convez.ai
MCP ServerDados & AnalyticsIntermediário
chrisns

UK Tenders MCP

chrisns/uk-tenders-mcp

Consulte toda contratação pública UK via IA — 677k processos, SQL e busca

Ver no GitHub
InicialPluginScore · de 0 a 100

Servidor MCP que indexa a totalidade de processos de contratação pública do Reino Unido (~677k) em BigQuery, permitindo consultas SQL e buscas por AI assistants. Integra dados de cinco portais oficiais (Find a Tender, Contracts Finder, Public Contracts Scotland, Sell2Wales, eTendersNI) seguindo o padrão Open Contracting (OCDS), com URL oficial em cada resultado.

Instalação

$npx convez-ai chrisns/uk-tenders-mcp

1 comando no terminal: detecta o formato do repositório (marketplace ou plugin avulso) e instala via Claude Code. Requer Node 18+ e o Claude Code instalado.

Métricas

Estrelas
2
Forks
1
Acompanhando
0
Saúde
71%
Último commit
há 2 dias
Versão
—
Issues abertas
0
Contribuidores
3

Casos de uso

  • Identificar frameworks existentes antes de novo processo licitatório
  • Benchmarking de preços pagos por fornecedores (Microsoft, AWS, etc) em governo
  • Análise de contratações por órgão, fornecedor, CPV, valor e período
  • Monitorar mudanças em editais (prazos, valores, cancelamentos) em tempo real

Veredito

Ferramenta altamente especializada e bem construída para quem trabalha com licitações públicas UK. A indexação em BigQuery + MCP é um padrão excelente; requer familiaridade com procurement, Open Contracting e SQL para máximo valor. Livre de autenticação e com endpoint HTTP pronto para produção.

README

A tradução para o português deste README ainda está na fila.

UK Tenders MCP

Query the UK public-procurement corpus — every tender, award and amendment — directly from your AI assistant, over MCP. Built on the official Open Contracting (OCDS) data, analytics-first, with the official notice URL on every result. ~677k procurement processes across all five UK portals.

Live MCP endpoint: https://tenders.run.cns.me/mcp

Data is an attributed mirror of source procurement portals under the Open Government Licence v3.0. It is not the authority of record — always verify critical details on the official notice.

Use it

Claude Code

claude mcp add --transport http uk-tenders https://tenders.run.cns.me/mcp

Headless (one-shot)

claude -p "Use the uk-tenders MCP server: which organisations bought from Amazon, by value?" \
  --mcp-config .mcp.json \
  --allowedTools "mcp__uk-tenders__query_sql,mcp__uk-tenders__search_tenders,mcp__uk-tenders__top_suppliers,mcp__uk-tenders__awarded_value_by_buyer"

(Allow-list the read-only tools rather than --dangerously-skip-permissions.)

Any MCP client

Point it at https://tenders.run.cns.me/mcp (Streamable HTTP). No authentication.

Tools

ToolWhat it does
search_tendersFilter by keyword, CPV, buyer, supplier, value, status, stage, regime, region, source, dates. Official URL on every result.
get_tenderOne process by OCID or notice id — compiled state + change timeline.
list_updatesWhat changed since a timestamp (deadline/value changes, cancellations, new awards).
aggregate_tendersFlexible analytics: count / sum / avg / min / max / median of awarded value, grouped by buyer, supplier, CPV, stage, status, regime, region, or time.
top_suppliers, awarded_value_by_buyer, awards_over_time, cpv_breakdownNamed shortcuts over the aggregate engine.
query_sqlRead-only BigQuery SQL over the public dataset (byte-capped; verbatim source data — may include published contact details). The power tool.
get_schemaTables/columns for query_sql.
get_statusPer-source freshness, coverage and health.

Every result carries the official notice URL; every response carries a per-source freshness signal.

Example questions

Reuse opportunities — call off, don't re-procure

The highest-value use for a civil servant: before launching a fresh procurement, see what peers already bought — to call off an existing framework, benchmark prices, or shortlist incumbents.

  • "Is there a framework I can call off for cloud/IT instead of running my own?" → Crown Commercial Service dominates CPV‑72 (IT) framework awards by a wide margin — the obvious call‑off route (G‑Cloud, Technology Services). cpv_breakdown / top_suppliers.
  • "What have peers paid Microsoft / AWS, to benchmark before I negotiate?" → 337 itemised award lines name a Microsoft supplier; 223 name Amazon (~£2.49bn of award ceilings, top buyer Home Office ~£784m). Use the per-notice values, not the headline sum (see Caveats).
  • "Who already supplies cyber-security / consultancy across government, and at what scale?" → top_suppliers + a keyword/CPV filter surfaces incumbents and the frameworks they sit on.
  • "Show me recent awards similar to what I'm about to buy" → search_tenders by keyword/CPV; each hit links to the notice (and thus its framework / call-off terms).

Spend & market analysis

  • "Which organisations bought from Amazon, and how much?" → 223 award lines, top buyer Home Office (~£784m of ceiling). "Top suppliers to the NHS", "how has cloud spend trended?" → top_suppliers, awards_over_time.

Devolved / regional

  • "Biggest IT contracts in Wales this year" → e.g. Welsh Government, Hwb Educational Platform for Wales (~£76m). Filter region='UKL', cpv_division='72'. ('UKN' = NI.)

Pipeline & live opportunities

  • "Active tenders closing in the next 60 days" → ~5,000+ live notices; sort by value, each with deadline + URL. Cross-source — e.g. a live cyber SOC tender from Cardiff & Vale Health Board (Sell2Wales) sits alongside MoD (FTS) and CCS (Contracts Finder).
  • "What changed this week — cancellations or new awards?" → list_updates.

Caveats (the model is told these)

  • Awarded value = contract/framework ceiling, not actual spend. Several notices carry placeholder ceilings (e.g. £9,999,999,999) or whole-framework maxima that massively inflate naive SUMs — round £10bn-per-supplier totals are a tell. Every result links to the official notice; verify critical figures there. Prefer per-notice values over aggregates.
  • Names aren't normalised at source ("Ministry of Justice" vs "Ministry of Justice."). Cross-source dedup groups processes (process_group, ~24k cross-source/duplicate clusters linked), not entity names.
  • Freshness varies by source — get_status reports it.

Architecture

Source portals ──Python adapters──► GCS raw archive ──► BigQuery uk_tenders_raw (write: full event log)
                                                                  │ compile (ocds-merge), verbatim
   AI assistant ──MCP──► Cloud Run (TS API, read-only SA) ──► BigQuery uk_tenders_public (read-only, query-serving)
                                                                  ▲
                              Cloud Run Job (Python) ◄── nightly Cloud Scheduler
  • Read path: MCP client → Cloud Run (TypeScript) → BigQuery public dataset.
  • Write path: nightly Cloud Run Job (Python) replays each source → raw event log → compiled, analytics-ready tables.
  • Faithful mirror: the public dataset reproduces the source OCDS verbatim — the index is not the authority of record and does not alter or remove content; every record carries the official notice URL and OGL v3.0 attribution.
  • Data isolation: the full raw event log and GCS archive stay in an access-controlled write tier the API service account cannot read — a least-privilege read boundary, so a compromised public API can only ever touch the query-serving dataset.

See docs/PRD.md, the ADRs, and the glossary in CONTEXT.md.

Data sources

SourceRecordsStatus / mechanism
Find a Tender (FTS)~109.6kLive — OCDS 1.1.x (ocds-h6vhtk), windowed updatedFrom replay
Contracts Finder~546.6kLive — OCDS (ocds-b5fd17) + bulk daily harvester, 2016→
Public Contracts Scotland~8.3kLive — Proactis API (ocds-r6ebe6)
Sell2Wales~2.1kPartial — HTML scrape, 2013→ (ocds-kuma6s). Its entire OCDS API (collection, single-notice, bulk) is broken upstream — nvarchar→float SQL error — so coverage is limited to what the public archived search exposes (~120/yr). Above-threshold Welsh notices are captured in full via FTS.
eTendersNI~10.7kSnapshot — HTML scrape, CAPTCHA solved by Gemini; refreshed manually from macOS (its WAF blocks Linux/cloud — see DEPLOYMENT.md)

All five portals are ingested into one canonical OCDS model with cross-source deduplication (process_group, blocking + fuzzy buyer/title/value matching). The nightly Cloud Run Job refreshes the four non-CAPTCHA sources + dedup at 02:30 Europe/London; eTendersNI is a periodic macOS refresh (scripts/refresh_etendersni.sh).

Develop

# ingestion (Python)
python3 -m venv .venv && .venv/bin/pip install -e 'ingestion[dev]'
PYTHONPATH=ingestion/src .venv/bin/python -m pytest ingestion/tests -q
# validate against live FTS without touching the cloud:
PYTHONPATH=ingestion/src .venv/bin/python -m uk_tenders_ingest --source fts --mode backfill \
  --from 2026-05-01 --to 2026-05-31 --dry-run --max-releases 100

# API (TypeScript)
cd api && npm install && npm test && npm run build
GCP_PROJECT=govreposcrape node dist/index.js   # local server against BigQuery (needs ADC)
node test/smoke.mjs http://localhost:8080/mcp

Deploy

See DEPLOYMENT.md. In short: ./scripts/deploy_api.sh (API) and ./scripts/deploy_ingest.sh (nightly job), or terraform/ for the full IaC.

License

Code: MIT. Data: Open Government Licence v3.0 (attribution preserved).

Ver o original no GitHub

Bom para

Relacionados