Datos.gob.do MCP Server

[English](README.md) · [Español](README.es.md)

---

datosgobdo-mcp

Ask an AI assistant a question about Dominican public data, and get an answer traced back to the government file it came from.

This is a Model Context Protocol server for datos.gob.do, the Dominican Republic's official open-data portal. It plugs into Claude Desktop, Claude Code, Cursor, ChatGPT Desktop or any MCP-compatible client, and lets the model search the catalog, download the actual files, parse them, and run real analysis — without you writing code, opening a URL, or downloading a spreadsheet.

Official source. The canonical repository is

alcastaro/datos.gob.do-MCP-server.

The only official distributions are the PyPI package

dominican-open-data-mcp

and the MCP Registry entry io.github.alcastaro/datos.gob.do-MCP-server.

Copies published elsewhere are not maintained by the author and may be

outdated or modified — verify against this repository before installing.

This README is written to be read in order. Part 1 needs no technical knowledge. Part 2 teaches what MCP actually is, using this server as the worked example. Parts 3 to 6 are the reference and the engineering detail. If you want the same material as a walkthrough, read the Tutorial (Español).

---

Contents

Part 1 — Start here (no technical knowledge needed)

1. What this is, in plain words
2. Quick start
3. The six guided prompts — start with /empezar_aqui
4. What you can ask
5. Read this before you quote a number

Part 2 — Understanding MCP (educational)

1. What is MCP? Tools, resources and prompts
2. What is datos.gob.do?

Part 3 — What this server exposes

1. Tools (24, plus 3 optional)
2. Resources (3) and one resource template
3. Prompts (6)
4. Which primitive to reach for

Part 4 — Why this server exists

1. How it compares with other CKAN MCP servers

Part 5 — Technical reference

1. Installation and client configuration
2. What the answers tell you about themselves
3. Security and environment variables
4. Architecture
5. Measured limitations

Part 6 — Development

1. Development, testing and the MCP Inspector
2. Contributing, credits, how to cite, licence
3. Privacy Policy

---

---

Part 1 — Start here

1. What this is, in plain words

The Dominican government publishes thousands of files: public payrolls, budget execution, hospital activity, migration flows, procurement contracts, electricity losses, forest fires. It is all public. Almost nobody reads it, because reading it means knowing which of 266 institutions published what, finding the file, downloading a spreadsheet with the header on row 3, and knowing what to do next.

This server hands that whole job to your AI assistant. You ask in your own words. The assistant finds the dataset, downloads the file from the institution's own server, figures out the columns, runs the sum or the average, and tells you the answer along with where it came from and what it had to leave out.

Three things worth knowing up front, because they shape everything else:

• It only reads. Nothing here can modify the portal or publish anything. There is no login and no password.
• It runs on your computer, alongside your assistant. Your questions do not pass through a server owned by this project.
• It tells you when the data is bad. Roughly half the catalog cannot be downloaded programmatically at all, and the tools say so instead of inventing a number. That is the point of the whole design, not a caveat buried at the bottom.

2. Quick start

You need uv, a small tool that runs Python programs without you installing anything else. On macOS or Linux, paste this into a terminal:

curl -LsSf https://astral.sh/uv/install.sh | sh

On Windows, follow the uv installation page.

Then tell your assistant about the server.

Claude Desktop. Open Settings → Developer → Edit Config, or edit the file directly:

• macOS: ~/Library/Application Support/Claude/claude_desktop_config.json
• Windows: %APPDATA%\Claude\claude_desktop_config.json

Paste this, replacing YOUR_USERNAME:

{

"mcpServers": { "datosgobdo": { "command": "/Users/YOUR_USERNAME/.local/bin/uvx", "args": ["dominican-open-data-mcp"] } } }

Use the full path to uvx — Claude Desktop does not read your shell's PATH. Then quit Claude Desktop completely (Cmd+Q on macOS, not just closing the window) and reopen it. Under Settings → Developer you should see datosgobdo running.

Nothing else is required. If you later want to change a setting — the network guard, the cache directory — it goes in an "env" block inside this file, not in your shell: see §13.

Claude Code. One line:

claude mcp add datosgobdo -- uvx dominican-open-data-mcp

Anything else. Same idea: register uvx as the command with dominican-open-data-mcp as its argument. The MCP clients directory lists which clients support which features. Full options — dev versions, local clones, hosted mode — are in §13.

3. The six guided prompts — start with /empezar_aqui

Twenty-four tools is not an invitation. Someone who has never seen this catalog has no way to know that payrolls, budget execution and public investment are the three things it covers best.

So the server ships six prompts: ready-made questions, written to encode the habits that took a full catalog audit to learn. In Claude Code and Claude Desktop they appear as slash commands. Type:

/empezar_aqui

and the assistant will introduce you to the portal, tell you what it covers well, propose three concrete questions you could ask next, and warn you up front about what cannot be downloaded.

The other five take one argument each:

If your client does not show prompts as slash commands, see its entry in the MCP clients directory — prompt support is optional for clients, and the MCP Inspector can always show and run them.

4. What you can ask

Plain questions, in Spanish or English. Some that work today:

*How many datasets are on the datos.gob.do portal, and which institutions publish the most?*

*Find the five most relevant budget datasets and tell me which institution publishes each one.*

*How much does the Judicial Branch spend on salaries?*

*How many active employees does the Ministry of Agriculture have in April 2026, broken down by employment status?*

That last one is worth pausing on, because it is the kind of question the whole analytics layer exists for. The Agriculture payroll is a CSV with 826,000 rows and 94 MB — far too big to paste into a conversation. The server downloads it once, converts it to a columnar cache, and answers with a grouped aggregation: 6 status types, roughly 8,915 employees. The first call takes about 14 seconds; every later question about the same file answers in under half a second.

*Compare FONDOMARENA's approved versus executed budget over the last three years.*

*What columns does the Ministry of Interior's vehicle-theft dataset have?*

*List the ten most recently updated datasets.*

Who this tends to be useful for: data journalists who would otherwise write a scraper; researchers who need programmatic access; transparency groups tracking budget execution and procurement; developers prototyping on public data; public officials checking what their own institution already publishes; and anyone curious about how the state actually operates.

5. Read this before you quote a number

This catalog has real defects, and they were measured — a census of the whole thing on 2026-08-08, one resource per dataset, 1,056 resources over real MCP sessions. Four findings change how you should read any figure you get from here:

About half the catalog cannot be downloaded by a program. 561 of 1,056 resources can be read (53.1 %), up from 540 in the census: 0.14.0's format work recovered 21 of them, re-measured against the live portal on 2026-08-13. The largest single cause of the rest is not this server and no version of it can fix it: 360 resources across 98 institutions sit behind a site configuration that refuses programmatic downloads of the files those same institutions publish as open data. From the same address, 21 other government hosts behind the same CDN answer normally — so it is per-site configuration, not our network. A further 15 links are dead and 6 files are unreadable at any encoding.

One in three multi-format datasets contradicts itself. Of 528 datasets whose formats could be compared, 176 disagree on row count or column count. One example: the Treasury's recaudaciones-sirite-2021-2025 has 971,818 rows as CSV and 197,338 as ODS. A citizen downloading the ODS and a journalist downloading the CSV would quote different numbers from the same official dataset. Practical rule: check more than one format before you publish a total.

Numbers are often stored as text. 93 of the 540 resources readable in that census hold numeric columns as text, usually because a handful of cells say N/A or #REF!. The tools read such a column as numbers where each value permits it and report what that cost — see §14. Read values_excluded before quoting the total.

No dataset declares how often it is updated. The periodicidad field is empty in all 1,056. A dataset labelled "2018-2026" may have been fed last month or frozen two years ago; you have to infer freshness from the last period that actually has data.

None of this is a reason not to use the catalog. It is a reason to cite it precisely — which is what /verificar_fuente and the self-describing response fields are for.

---

---

Part 2 — Understanding MCP

6. What is MCP? Tools,