrades are a reviewed change for
the same reason: any protocol-affecting bump is called out in
CHANGELOG.md.
Testing
PYTHONPATH=src pytest tests/ -m "not live" # offline, used in CI
PYTHONPATH=src pytest tests/ -m "live" # hits the real endpoint
python -m ruff check src tests
The live tests earn their place: the observationSet indirection (a cube's
observations hang off cube:observationSet, never directly off the cube) is a
structural assumption that a mock cannot validate. It is covered by a live test.
Contributing
See CONTRIBUTING.md for the ground rules (read-only, one
egress host, anchored queries) and the local dev loop. Further reading:
EXAMPLES.md for use cases by audience with the tool-selection
table, docs/roadmap.md for the project phase, and
PUBLISHING.md for the PyPI / MCP Registry release process.
Security
See SECURITY.md for the security posture and how to report a
vulnerability.
License
MIT License — see LICENSE. The LINDAS data remains subject to the
licence each publisher declares on the cube.