convez.ai
MCP ServerSegurançaAvançado
TomEageer

MCP Warden

TomEageer/pulse

Gateway de segurança offline para servidores MCP com detecção em tempo real

Ver no GitHub
InicialPluginScore · de 0 a 100

Gateway de segurança e scanner em tempo de execução para servidores MCP que bloqueia envenenamento de ferramentas, rug pulls, injeção de prompt e exfiltração de credenciais. Funciona offline, sem dependências, implementado em Go puro.

Instalação

$npx convez-ai TomEageer/pulse

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
57%
Último commit
há 5 meses
Versão
—
Issues abertas
0
Contribuidores
1

Casos de uso

  • Auditoria de servidores MCP antes da integração em CI/CD
  • Detecção de rug pulls comparando hash de ferramentas aprovadas
  • Bloqueio de exfiltração de credenciais (AWS, GitHub, OpenAI, JWT)
  • Proteção contra injeção de prompt em resultados de ferramentas

Veredito

Ferramenta essencial para quem opera servidores MCP em produção. Aborda um gap real de segurança com abordagem pragmática (offline, zero-dependency), mas reconhece honestamente suas limitações — detecção baseada em padrões, não contra obfuscação avançada. Recomendado como defesa em profundidade.

README

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

mcp-warden

A security gateway and scanner for Model Context Protocol (MCP) servers. Runs fully offline — no LLM, no telemetry, no data leaves the process.

⚠️ Early development. APIs and detection rules will change.

Why this exists

MCP adoption outran its security tooling. Through 2026 researchers disclosed 40+ CVEs against MCP implementations, and scans keep finding that a large fraction of exposed MCP servers ship with no authentication at all. The common attacks are well known: tool poisoning (hidden instructions in a tool's metadata), rug pulls (a trusted tool silently changing its definition after approval), prompt injection delivered through tool results, and credential exfiltration.

The existing open-source tools (mcp-scan, Cisco's scanner, mcp-audit, …) are static — they inspect configs and metadata before/at install time. They do not see what actually flows at call time. The gateways (MCPX, ContextForge, …) focus on routing, auth, and governance, not threat enforcement.

mcp-warden fills the runtime gap: it sits inline between an MCP client and server and inspects live JSON-RPC traffic, blocking the message when a critical rule fires.

What it checks

ClassWhenRule
Tool poisoningtools/listInstruction-like phrases or invisible-unicode in tool metadata
Rug pulltools/listA tool's hash changed since you pinned it (warden pin)
Prompt injectiontools/call resultInjection phrases / invisible unicode in returned content
Data exfiltrationtools/call args and resultsAWS / GitHub / Slack / OpenAI / Google keys, private keys, JWTs

Detection is deterministic and dependency-free (Go standard library only) — appropriate for a component in the trust path.

Install / build

go build -o warden ./cmd/warden   # Go 1.25+

Use

Audit a server (CI-friendly — exits non-zero on a critical finding):

warden scan https://my-mcp-server.example.com/mcp

Pin the approved tool set, then detect later rug pulls:

warden pin  https://my-mcp-server.example.com/mcp --out warden.lock.json
warden scan https://my-mcp-server.example.com/mcp --lock warden.lock.json

Run the inline gateway (warden.json):

{
  "listen": "127.0.0.1:7000",
  "upstream": "https://my-mcp-server.example.com/mcp",
  "lockfile": "warden.lock.json",
  "auditLog": "warden-audit.jsonl",
  "blockOnCritical": true
}
warden proxy --config warden.json
# point your MCP client at http://127.0.0.1:7000 instead of the upstream

When a critical rule fires, Warden returns a JSON-RPC error to the client instead of forwarding the malicious message, and records the decision to the audit log.

Limitations (honest)

  • Streamable HTTP transport only (stdio proxying not yet implemented).
  • Pattern-based detection catches known shapes, not novel obfuscation; treat it as defense-in-depth, not a guarantee. An optional LLM-judge mode is on the roadmap but will remain off by default to preserve the offline guarantee.
  • SSE response bodies are scanned as text (coarse) rather than fully parsed.

Contributions and rule submissions welcome — see CONTRIBUTING.md.

License

MIT — see LICENSE.

Ver o original no GitHub

Bom para

Relacionados