Querido Diario MCP Server

Querido Diário MCP Server

A local-first Model Context Protocol server that gives AI clients structured, read-only access to Brazilian municipal official gazettes indexed by Querido Diário.

No API key · Runs locally · Read-only · No telemetry · Open source

🇧🇷 Versão em português

The package is published on PyPI and the MCP Registry.

What it does

Querido Diário provides searchable municipal official gazettes through a public API. This server exposes a small, typed MCP interface on top of that API so an MCP client can:

• resolve Brazilian municipalities and their IBGE territory IDs;
• search indexed official gazettes by text, municipality and date range;
• receive structured results that are easier for agents to inspect and combine with other tools.

The server runs as a local stdio subprocess and only performs read-only HTTPS requests to the public Querido Diário API.

Quick start

Requires Python 3.12+ and uv.

No repository clone is required:

uvx querido-diario-mcp-server

This starts the MCP server on stdio. It is intended to be launched by an MCP client rather than used as an interactive terminal application.

Client configuration

All supported clients use the same command:

uvx querido-diario-mcp-server

Claude Desktop / Claude Code

{
  "mcpServers": {
    "querido-diario": {
      "command": "uvx",
      "args": ["querido-diario-mcp-server"]
    }
  }
}

Use this in claude_desktop_config.json for Claude Desktop or a project-level .mcp.json for Claude Code.

Cursor

{
  "mcpServers": {
    "querido-diario": {
      "command": "uvx",
      "args": ["querido-diario-mcp-server"]
    }
  }
}

Add it to .cursor/mcp.json or your global Cursor MCP configuration.

Codex CLI

[mcp_servers.querido-diario]
command = "uvx"
args = ["querido-diario-mcp-server"]

Other MCP clients that support local stdio servers can launch the same command.

Tools

The server intentionally exposes a small, read-only tool surface.

search_gazettes uses the simple-query-string syntax supported by the upstream search API. Examples include +required, -excluded and "exact phrase".

Example

A user can ask:

Find mentions of ACME Ltda in Porto Alegre gazettes
from January through July 2026.

An MCP client can then execute:

1. search_cities(city_name="Porto Alegre")
   -> territory_id: "4314902"

2. search_gazettes(
       query='"ACME Ltda"',
       territory_ids=["4314902"],
       published_since="2026-01-01",
       published_until="2026-07-31",
   )

The result is returned as structured tool output for the client to summarize, filter or combine with other tools.

Architecture

MCP client
   |
   | stdio / MCP
   v
server.py
   |
   | validated tool arguments
   v
client.py
   |
   | typed read-only HTTPS requests
   v
Querido Diário public API

The implementation keeps the protocol boundary separate from the HTTP integration:

src/querido_diario_mcp_server/
    __init__.py   package version and CLI entry point
    config.py     environment-driven configuration
    errors.py     integration error hierarchy
    models.py     typed Pydantic domain models
    client.py     async HTTP client with no MCP dependency
    server.py     MCP lifespan, validation and tool definitions

A single httpx.AsyncClient connection pool is created for the server lifecycle and reused across tool calls. The HTTP client can be tested independently of MCP, while integration tests exercise the real MCP server in process.

Design principles

• Read-only by design. The integration only calls a fixed set of upstream GET endpoints.
• Local-first. There is no hosted middleware, proprietary backend or account requirement.
• Typed boundaries. API responses and MCP outputs use explicit Pydantic models.
• Small tool surface. The server exposes only the operations required to discover cities and search gazettes.
• Explicit failures. Upstream failures are converted to concise MCP tool errors rather than raw tracebacks or HTML responses.
• No arbitrary URL fetching. Tool callers cannot turn the server into a generic HTTP or SSRF primitive.
• No telemetry. The server does not collect usage data.

Security

The server never automatically follows gazette URLs returned by the upstream API and does not accept arbitrary target URLs from tool callers. Territory IDs, dates, pagination and sort options are validated before use.

For vulnerability reporting and the supported security scope, see SECURITY.md.

Configuration

Normal users do not need to set any environment variables.

Development

Clone the repository only if you want to contribute or work on the implementation:

git clone https://github.com/lucaspmgomess/querido-diario-mcp-server.git
cd querido-diario-mcp-server
uv sync

Run the same checks enforced by CI:

uv run ruff check .
uv run ruff format --check .
uv run pyright
uv run pytest --cov

Run the server from the checkout:

uv run querido-diario-mcp-server

Inspect the MCP tools interactively:

uv run mcp dev src/querido_diario_mcp_server/server.py:mcp

Test strategy

Unit tests mock the HTTP boundary with httpx.MockTransport, so the automated suite does not depend on the public API.

Coverage includes:

• successful city and gazette searches;
• query serialization and repeated territory IDs;
• date ranges, pagination and sorting;
• empty result sets;
• 4xx and 5xx upstream failures;
• malformed responses;
• connection errors and timeouts;
• MCP tool discovery, schemas and structured outputs;
• validation errors and upstream failures at the MCP boundary.

Manual smoke-test scripts are available for validating the real production API and the full MCP-to-API path:

uv run python scripts/smoke_test_api.py
uv run python scripts/smoke_test_mcp.py

Project status

The project is currently in beta. The MCP tool surface is intentionally small and changes are kept conservative. Release history is documented in CHANGELOG.md.

The following are intentionally outside the current scope:

• arbitrary URL fetching;
• automatic PDF or full-text download;
• OCR;
• write operations;
• crawling and background jobs;
• built-in LLM summarization;
• a web interface.

Relationship to Querido Diário

This is a community-built, unofficial project.

Querido Diário is maintained by Open Knowledge Brasil and its community. This repository is not an official Open Knowledge Brasil project, is not endorsed or maintained by the organization, and only consumes the project's public API.

Relevant upstream repositories:

• okfn-brasil/querido-diario
• okfn-brasil/querido-diario-api
• okfn-brasil/querido-diario-deployment

Additional notes about upstream endpoint history are kept in docs/upstream-api-history.md.

Contributing

Issues and pull requests are welcome. See CONTRIBUTING.md for the development workflow, project scope and review checklist.

License

MIT