convez.ai
MCP ServerSegurançaIntermediário
rogerchappel

MCPSeal

rogerchappel/mcpseal

Linter de segurança para configs MCP sem acesso à rede

Ver no GitHub
BomPluginScore · de 0 a 100

Linter local de configurações MCP que analisa riscos de segurança em definições de servidores de agentes antes da conexão. Detecta comandos perigosos, exposição de variáveis de ambiente, acesso amplo ao filesystem e descrições suspeitas de ferramentas, com saída determinística em Markdown/JSON e redação de secrets habilitada por padrão.

Instalação

$npx convez-ai rogerchappel/mcpseal

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
0
Forks
0
Acompanhando
0
Saúde
100%
Último commit
há 3 semanas
Versão
—
Issues abertas
3
Contribuidores
1

Casos de uso

  • Auditar configurações MCP antes de implantar agentes
  • Validar segurança em CI/CD com gates explícitos (--fail-on)
  • Revisar exposição de credenciais e acesso ao filesystem em servidores MCP
  • Gerar relatórios de risco em Markdown para documentação de segurança

Veredito

Ferramenta essencial para validar segurança de configurações MCP antes de ativar agentes. Abordagem local-first sem telemetria é adequada para CI/CD, mas lembre-se que é linter heurístico, não sandbox—combine com code review e permissões do SO.

README

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

MCPSeal

MCPSeal is a no-network MCP config linter. It explains risky server commands, environment exposure, broad filesystem access, and suspicious tool descriptions before an agent connects to those tools.

It is built for local review and CI evidence: deterministic Markdown/JSON output, redaction on by default, and explicit --fail-on gates.

Why this exists

MCP configs are powerful launch plans. A single server entry can grant an agent shell execution, broad filesystem reads, Docker access, or credential-bearing environment variables. MCPSeal gives developers a fast preflight that is easy to run, diff, and automate without uploading configs anywhere.

Quick start

npm install
npm run build
node dist/src/cli.js scan examples/risky-mcp.json
node dist/src/cli.js check examples/risky-mcp.json --fail-on secret-env,broad-fs

After installation from a package, use the binary directly:

mcpseal scan ~/.config/my-agent/mcp.json --out mcpseal-report.md
mcpseal check ./examples/risky-mcp.json --fail-on secret-env,broad-fs,risky-command

Commands

See docs/CLI.md for the full CLI reference.

scan

Produces Markdown by default:

mcpseal scan <file-or-dir...> --out report.md
mcpseal scan <file-or-dir...> --format json

check

Produces JSON by default and exits with code 2 when an explicit gate fails:

mcpseal check <file-or-dir...> --fail-on secret-env,broad-fs

Valid --fail-on categories:

  • risky-command
  • secret-env
  • broad-fs
  • tool-description
  • config-shape

What MCPSeal checks

See docs/RULES.md for the detailed rule catalog.

  • Risky commands: shells/interpreters, package runners, network/destructive command tokens.
  • Secret env: keys or values that look like tokens, passwords, private keys, or cloud credentials.
  • Broad filesystem: root/home/tmp-style access, broad mount flags, Docker socket references.
  • Tool descriptions: missing/weak descriptions or prompt-risk language.
  • Config shape: invalid JSON, missing commands, or files without recognizable server definitions.

Redaction

Redaction is enabled by default. Secret-looking environment values and common token formats are replaced with [REDACTED] in output.

Use --no-redact only for local debugging when you are sure the output will not be shared:

mcpseal scan examples/risky-mcp.json --no-redact

Safety model

  • No telemetry.
  • No external network calls in scan/check.
  • No hidden state or SaaS account.
  • No input mutation.
  • Deterministic timestamp (1970-01-01T00:00:00.000Z) for stable diffs.
  • Non-zero policy gates only happen when requested through --fail-on.

Examples

Checked-in fixtures live under examples/:

node dist/src/cli.js scan examples/risky-mcp.json --out .tmp/risky.md
node dist/src/cli.js check examples/safer-mcp.json --fail-on secret-env,broad-fs,risky-command

Development

npm test
npm run check
npm run build
npm run smoke
bash scripts/validate.sh

Limitations

MCPSeal is a linter, not a sandbox. It uses deterministic heuristics and cannot prove a server is safe. Treat findings as review prompts, keep server packages pinned, and combine this with normal code review and OS-level permission boundaries.

Package contents

The npm package allowlist includes the runtime files plus the public support documents needed for release review: README.md, LICENSE, SECURITY.md, CHANGELOG.md, CONTRIBUTING.md, CODE_OF_CONDUCT.md. Run npm run package:smoke or npm pack --dry-run before publishing to confirm those files are still present in the tarball.

Verification

Run the release-readiness checks that match this package before publishing or opening a release PR.

  • npm run release:check - run the full release gate
Ver o original no GitHub

Bom para

Relacionados