# Querying SIDRA through MCP in Claude and ChatGPT — IBGE inside the assistant, with provenance Sidney Bissoli 2026-09-02 - [MCP in three lines](#mcp-in-three-lines) - [Connecting, client by client](#connecting-client-by-client) - [claude.ai (and Claude Desktop, through the same account)](#claudeai-and-claude-desktop-through-the-same-account) - [Claude Code](#claude-code) - [Claude Desktop (local server)](#claude-desktop-local-server) - [ChatGPT](#chatgpt) - [Cursor, VS Code and Gemini CLI](#cursor-vs-code-and-gemini-cli) - [One question, end to end](#one-question-end-to-end) - [Step 1 — `ibge_sidra_tabelas`: which table?](#step-1--ibge_sidra_tabelas-which-table) - [Step 2 — `ibge_sidra_metadados`: is it the right one?](#step-2--ibge_sidra_metadados-is-it-the-right-one) - [Step 3 — `ibge_sidra`: the data](#step-3--ibge_sidra-the-data) - [What travels with the number](#what-travels-with-the-number) - [In Deep Research: `search` and `fetch`, live](#in-deep-research-search-and-fetch-live) - [Before publishing a number](#before-publishing-a-number) *Published 2 September 2026. Every figure here was captured live, through the production server, on the publication date; none is illustrative. The source URL sits next to each one.* Ask an AI assistant *“what is Brazil’s unemployment rate?”* and it answers with a plausible number from training: maybe the right one, maybe two years old, with no source. Connect the same assistant to SIDRA — IBGE’s statistical table bank — and the answer changes in kind: the table, the quarter, the URL that reproduces the query and the extraction date come with it. Anyone can retrace the path. This text is the **how**. It shows what it takes to connect [`ibge-br-mcp`](https://github.com/SidneyBissoli/ibge-br-mcp) — an [MCP](https://modelcontextprotocol.io) server over IBGE’s public APIs — in each client that speaks the protocol (claude.ai, Claude Desktop, Claude Code, ChatGPT, Cursor, VS Code, Gemini CLI), then asks one real question end to end, with the output of every step pasted here. On *which* table to pick and how to be sure it is the right one, I wrote earlier: [Finding the right SIDRA table](../../..\en/blog/posts/sidra-tabela-certa/). The two texts complete each other; this one is about the tool. ## MCP in three lines The [Model Context Protocol](https://modelcontextprotocol.io) is an open standard for an AI assistant to call outside tools. A **server** exposes the tools (here: find a table, read its metadata, query the data); a **client** — the app you talk to — calls them when the question asks for it. The model decides which tool and which parameters; the server runs the real query against IBGE’s API and returns the result with provenance. `ibge-br-mcp` comes in two forms, and the only difference is where it runs: | Form | Address | When | |:---|:---|:---| | **Remote** (hosted, nothing to install) | `https://ibge.sidneybissoli.com/mcp` | claude.ai, ChatGPT, and any client that accepts a remote server. No key, no sign-up. | | **Local** (runs on your machine, via Node.js) | `npx -y ibge-br-mcp` | Claude Desktop, Claude Code, editors — when you prefer the calls to leave from your own computer. | Same tools in both; same APIs behind them. The remote form is a Cloudflare Worker bound to the package published on [npm](https://www.npmjs.com/package/ibge-br-mcp); the local form is the package itself. ## Connecting, client by client Each block below was checked against the client’s official documentation on the publication date; where I ran the steps myself, I say so. ### claude.ai (and Claude Desktop, through the same account) Custom connectors over remote MCP are available on the Free, Pro, Max, Team and Enterprise plans (Free allows one connector), per [Anthropic](https://support.claude.com/en/articles/11175166-getting-started-with-custom-connectors-using-remote-mcp). 1. **Settings → Connectors → Add custom connector.** 2. Name `IBGE`, URL `https://ibge.sidneybissoli.com/mcp`. Leave the advanced settings (OAuth Client ID/Secret) blank: the server requires no authentication. 3. In the conversation, **+** → **Connectors** → switch IBGE on. The connector lives in the account, not in the app: Claude Desktop and Cowork see it without any further setup. This is the path I use daily — every number in this text came through it. ### Claude Code One line in the terminal, for the remote or the local form: ``` bash claude mcp add --transport http ibge https://ibge.sidneybissoli.com/mcp claude mcp add --transport stdio ibge-local -- npx -y ibge-br-mcp ``` `--scope project` writes a `.mcp.json` at the project root, which can be committed and applies to everyone who opens the folder: ``` json { "mcpServers": { "ibge": { "type": "http", "url": "https://ibge.sidneybissoli.com/mcp" } } } ``` `claude mcp list` shows the status; so does `/mcp` inside a session. [Documentation](https://code.claude.com/docs/en/mcp). ### Claude Desktop (local server) Besides the account connector, Desktop runs local servers from a file: **Settings → Developer → Edit Config** opens `claude_desktop_config.json` (`~/Library/Application Support/Claude/` on macOS, `%APPDATA%\Claude\` on Windows). Requires Node.js 18 or newer. ``` json { "mcpServers": { "ibge-br-mcp": { "command": "npx", "args": ["-y", "ibge-br-mcp"] } } } ``` Restart the app; the server shows up under **+ → Connectors**. The `-y` matters: without it, `npx` stops on first run waiting for a confirmation nobody sees. [Documentation](https://modelcontextprotocol.io/docs/develop/connect-local-servers). ### ChatGPT ChatGPT has two paths, and the server serves both. I ran both on 2 September 2026, with the remote URL, no authentication. **Developer mode** (Pro, Plus, Business, Enterprise and Education plans, on the web, per [OpenAI](https://developers.openai.com/api/docs/guides/developer-mode)): 1. **Settings → Security and login → Developer mode**, on. 2. **Settings → Apps → Create**: name `IBGE`, a description, MCP server URL `https://ibge.sidneybissoli.com/mcp`, Authentication **No authentication**, tick that you trust the app → **Create**. The URL was accepted right away. 3. In the conversation, **+ → Developer mode**, tick IBGE and ask explicitly: *“use the IBGE app to…”*. Any tool on the server can be called; the `ibge_*` ones are the ones that bring data. **Deep Research.** ChatGPT’s deep research only accepts servers that expose two tools with names fixed by OpenAI, `search` and `fetch` ([doc](https://developers.openai.com/api/docs/mcp)). `ibge-br-mcp` exposes them since version 4.3.0, on top of the `ibge_*` tools: `search` ranks the question against SIDRA tables, municipalities and the known indicators; `fetch` returns the document as Markdown with the public URL ChatGPT cites. With the app created above, **+ → Deep research → Sources** lists IBGE; tick it. In the run I did, the report called `search`, picked Table 6579 and assembled the population of the 34 municipalities of the Belo Horizonte metropolitan region, each line with its source. ### Cursor, VS Code and Gemini CLI Same idea, different files. All three accept the remote form directly. **Cursor** — `.cursor/mcp.json` in the project or `~/.cursor/mcp.json` ([doc](https://cursor.com/docs/context/mcp)): ``` json { "mcpServers": { "ibge": { "url": "https://ibge.sidneybissoli.com/mcp" } } } ``` **VS Code** (Copilot) — `.vscode/mcp.json`; the key is `servers`, not `mcpServers` ([doc](https://code.visualstudio.com/docs/copilot/chat/mcp-servers)): ``` json { "servers": { "ibge": { "type": "http", "url": "https://ibge.sidneybissoli.com/mcp" } } } ``` **Gemini CLI** — `~/.gemini/settings.json`; for Streamable HTTP the key is `httpUrl` (`url` is the older SSE transport) ([doc](https://geminicli.com/docs/tools/mcp-server/)): ``` json { "mcpServers": { "ibge": { "httpUrl": "https://ibge.sidneybissoli.com/mcp" } } } ``` For the local form in any of them: `"command": "npx", "args": ["-y", "ibge-br-mcp"]` in place of the URL. ## One question, end to end Once connected, the flow is conversation. The question I asked, in claude.ai, with the connector on: > What is the unemployment rate by state in the latest available quarter > — and where is it highest? The assistant does not know by heart which table answers that, and should not guess. The server’s instructions walk it through SIDRA’s three steps — the same ones as in the [earlier text](../../..\en/blog/posts/sidra-tabela-certa/): find the table, read the metadata, query. Below, every call it made and what came back. ### Step 1 — `ibge_sidra_tabelas`: which table? Searching `desocupação` (unemployment): 33 tables, from different surveys — the discontinued Monthly Employment Survey, the annual PNAD Contínua, the SDG indicators. Narrowing with `pesquisa: "trimestral"` (quarterly) leaves 7, all from the quarterly PNAD Contínua: | Code | Name | |---:|:---| | **4099** | Unemployment and labour underutilisation rates, in the reference week, persons aged 14 and over | | 6396 | … by sex | | 6397 | … by age group | | 6468 | Unemployment rate — total, coefficient of variation, change on the previous quarter and on the same quarter a year earlier | | 6467, 6483, 6484 | Unemployment level; combined underutilisation rates | Source: `https://servicodados.ibge.gov.br/api/v3/agregados`. 4099 is the candidate: no breakdown by sex or age, and the rate itself rather than its change. ### Step 2 — `ibge_sidra_metadados`: is it the right one? The metadata for 4099 say what the table is before any number: - Survey: **quarterly PNAD Contínua**; subject Labour; quarterly, from Q1 2012 to **Q2 2026**. - Territorial levels: Brazil, Major Region, **Federation Unit** (state), Metropolitan Region, Municipality, Integrated Development Region. - Variable **4099** — Unemployment rate, in the reference week, persons aged 14 and over, in **%**. The other seven are coefficients of variation and combined underutilisation rates. Source: `https://servicodados.ibge.gov.br/api/v3/agregados/4099/metadados`. This is where it is settled: states are there, the latest quarter is there, the variable is the rate and not the level or the change. Query it. ### Step 3 — `ibge_sidra`: the data The call, as the assistant assembled it: ibge_sidra(tabela="4099", variaveis="4099", nivel_territorial="3", periodos="last", estatisticas=true, topN=5) `estatisticas: true` makes the server compute the distribution over every row before returning — 27 federation units, one call: | | | |:---|---:| | Period | Q2 2026 | | Median across the 27 states | 5.6% | | Minimum | Santa Catarina, 2.1% | | Maximum | **Amapá, 9.8%** | | Highest | Amapá 9.8 · Bahia 9.1 · Piauí 8.3 · Pernambuco 8.3 · Alagoas 7.9 | | Lowest | Santa Catarina 2.1 · Mato Grosso 2.2 · Espírito Santo 2.3 · Rondônia 2.6 · Mato Grosso do Sul 2.7 | Source: `https://apisidra.ibge.gov.br/values/t/4099/n3/all/v/4099/p/last`. A second call with `nivel_territorial="1"` gives Brazil in the same quarter: **5.4%** (source: `https://apisidra.ibge.gov.br/values/t/4099/n1/all/v/4099/p/last`). ### What travels with the number Every server response carries a provenance block. Step 3’s, as returned: ``` json { "source": "IBGE — SIDRA (Banco de Tabelas Estatísticas)", "source_url": "https://apisidra.ibge.gov.br/values/t/4099/n3/all/v/4099/p/last", "data_vintage": "2º trimestre 2026", "retrieved_at": "2026-09-02T20:15:51-03:00", "citation": "Fonte: IBGE — SIDRA, Tabela 4099 (PNAD Contínua - Taxa de desocupação (trimestral)), https://apisidra.ibge.gov.br/values/t/4099/n3/all/v/4099/p/last, extraído em 02/09/2026.", "license": "Dados abertos do Poder Executivo federal (Lei 12.527/2011; Decreto 8.777/2016)" } ``` The `source_url` opens in a browser and returns the same 27 values. What the assistant writes for you is interpretation; this is the data. ## In Deep Research: `search` and `fetch`, live The two tools of OpenAI’s contract follow the same principle, in a shape ChatGPT can cite. Called through the same connector, on the same date: `search("taxa de desocupação trimestral")` returned ten results, each with `id`, `title` and `url`; the first ones: | id | title | url | |:---|:---|:---| | `ind:desemprego` | Taxa de Desocupação (indicador desemprego) | sidra.ibge.gov.br/tabela/4099 | | `sidra:6468` | Tabela 6468 — Taxa de desocupação … variações em relação ao trimestre anterior | sidra.ibge.gov.br/tabela/6468 | | `sidra:4099` | Tabela 4099 — Taxas de desocupação e de subutilização da força de trabalho … | sidra.ibge.gov.br/tabela/4099 | `fetch("sidra:4099")` returned the document: the table’s metadata as Markdown (survey, period 201201–202602, levels, the eight variables) and the canonical URL `https://sidra.ibge.gov.br/tabela/4099` — which is what shows up in the report’s citation. The Deep Research report is written by the model; what holds it up is that `url` on every source. ## Before publishing a number The tool removes the manual work; it does not remove judgement. Three checks that stay yours, with or without an assistant: 1. **Which survey the table belongs to.** Step 2 said *quarterly PNAD Contínua*. An unemployment rate from the **annual** PNAD Contínua (Table 4562) or from the old Monthly Employment Survey (2179) has a different coverage and period — and would show up in the same search as step 1. The [earlier text](../../..\en/blog/posts/sidra-tabela-certa/) is about this. 2. **Which period.** `data_vintage` says *Q2 2026*; `periodos: "last"` takes the latest published, which changes with every release. A number without its quarter is a loose number. 3. **Which unit and level.** Rate in %, persons aged 14 and over, by state. The median across states (5.6%) is not Brazil’s rate (5.4%): the first counts each state as one; the second is weighted by the labour force. The provenance block exists so that these checks can be made by whoever reads, not only by whoever wrote. ------------------------------------------------------------------------ *Server: [`ibge-br-mcp`](https://github.com/SidneyBissoli/ibge-br-mcp) · hosted endpoint at `https://ibge.sidneybissoli.com/mcp` · npm package [`ibge-br-mcp`](https://www.npmjs.com/package/ibge-br-mcp) · MIT. IBGE data under Brazil’s open-data framework (Law 12.527/2011 and Decree 8.777/2016): free use, with the obligation to credit the source.*