Senado Br Mcp Cloudflare MCP Server

Senado Brasil MCP Server

🇧🇷 Leia em Português

A public, hosted MCP server that gives AI assistants live, structured access to Brazilian Senate open data — no installation, no account, no API key. Point your MCP client at the hosted endpoint and start asking about senators, bills, votes, expenses, and more. It runs on Cloudflare Workers over Streamable HTTP.

It exposes 67 tools, 4 prompts, and 5 resources across two domains:

• Legislative — senators; bills and their tramitation; votes; committees; plenary sessions, results and presidential vetoes; party-bloc voting orientation; speeches and stenographic transcripts; blocs and leadership; federal legislation; and citizen participation via the e-Cidadania portal.
• Administrative — CEAPS parliamentary-quota expenses; housing allowance; civil servants and payroll; overtime; interns; procurement contracts and biddings; outsourced staff; petty-cash funds; and budget execution.

Data comes from three official sources — the legislative open-data API, the administrative open-data API, and the e-Cidadania portal. All tool responses are in Portuguese (pt-BR). See CHANGELOG.md for release history.

See it in action

Point a client at the endpoint and ask in plain language — English or Portuguese:

• "How did São Paulo's senators vote in the most recent floor votes?" → senado_search_votacoes
• "Show the legislative progress of PEC 45/2019 (a constitutional amendment proposal)." → senado_buscar_materias + senado_obter_materia
• "How much was spent on the CEAPS parliamentary allowance in 2024, broken down by expense type?" → senado_ceaps

The answers come live from the Senate's official open-data APIs — exact figures with provenance, not numbers guessed from training data.

Use it (hosted — no setup)

This is a remote, hosted, open-access server. To use it, point any MCP client at the Streamable HTTP endpoint — no install, no account, no API key, no configuration:

https://senado.sidneybissoli.com/mcp

OpenAI / ChatGPT app surface

For OpenAI Apps SDK submission and review, the Worker also exposes a curated MCP surface:

https://senado.sidneybissoli.com/mcp/openai-app-v2

This endpoint intentionally keeps the full public MCP server intact at /mcp, but limits tool discovery to 27 high-signal, intent-oriented tools for ChatGPT app use. /mcp/openai-app remains available as a legacy alias, but new ChatGPT app configurations should use /mcp/openai-app-v2 so clients fetch the current tool schema. The tools still call the same handlers and return the same provenance envelope; only the advertised surface is narrower. Any ChatGPT app listing should present this as an independent open-data research app, not as an official Senate, OpenAI or ChatGPT connector.

For ChatGPT Apps, those 27 tools also advertise a shared MCP Apps UI template at ui://senado-br-mcp/openai-app-dashboard-v2.html. The self-contained widget renders the returned structuredContent as a compact dashboard with metrics, main records, and source/provenance, without adding another model-visible data tool.

Public legal URLs for app review:

• Privacy policy: https://senado.sidneybissoli.com/privacy
• Terms of use: https://senado.sidneybissoli.com/terms

Install (any client)

For clients that launch MCP servers as a command — and for one-command setup — use the mcp-remote bridge. No build, no config, no key:

npx -y mcp-remote https://senado.sidneybissoli.com/mcp

• One-click (LobeHub): open the server page and click Install.
• Native remote URL (Claude Desktop/Code and other Streamable-HTTP clients): see Connecting MCP Clients.

Everything below Architecture (Prerequisites, Setup, Deploy) is only for optionally self-hosting your own instance — it is not required to use this public server.

Run locally (npx · stdio)

Prefer not to route queries through a third-party host (e.g. a newsroom policy)? The same server also runs as a local stdio process that talks directly to the official government APIs — same 67 tools, same provenance envelope, no Cloudflare in the loop. This is the npm/stdio channel, published as senado-br-mcp.

Point a command-based client (Claude Desktop/Code, etc.) at the package — npm fetches and runs it, no clone or build:

{
  "mcpServers": {
    "senado-br": {
      "command": "npx",
      "args": ["-y", "senado-br-mcp"]
    }
  }
}

To run it directly or hack on it, use the source checkout instead:

git clone https://github.com/SidneyBissoli/senado-br-mcp-cloudflare
cd senado-br-mcp-cloudflare
npm install
npm run build
node dist/cli.js   # serves MCP over stdio (Ctrl+C to stop)

Parity with the hosted server: the legislative and administrative tools are identical (same upstream APIs, same throttle/cache/provenance) — locally the L1 Cloudflare cache is a no-op, but the L0 in-memory cache still works, so results are the same. The only difference is the e-Cidadania list/corpus tools: without D1 they fall back to a live scrape of the ~5 REST highlights, flagged via meta.fonte / possivelDesatualizacao; the detail tools (obter_*) are identical. Logs go to stderr — stdout carries only the JSON-RPC protocol stream.

Agent Skill (optional)

This repo bundles a Claude Agent Skill at .claude/skills/senado-br/ that teaches Claude when to reach for this server and how to use its 67 tools well — a themed tool map, common question→tool playbooks, the provenance contract, and gotchas (dates, the codigoMateria bridge, e-Cidadania's open-set listing, pagination). It points back to the server's own senado://catalogo / senado://guia resources rather than duplicating them.

Claude Code auto-discovers it when you work in this repo. To use it elsewhere, copy .claude/skills/senado-br/ into your ~/.claude/skills/, or zip the folder and upload it in claude.ai (Settings → Features). The skill assumes the senado-br MCP server is connected (hosted or via npx).

Architecture

• Runtime: Cloudflare Workers (ESM)
• Transport: Streamable HTTP (MCP spec 2025-03-26) via createMcpHandler from agents/mcp
• Protocol: MCP over JSON-RPC — /mcp handles the full public server; /mcp/openai-app-v2 exposes a curated 27-tool profile plus a shared MCP Apps widget for OpenAI app review/submission (/mcp/openai-app remains as a legacy alias)
• SDK: @modelcontextprotocol/sdk 1.26.0+ (per-request McpServer instances)
• Validation: Zod schemas for all tool inputs
• Caching: 2-layer (L0 memory + L1 Cache API) with SHA-256 keying
• e-Cidadania store: D1 database refreshed by a Cron Trigger (every 2h) — list tools read from D1 with a live-scrape fallback and a staleness flag; detail tools stay live with write-through (see e-Cidadania)
• Rate limiting: Token bucket — global (8 req/s) + per-client (2 req/s)
• Upstream throttle: Max 6 concurrent requests, 10s timeout, retry with exponential backoff
• Auth: Optional Bearer token (set the API_KEY secret; open access when unset). Constant-time comparison.
• Observability: Structured JSON logging + in-memory counters at /metrics; per-tool call telemetry (selection, error rate, cache-vs-live) in Cloudflare Analytics Engine, PII-free
• Liveness: Runs on Cloudflare's own global network behind a custom domain — no third-party host that can go dark. Public /health and /status (version + last-deploy id/timestamp) make uptime and the current build verifiable; the status badge above pings the live endpoint
• Tests: Vitest unit tests for parsers, helpers, cache, throttle, and auth

Self-hosting (optional)

Not needed to use the server — it is already hosted at https://senado.sidneybissoli.com/mcp (open access). Follow this section only if you want to run your own private instance.

Prerequisites

• Node.js 20+
• Wrangler CLI v4+
• Cloudflare account

Setup

1. Install dependencies

npm install

2. Create KV namespace

# Create the KV namespace
wrangler kv namespace create CACHE_KV

# Note the ID from the output, e.g.:
# { binding = "CACHE_KV", id = "abc123..." }

3. Configure wrangler.toml

Replace the placeholder KV namespace ID:

[[kv_namespaces]]
binding = "CACHE_KV"
id = "YOUR_KV_NAMESPACE_ID_HERE"

Optionally set ALLOWED_ORIGIN to restrict CORS:

[vars]
ALLOWED_ORIGIN = "https://your-app.example.com"

The e-Cidadania pipeline needs a D1 database and a Cron Trigger (both already declared in wrangler.toml — replace the database ID):

[[d1_databases]]
binding = "ECIDADANIA_DB"
database_name = "senado-ecidadania"
database_id = "YOUR_D1_DATABASE_ID_HERE"

[triggers]
crons = ["0 */2 * * *"]

Create the database (paste the returned ID above) and apply the schema:

npx wrangler d1 create senado-ecidadania
npx wrangler d1 migrations apply senado-ecidadania --remote

The list tools fall back to live scraping when D1 is empty, so the server works before the first Cron run.

4. (Optional) Enable authentication

wrangler secret put API_KEY
# Clients must then send: Authorization: Bearer <key>
# When API_KEY is not set, the server is open access.

5. Local development

npm run dev
# Dev server runs locally on port 8787 (local only).
# The public MCP endpoint is https://senado.sidneybissoli.com/mcp

6. Tests and typecheck

npm test             # run all tests once
npm run test:watch   # watch mode
npm run typecheck    # tsc --noEmit

7. Deploy

npm run deploy
# Serves at https://senado.sidneybissoli.com (custom domain) and
# https://senado-br-mcp.sidneybissoli.workers.dev (workers.dev fallback)

Endpoints

MCP Request Examples

All requests go to POST /mcp with JSON-RPC 2.0 format.

List available tools

{
  "jsonrpc": "2.0",
  "id": 1,
  "method": "tools/list"
}

Call a tool — List senators from SP

{
  "jsonrpc": "2.0",
  "id": 2,
  "method": "tools/call",
  "params": {
    "name": "senado_listar_senadores",
    "arguments": {
      "uf": "SP",
      "emExercicio": true
    }
  }
}

Call a tool — Search bills by keyword

{
  "jsonrpc": "2.0",
  "id": 3,
  "method": "tools/call",
  "params": {
    "name": "senado_buscar_materias",
    "arguments": {
      "palavraChave": "inteligência artificial",
      "tramitando": true
    }
  }
}

Call a tool — Get recent plenary votes

{
  "jsonrpc": "2.0",
  "id": 4,
  "method": "tools/call",
  "params": {
    "name": "senado_search_votacoes",
    "arguments": {
      "dias": 7
    }
  }
}

Call a tool — Most popular citizen ideas

{
  "jsonrpc": "2.0",
  "id": 5,
  "method": "tools/call",
  "params": {
    "name": "senado_ecidadania_listar_ideias",
    "arguments": {
      "ordenarPor": "apoios",
      "ordem": "desc",
      "status": "aberta"
    }
  }
}

Upstream API Endpoints

The server consumes two classes of upstream endpoints from the Senado API:

Legacy endpoints (.json suffix, PascalCase responses)

Used by Groups A, E, F, H, I, J, K, L, M, N. The .json suffix is appended automatically by upstream.ts. None of these is marked deprecated upstream.

v3 endpoints (flat JSON arrays/objects, camelCase)

Used by Groups B, C, D. Dates must be in ISO format (YYYY-MM-DD) — tools accept YYYYMMDD and convert. The codigoMateria query param bridges legacy matéria codes to v3 processes.

Administrative API (adm.senado.gov.br/adm-dadosabertos, flat snake_case JSON)

Used by Groups O, P, Q, R via admFetch (no .json suffix; HTTP 404 treated as empty collection). Base URL configurable via SENADO_ADM_BASE_URL.

e-Cidadania (D1-backed, Cron-refreshed)

The e-Cidadania list data is persisted in a D1 database (ecidadania_current/_history/_scrape_runs, discriminated by entidade; plus ecidadania_comentarios for the audiência comment level and ecidadania_detalhe_cursor for the resumable detail backfill — added in schema v2) and read from there instead of being scraped on every call. Three cadences write into it:

• a daily off-Worker GitHub Action owns the full corpus of the three live entities (consultas, eventos, ideias; see below) — the source of truth. Daily (not weekly) because the first-seen series MIN(scraped_at) is the only measurable entry-rhythm signal and every skipped day permanently shortens it (ROADMAP Etapa 2, decisão D3);
• a weekly ingest Action (.github/workflows/verify-consultas-votos.yml — historical filename) for the consultas_votos acervo: the Senado republishes the Arquimedes CSV periodically (confirmed 2026-07-20), so the weekly run re-ingests the current vintage under the same anomaly guards as the other corpora (see below);
• an in-Worker Cron Trigger (0 */2 * * *, src/scraper/pipeline.ts → refreshEcidadania) does only a targeted metric splice of the ~5 REST highlights per live entity (restcolecaomaismateria/ideia/audiencia — votos/comentários/apoios), recorded as ok-metrica so it never re-breaks the corpus baseline and never touches the long tail. In v2 the eventos splice preserves the corpus's canonical comment count (the daily crawl is the source of truth for comentarios, so the splice can't ping-pong it against the degraded REST count).

Both writers build payloads through the canonical buildXResumo builders + shared contentHash, so their rows are byte-identical. Each write:

• upserts ecidadania_current (one row per item — what the tools read),
• appends ecidadania_history only when an item's content_hash changes (time-series-ready),
• records each run in ecidadania_scrape_runs.

An anomaly guard (src/scraper/anomaly.ts, classifyRun) ensures a failed or anomalous corpus run (zero rows, or fewer than ECIDADANIA_CORPUS_MIN_PCT% of the last good run) never overwrites the last good state.

The list / analysis tools (listar_*, consultas_analise, sugerir_tema_enquete, consultas_votos) read from D1 via resolveList (src/scraper/store.ts): D1-first. Because every entity is now a full corpus, a stale corpus is served from D1 flagged (possivelDesatualizacao: true) rather than collapsing to the ~5-item live highlights (the original coverage bug); the live scrape is reserved for an empty D1 (cold start, before the first weekly run). Staleness uses ECIDADANIA_CORPUS_STALE_MAX_MIN (~10 days). Every list response carries an additive meta (fonte, lastScrapedAt, possivelDesatualizacao) so callers always see the data's real age and never get stale data silently.

The detail tools (obter_*) stay live (HTML scraped with CSS-class-targeted regex) for freshness, and write their richer payload through to ecidadania_detalhe fire-and-forget (deduped by content_hash), so detail history accrues without adding latency to the response.

Full-corpus ingestion (off-Worker)

The three live e-Cidadania corpora are owned by the daily Action (.github/workflows/ingest-ecidadania.yml), each with its own scripts/ingest-ecidadania/index-*.ts orchestrator emitting batched out-*.sql the apply step bulk-loads:

• consultas — open consultations (detailed below). In v2 each crawled matter is also enriched from its detail page (visualizacaomateria) for autoria/relator; those are immutable, so only rows not yet enriched are fetched.
• eventos — audiências/eventos from the principalaudiencia?p=N HTML listing; status comes straight from the listing block (no /processo bridge). In v2 every event is enriched from its detail page (canonical data/hora + comissaoNomeCompleto/local/descricao/pauta/convidados/videoUrl) and its AJAX comment fragment (canonical count + one ecidadania_comentarios row per comment, diffed against the stored hashes and emitted as out-eventos-comentarios-*.sql).
• ideias — ideias legislativas (~113.7k) from pesquisaideia?situacao=N&p=M, crawled per situacao bucket (the listing has no inline status) and emitted in ~10k-statement batches. In v2 the listing crawl preserves the immutable detail fields, and a separate resumable backfill (index-ideias-detalhe.ts, run via ingest:ecidadania:ideias-detalhe) fills them a chunk per run — because ~113.7k detail fetches don't fit one Action, it persists a cursor in ecidadania_detalhe_cursor and wraps around at the end.

The fourth entity, consultas_votos, is a separate historical acervo of votes-by-UF parsed from the ~33 MB Arquimedes CSV (Proposições-com-votos.csv), aggregated to one record per matéria with a votosPorUf breakdown. The CSV's "dados atualizados até" stamp becomes the provenance reference_period; it is excluded from the row hash (consultaVotoCore) so a re-ingest with unchanged votes doesn't churn _history. STATUS ATUAL is uniformly "Descontinuado", hence archival, not a migration of the open consultations. Served by senado_ecidadania_consultas_votos with provenance pointing at the CSV (ECIDADANIA_ARQUIMEDES). It is excluded from the daily job and owned by its own weekly ingest Action (.github/workflows/verify-consultas-votos.yml — the filename keeps the historical verify- prefix): the acervo was originally treated as a frozen single vintage (ROADMAP Etapa 2, decisão D1) and the weekly run only verified it, but on 2026-07-20 the Senado republished the CSV as a fresh vintage (+43 matérias, 648 updated), so the scheduled run now re-ingests the current vintage under the standard anomaly guards (empty/truncated CSV and the catastrophic floor still fail without writing; force dispatch overrides the floor). The script's verify mode (INGEST_CONSULTAS_VOTOS_VERIFY=1 / --verify) remains available as an on-demand integrity check.

The consultas job is the reference implementation:

consultas covers the full set of OPEN consultations — every matter currently in tramitação (~7.7k), not just the ~5 highlights. Confirmed on the first run: the pesquisamateria listing is in-tramitação-only, so closed/historical consultations are not captured by this source (a pre-ingestion historical backfill is out of scope). Three settled design decisions:

1. Decoupled ingestion. The open set is acquired by an off-Worker TypeScript job (scripts/ingest-ecidadania/, run by a daily GitHub Action — .github/workflows/ingest-ecidadania.yml) that paginates the HTML listing (pesquisamateria?p=1..N, the only full-coverage source for open consultations) for ids + vote counts and bulk-loads D1; the Worker only reads. The brittle, long crawl is kept out of the request/Cron path.
2. Status from /processo, not HTML. A consultation runs from presentation until the end of tramitação, so status is a function of the matter: aberta ⟺ the codigoMateria is in the /processo tramitando=S set, derived from robust JSON (never scraped). Every consultation enters as aberta (the listing only yields in-tramitação matters); on each complete run the job re-derives status for all stored rows by /processo membership (not by listing-absence, which can be transient), so a consultation whose matter leaves tramitação flips to encerrada. The encerrada/todas sets therefore grow over time; consultations that closed before the first ingestion aren't captured (out of scope). The list/analysis tools default to status: aberta.
3. Two reconciled cadences (one shared writer contract). The job reuses contentHash + the ConsultaResumo builder + classifyRun from src/scraper/, so its rows are byte-identical to the Cron's. The daily job owns the long tail; the 2h Cron keeps the ~5 hot/open highlights fresh via a targeted metric splice (recorded as ok-metrica, bypassing the corpus classifyRun baseline). Corpus freshness (possivelDesatualizacao) is computed from the last status='ok' run and uses a larger window (ECIDADANIA_CORPUS_STALE_MAX_MIN), and a stale consultas corpus is served from D1 flagged rather than collapsing back to the live highlights.

Write guards on the load: an incomplete crawl (any page failed) or an incomplete /processo status universe writes only an erro run row; even a complete crawl is rejected by a catastrophic floor (ECIDADANIA_CORPUS_MIN_PCT, default 80% of the last good corpus) to guard against a degraded page — overridable with --force / INGEST_FORCE=1 for a legitimate large shrink. Run daily via the Action, or manually:

CLOUDFLARE_API_TOKEN=… npm run ingest:ecidadania                 # writes scripts/ingest-ecidadania/out.sql
npx wrangler d1 execute senado-ecidadania --remote --file=scripts/ingest-ecidadania/out.sql

Caching

Layer architecture

Cache categories

POST caching approach

MCP uses POST for all tools/call requests. Caching POST responses is not natively supported by the Cache API, which requires GET requests. The solution:

1. Hash parameters — Tool name + sorted parameters are hashed with SHA-256
2. Synthetic GET key — A synthetic URL https://senado-br-mcp.internal/__cache/{tool}/{hash} is constructed
3. Cache API match/put — The synthetic GET URL is used with caches.default.match() and caches.default.put(), allowing standard Cache API operations on POST-originated data

This caching happens at the tool level (inside each tool's callback), not at the MCP transport level.

Provenance

Every tool attaches a provenance envelope so a result is traceable back to its official source — provenance is treated as a first-class part of the answer, not an optional extra (the audience is journalists and political-science researchers, for whom an un-sourced figure is unusable). The envelope lives in structuredContent.provenance (parseable by clients, and validated server-side at build time via ProvenanceSchema.parse — note the advertised per-tool output schema is permissive, so it is not what enforces the envelope's shape) and is mirrored as a one-line source footer in the text content for clients that only render text — the data JSON itself is not duplicated with the envelope, to keep the per-response token cost low (a fixed ≈170 chars).

Coverage spans all four upstream sources, each with its own source/citation/license (in src/utils/provenance.ts):

• Senado Federal — Dados Abertos (Legislativo) — legis.senado.leg.br/dadosabertos
• Senado Federal — Dados Abertos (Administrativo) — adm.senado.gov.br/adm-dadosabertos
• Senado Federal — Execução Orçamentária e Financeira — Arquimedes/Financeiro feed at senado.gov.br
• Senado Federal — Portal e-Cidadania — www12.senado.leg.br/ecidadania

Fields (per response — one tool, one source):

In addition to the provenance envelope, structuredContent carries a top-level attribution list — the distinct source URLs behind the response. This mirrors the naming proposed in modelcontextprotocol#711 (where attribution is a list of source references at the response level), so the server stays forward-compatible if that RFC lands; the richer provenance object remains this server's own extension.

Out-of-band mirror in _meta. The same provenance and attribution are also mirrored on the result's _meta under namespaced keys (com.sidneybissoli.senado/provenance and com.sidneybissoli.senado/attribution). The MCP spec keeps _meta for metadata about a result that should not steer the model, which is where the still-incubating trust/attribution work from #711 (split into an experimental extensions track, not yet in core) points; mirroring there gives audit and UI consumers the provenance without reading the model-facing data channel, at no model-token cost, while structuredContent keeps it visible so the model can cite the source. The mirror survives the ChatGPT-app profile minimizer (which only strips structuredContent.meta).

Field-level granularity. Most tools are single-source, so one envelope suffices. The few that merge slices in one response fill provenance.field_sources — a list of { fields, source_url, retrieved_at, … } attributing specific output fields to their real origin. Example: senado_obter_materia secao=detalhe fuses /processo/{id} (the top-level source) with the ementa from /processo and the relator from /processo/relatoria, each carrying its own retrieved_at.

retrieved_at fidelity is provided by the cache layer (cachedFetchWithMeta), which persists the fetch timestamp alongside the value, so it reflects the real upstream extraction even on a cache hit. Two exceptions report an honest live timestamp instead: the e-Cidadania list tools (read from D1) use the corpus's lastScrapedAt — the true age of the stored data — while e-Cidadania detail tools, scraped live, use the fetch time and a level-3 canonical item URL. The only path that falls back to the build-time default is the in-code static reference catalog (senado_tabelas_referencia tipos-materia), which has no upstream extraction instant.

Coverage is universal: all 67 tools carry the envelope (verify with grep -c 'resultWithProvenance(' src/tools/*.ts). The ⊕ marks in the inventory below denote the original pilot tools (votes, bills, processes); the envelope now extends to every tool, so the marks are historical.

Citable dataset (e-Cidadania participation)

Beyond the live server, this project publishes a frozen, versioned, citable dataset of the e-Cidadania participation layer (public consultations, legislative ideas, interactive events + their comments, historical votes by state) — the layer the R package congressbr never covered. Each value carries a per-field provenance envelope ({ value, sourceEndpoint, sourceField, retrievedAt, license, schemaVersion }); the data license (Dados Abertos do Senado Federal) is kept separate from the code license (MIT).

• How to cite — CITATION.cff (dataset; cite the version-DOI of the snapshot you used, the concept-DOI for the dataset across versions).
• What's in each release — CHANGELOG-dataset.md (cumulative, append-only; binds each release to its schemaVersion).
• Variable dictionary & field provenance — docs/dataset-dictionary.md (generated from src/dataset/schema.ts, the single source of truth).
• Data license — LICENSE-DATA.md.
• Cutting a release (freeze → checksums → GitHub Release → Zenodo DOI) — docs/release-runbook.md; machinery in src/dataset/, scripts/build-dataset/, and .github/workflows/release-dataset.yml.

Inventory — schema v2 (schemaVersion 2.0.0)

Each release ships one NDJSON resource per entity (one HarmonizedRecord per line: identity + a provenance envelope per field), plus a datapackage.json manifest and a copy of the dictionary. Five resources:

ⁿ = new/reopened in v2 · ᶜ = source corrected to the canonical one in v2. What v2 (2.0.0) changed — the ingestion moved from listing-only to listing + detail (+ AJAX comments for events):

Documentation truncated — see the full README on GitHub.