Idescat MCP Server

Idescat MCP server

MCP (Model Context Protocol) server that exposes tools to query the official statistics of Catalonia, published by Idescat (Institut d'Estadística de Catalunya), from any MCP client (Claude, Windsurf, Cursor, VS Code, etc.).

• Public endpoint: https://api.idescat.cat/mcp
• Transport: Streamable HTTP (MCP spec 2026-07-28, per-request metadata and server/discover, with legacy initialize compatibility for 2025-11-25 and earlier), stateless — each POST request is independent, with no sessions. POST responses use JSON or SSE according to the client's Accept header; no persistent SSE stream is offered via GET.
• Authentication: none required (the data is public).
• Data: official Idescat statistics in JSON-stat; labels and descriptions are in Catalan; territorial scope is Catalonia and its territorial breakdowns (counties, municipalities, provinces, etc.).

Installation

The public endpoint is https://api.idescat.cat/mcp. Add one of the snippets below to your MCP client's configuration. Different clients expect a different key (serverUrl, url, ...) and sometimes a type.

Windsurf/Devin, Antigravity (serverUrl)

{

"mcpServers": { "idescat": { "serverUrl": "https://api.idescat.cat/mcp" } } }

Cursor, Roo Code, Zoo Code (type: streamable-http)

{

"mcpServers": { "idescat": { "type": "streamable-http", "url": "https://api.idescat.cat/mcp" } } }

Cline (type: streamableHttp)

{

"mcpServers": { "idescat": { "type": "streamableHttp", "url": "https://api.idescat.cat/mcp" } } }

VS Code and the emerging standard form (servers, type: http)

{

"servers": { "idescat": { "type": "http", "url": "https://api.idescat.cat/mcp" } } }

Claude.ai (custom connector)

Claude.ai (web and desktop apps) can reach the remote server directly as a custom connector, no configuration file needed.

1. Open Customize → Connectors
2. Click Add custom connector.
3. Fill in the fields:

• Name: Idescat
• Remote MCP server URL: https://api.idescat.cat/mcp

1. Click Add. No authentication (OAuth / API key) is needed, since the data is public.
2. Back in a chat, enable the Idescat connector from the tools/attachments menu. Its tools and resources are now available to the assistant.
3. (Optional) In the connector's tool settings, switch each Idescat tool from Needs approval to Always allow so the assistant can run them without asking for confirmation every time. All the tools are read-only, so this is safe.

Claude Desktop and stdio-only clients (bridge with mcp-remote)

Clients that only speak stdio can reach the remote server through the mcp-remote bridge:

{

"mcpServers": { "idescat": { "command": "npx", "args": ["-y", "mcp-remote", "https://api.idescat.cat/mcp"] } } }

Not sure which key your client expects? (catch-all)

Most clients ignore keys they don't recognise, so this superset works with many of them at once:

{

"mcpServers": { "idescat": { "url": "https://api.idescat.cat/mcp", "serverUrl": "https://api.idescat.cat/mcp", "type": "streamable-http", "streamableHttp": "https://api.idescat.cat/mcp" } } }

Protocol versions

The server is dual-era (MCP versioning):

Modern requests must also send matching headers: MCP-Protocol-Version (equal to the _meta version), Mcp-Method (equal to the JSON-RPC method), and Mcp-Name for tools/call (the tool name) and resources/read (the URI; UTF-8 values may use the =?base64?...?= sentinel). Origin is always accepted — the data is public and the API answers CORS *.

Modern clients may open subscriptions/listen: the server answers on an SSE stream with notifications/subscriptions/acknowledged carrying an empty filter (no notification types are supported — the server is stateless and the listings don't change), followed by the gradual close defined by the spec. Clients should rely on ttlMs instead.

Available tools

Available resources

Beyond the tools, the server exposes MCP resources — content a client can attach to the context without executing any action.

Static resources (resources/list):

Resource templates (resources/templates/list) — parametrised URIs:

The transport is stateless, so subscribe and resources/list_changed

notifications are not offered (there is no persistent connection).

Server-side calculations

Language models are notoriously unreliable at mental arithmetic on long series of values. get_table_data and render_table therefore accept three optional parameters that make the server compute the result deterministically instead of leaving the model to add, subtract or divide the returned values itself:

{ "filters": { "AGE": ">=65", "SEX": "TOTAL" }, "last": 1 }

{ "sum": ["AGE"] } { "derive": { "share": "AGE", "percent_change": "previous" } }

• Age-range filters. For dimensions coded Y000, Y005_009, Y_GE100, etc., a filter value can be a range expression (">=65", ">65", "<16", "<=15", "16-64"); the server resolves it to the exact list of codes and fails with range_not_representable when the published groups cannot express the bound exactly.
• `sum` collapses the listed dimensions by adding up their categories; the result is a single SUM category. TOTAL is excluded automatically.
• `derive` appends derived metrics to the metric dimension: share ( _SHARE = 100 × value / value at TOTAL of the given dimension), difference ( _DIFF vs. the previous or a given period) and percent_change ( _PCT = 100 × (value − base) / base).

The engine relies on Idescat's conventions: additivity is declared by extension["%"] on the metric dimension (only additive metrics can be summed or shared), TOTAL is the total code, and non-computable cells come back as null with status d. Everything applied is recorded in extension.derived (ranges, summed codes, formulas, warnings) and render_table prints it under the table as a Calculations: block.

Unsafe requests are refused: metric_not_additive, geo_not_summable (the territorial dimension is never summed — use the geo parameter with the division that already publishes the aggregate), range_not_representable, incomplete_data and total_not_available.

Agent skills

Two skills in the Agent Skills open format (a directory with a SKILL.md file: YAML frontmatter + Markdown instructions) teach agents how to use Idescat data correctly. They are published on the website and advertised in the ARD manifest:

Repository files

The following files define the server's guidance and resources and are served, unchanged, to MCP clients:

• [`instructions.md`](instructions.md) — usage guidance the server sends to clients during the handshake and exposes as the idescat://instructions resource. It explains how to choose a statistic, the data-retrieval flow, how to interpret JSON-stat, and the source-attribution rules.
• [`resources.json`](resources.json) — declaration of the MCP resources and resource templates the server exposes (the two tables in the Available resources section above), including each URI, title, description, MIME type, and its backing file or tool.
• [`resources/geo.md`](resources/geo.md) — glossary of territorial codes (cat, prov, at, com, mun, dis, sec) used by the geo parameter, exposed as the idescat://glossary/geo resource.

The repository also includes:

• [`server.json`](server.json) — the server's metadata manifest in the standard server.json format, used to publish the server to the official MCP Registry.
• [`jsonstat/SKILL.md`](jsonstat/SKILL.md) and [`workflow/SKILL.md`](workflow/SKILL.md) — the two agent skills, also published on the website under https://www.idescat.cat/dev/api/taules/.

Discovery (Agentic Resource Discovery)

The server is advertised using the Agentic Resource Discovery (ARD) specification (v0.91) so agents can find it automatically:

• An ARD manifest is published at https://www.idescat.cat/.well-known/ard.json and https://api.idescat.cat/.well-known/ard.json (served with Access-Control-Allow-Origin: * on both hosts). It carries a top-level specVersion so the same document is valid both as an ARD manifest and as an AI Catalog file: /.well-known/ai-catalog.json is served direct