Ratchet MCP
gorrie/ratchet-mcp
Dataset citado sobre redes institucionais e pessoal recorrente na administração US
Ver no GitHubServidor MCP que expõe um dataset curado de 401 pessoas e 351 instituições na grade de controle regulatório/financeiro/legal dos EUA, com dupla citação em fontes primárias. Permite consultar padrões de pessoal recorrentes e conexões entre atores políticos, financeiros e de segurança. Inclui 9 ferramentas de query e visualizador web D3 opcional.
Instalação
npx convez-ai gorrie/ratchet-mcp1 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
Casos de uso
- Rastrear padrões de rotação de pessoal entre administrações
- Mapear conexões entre atores em governo, finanças e segurança
- Pesquisar sobreposição institucional entre pessoas e entidades
- Análise de política pública e captura regulatória
Veredito

Ferramenta sólida de pesquisa institucional com rigor editorial raro: cada registro exige 2+ fontes primárias e passa por CI. Ideal para jornalistas, pesquisadores e analistas políticos que precisam de dados verificados sobre redes de poder. Não é para iniciantes — requer mentalidade investigativa e disposição para lermos a documentação de escopo.
README
A tradução para o português deste README ainda está na fila.
Ratchet MCP
Open-source Model Context Protocol server for a curated, source-cited dataset of named persons and institutions across the US legal / regulatory / financial / multilateral control grid — tagged with recurring staffing plays and control-grid mechanisms.
Companion to The Ratchet: How Safety Infrastructure Became the Control Grid (Evil Robots Series Book 2, 4LULZ). The book argues that a small recurring cohort runs identifiable, repeating personnel patterns across US administrations. This MCP server makes those patterns directly queryable.
What's in it
- 401 named persons — every record carries ≥2 primary sources and tags from the closed plays and actors vocabularies.
- 351 institutions — every record carries ≥1 source.
- 844 edges — person → institution adjacencies.
- MCP server in Python with 9 tools:
query_cohort,get_entity,who_connects,find_overlap,list_plays_for,list_players_for,find_in_administration,enrich_from_littlesis,littlesis_relationships. - Optional web viewer — D3 force-directed graph at
http://localhost:8088.
Quickstart
git clone https://github.com/gorrie/ratchet-mcp
cd ratchet-mcp
docker compose up --build
# Web viewer (optional):
docker compose --profile web up --build
Wire to Claude Desktop by adding to claude_desktop_config.json:
{
"mcpServers": {
"ratchet": {
"command": "docker",
"args": ["compose", "-f", "/absolute/path/to/ratchet-mcp/docker-compose.yml",
"run", "--rm", "-T", "mcp"]
}
}
}
Three example prompts to demo the surface — see docs/PROMPTS.md for the full set:
- Trump II Vault: "Who ran the Vault play under the Trump II administration?" → 4-5 named Treasury revolvers with sourced positions.
- Anthropic → State: "What connects Anthropic to the State Department?" → multi-hop path through senior policy staff with documented cross-affiliations.
- Two Cranks Same Hand: "Find people who touched both
tapandbackdooractors." → Hayden, Cheney, Bolton.
Scope and discipline
This is an open dataset held to publication standards:
- Every record has ≥2 primary sources. No record makes it past CI with Wikipedia as its only source. See docs/CITATIONS.md.
- No motive imputation, no characterizations, no private-life
references, no unproven allegations. The
rolefield is positions only. See the anti-defamation policy in docs/CITATIONS.md. - Foundations are in, bloodline conspiracy framing is out. This is a dataset about documented institutional positions. See docs/SCOPE.md.
To propose additions or corrections, read docs/CONTRIBUTING.md first. PRs that don't meet the citation and tag-justification standards get blocked at CI.
Licensing
MIT for everything — code and dataset. See LICENSE. Use it, fork it, ship it. Citations within each dataset record point at the underlying primary sources independently of how this repo is licensed.
LittleSis (linked from enrich_from_littlesis) is a runtime API, not
an upstream we copy from. Records in server/data/ are independently
sourced from gov records / Wikidata / official org bios / academic
papers / archived press releases — LittleSis appears as one citation
among several, not as the dataset's origin.
Layout
ratchet-mcp/
├── docs/ # SCOPE, PLAYS, ACTORS, PROMPTS,
│ # CITATIONS, CONTRIBUTING
├── server/
│ ├── ratchet_mcp/ # Python package
│ │ ├── data.py # JSONL loader / Graph
│ │ ├── queries.py # Query primitives
│ │ ├── littlesis.py # LittleSis API client
│ │ └── server.py # MCP tool surface
│ ├── data/ # JSONL data (single source of truth)
│ ├── tests/ # pytest suite
│ ├── pyproject.toml
│ └── Dockerfile
├── web/ # Optional D3 viewer
│ ├── Dockerfile
│ ├── index.html
│ ├── viewer.js
│ ├── viewer.css
│ └── nginx.conf
├── scripts/
│ └── extract_from_graphjs.js # one-time extraction from upstream graph.js
├── docker-compose.yml
├── LICENSE # MIT (code + data)
└── README.md
Status
v0.1 — initial public release (2026-06-09). Dataset at 401 persons /
351 institutions / 844 edges, covering SCOPE.md cohorts A through F
plus the foreign-influence / Five Eyes / UK-EU / USAID-DRG / Gulf
customer / climate-ESG / health-governance / nation-state comparative
clusters and the historical-ideologues layer. The book The Ratchet
ships shortly; this dataset is the queryable form of its institutional
infrastructure argument.
Longitudinal cadence. Re-sampled periodically. Each cut is tagged
(vX.Y.Z) so peer experimenters can pin a specific version and
reproduce; the time-series of cuts is the bias-drift instrument.
Server tool surface stable. CI gates citation, source-count, and
tag-justification rules (see .github/workflows/ci.yml).