Bouncer — Verificador Estático de Conformidade Regulatória
nugehs/bouncer
Valida controles regulatórios exigidos no código sem usar IA
Ver no GitHubMCP server que verifica automaticamente se os controles exigidos por regulações (UK Online Safety Act, ICO Children's Code) existem no código-fonte. Funciona como regra determinística sem dependência de LLM, rodando em CI/CD e garantindo conformidade antes do deploy.
Instalação
npx convez-ai nugehs/bouncer1 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
- Verificar conformidade com UK Online Safety Act em aplicações
- Validar implementação de proteções para crianças (AADC)
- Automatizar auditorias de conformidade em pipelines CI/CD
- Gerar relatórios de controles regulatórios para compliance
Veredito

Ferramenta especializada e bem arquitetada para verificação determinística de conformidade regulatória. Preenche um nicho claro que scanners de segurança tradicionais (semgrep, CodeQL) não cobrem — prova que controles obrigatórios *existem*, não apenas que vulnerabilidades não existem. Excelente para times que precisam demonstrar controle regulatório auditável.
README
A tradução para o português deste README ainda está na fila.
bouncer
Static compliance-controls checker — the controls a regulation requires, verified in your code.
nugehs.github.io/bouncer-web (site)

bouncer verifies that the controls a regulation requires actually exist in your code — UK Online Safety Act, ICO Children's Code (AADC) — expressed as deterministic rule packs. It runs in CI, exits non-zero when a required control is missing, and needs no LLM.
It checks IDs at the door so non-compliant code doesn't get in.
bouncer is an engineering aid, not legal advice. A green report means the coded controls a rule looks for were found — it is not a substitute for a compliance / DPO review.
Why
Regulators now expect demonstrable controls: age assurance, high-privacy defaults for children, report/block affordances on user-generated content, a DPIA, a risk assessment. Those are concrete things that either exist in a codebase or don't. bouncer turns a regulation into a set of static checks over your repo, the same way tieline turns an API contract into drift checks — the engine knows nothing about the law; the rule packs do.
bouncer vs semgrep / policy-as-code
Scanners like semgrep, CodeQL, or Snyk answer "is there bad code here?" — they
hunt for vulnerabilities and dangerous patterns that shouldn't exist. bouncer
answers the opposite question: "does the code the regulation requires actually
exist?" — age assurance on sign-up, report/block on UGC surfaces, high-privacy
defaults for children. A repo can be vulnerability-free and still fail every one
of those obligations. Policy-as-code tools (OPA/Rego, Conftest) gate configs and
infrastructure against policy; bouncer gates application source against
regulatory rule packs, with file:line evidence for every control and an honest
unknown when a surface can't be located. In short: semgrep finds
vulnerabilities; bouncer proves required controls exist. They complement each
other — run both.
Install
npx @nugehs/bouncer init
npx @nugehs/bouncer check
Or clone and run with plain Node (zero runtime dependencies, Node ≥ 18).
Usage
bouncer init [path] # write a starter bouncer.config.json
bouncer check # run packs, print report, exit 1 on a missing control
bouncer check --pack uk-aadc # restrict to one pack
bouncer check --status fail # show only the failures
bouncer report --out report.html # self-contained HTML audit report
bouncer list # every rule the configured packs apply
bouncer explain <ruleId> # what a rule requires + how it is checked
bouncer packs # rule packs shipped with bouncer
bouncer doctor # sanity-check config, adapter, packs
bouncer mcp # start the MCP server (stdio)
Verdicts
| Verdict | Meaning |
|---|---|
| pass | the required control was found (evidence: file:line) |
| fail | the surface exists, but no evidence of the control was found |
| unknown | the surface could not be located in this repo — can't determine, not a pass |
unknown is deliberate: bouncer never reports a green pass for a surface it could
not find. Missing surface → honest "can't determine".
Configuration
bouncer.config.json:
{
"target": {
"adapter": "next",
"repo": "./my-app",
"roots": ["app", "src", "components"]
},
"packs": ["uk-osa", "uk-aadc"],
"packDirs": [],
"ignore": [],
"failOn": ["fail"]
}
adapter— how regulation surfaces (sign-up, profile, chat, livestream…) map onto files for your stack.
Adapters shipped today:
next(App Router) andreact-native. That's it — if your stack isn't covered, an adapter is a single small file mapping surface aliases to file globs (seesrc/lib/adapters/next.js). Adapter PRs are very welcome —nuxt,sveltekit,remix,flutter,djangoare all natural candidates.
packs— which rule packs to run. Built-ins:uk-osa,uk-aadc.packDirs— extra directories of your own*.jsonpacks.ignore— rule ids to skip.failOn— which buckets makecheckexit non-zero (default["fail"]).
Rule packs
A pack is JSON. Each rule maps a legal standard to a static assertion over a surface:
{
"id": "aadc.geolocation-default-off",
"standard": "Standard 10 — Geolocation",
"severity": "high",
"surface": "profile",
"intent": "Geolocation must default to off for children.",
"fix": "Default any location-sharing setting to off.",
"assert": {
"find": "(geo|location)[^\\n;,]{0,30}(default|initial)[^\\n;,]{0,15}(false|off)",
"in": ["profile", "any"],
"expect": "present"
}
}
Assertion nodes:
{ "find": "<regex>", "in": "<surface|glob>", "expect": "present|absent" }{ "allOf": [ … ] }·{ "anyOf": [ … ] }·{ "not": … }
in accepts a surface alias (resolved by the adapter), an array of aliases/globs,
or a raw glob. expect: "absent" flips the meaning — a match is a violation
(used for nudge patterns, self-declared age checkboxes, etc.).
Surfaces (next adapter)
any, signup, auth, profile, chat, livestream, ugc, governance.
MCP
bouncer is also an MCP server (stdio), so an agent can pull the same deterministic results:
| Tool | Purpose |
|---|---|
compliance_check | run packs, return per-control verdicts + evidence |
list_rules | list rules the configured packs apply |
explain_rule | a rule's standard, intent, fix, and how it is checked |
list_packs | available rule packs |
// .mcp.json
{ "mcpServers": { "bouncer": { "command": "npx", "args": ["-y", "@nugehs/bouncer", "mcp"] } } }
CI
- run: npx @nugehs/bouncer check
Fails the build when a required control goes missing — e.g. someone removes an age-gate or a report button from a UGC surface.
Tests
npm test # node --test — zero dependencies, nothing to install
The suite runs on Node's built-in test runner against throwaway fixture repos:
glob/brace expansion, every assertion probe (find, allOf/anyOf/not,
allInFile + within windows, expect: "absent"), the pass/fail/unknown
verdict semantics, and pack loading. CI runs it on Node 18, 20, and 22.
License
MIT
Part of the toolchain
bouncer is one of four tools that form a deterministic trust layer for AI-assisted development. Each answers a question people keep handing to an LLM — with static analysis instead.
- repoctx — context: what does this change actually touch?
- tieline — contracts: did the front end and back end quietly stop agreeing?
- bouncer (this tool) — compliance: could you defend this to Ofcom?
- aiglare — governance: where can the model do something you can't undo?
More at segunolumbe.com. static analysis, never the model.