convez.ai
MCP Server
saqreed

argusgate

saqreed/argusgate

Ver no GitHub
BomPluginScore · de 0 a 100

ArgusGate is an open-source security scanner and policy gateway for Model Context Protocol servers. It helps developers and DevOps/SecOps teams inspect MCP tool metadata, detect suspicious tool descriptions, identify risky capabilities, validate security policies, and generate CI-friendly reports before connecting AI agents to internal systems.

Instalação

$npx convez-ai saqreed/argusgate

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
3
Forks
0
Acompanhando
0
Saúde
85%
Último commit
há 1 mês
Versão
v0.3.1
Issues abertas
6
Contribuidores
2

README

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

ArgusGate

ArgusGate is an open-source security scanner and policy gateway for Model Context Protocol servers.

It helps developers and DevOps/SecOps teams inspect MCP tool metadata, detect suspicious tool descriptions, identify risky capabilities, validate security policies, and generate CI-friendly reports before connecting AI agents to internal systems.

ArgusGate is experimental. It performs heuristic static analysis. It is not a complete security boundary and does not replace sandboxing, least-privilege credentials, network isolation, code review, or runtime monitoring.

Why It Exists

MCP servers can expose tools that read files, run commands, query databases, automate browsers, call APIs, and operate infrastructure. A malicious or compromised server can also hide instructions in tool descriptions or leak secrets through config and metadata.

ArgusGate v0.1.x focuses on a small useful baseline: local offline scans, policy checks, readable terminal output, JSON reports, and CI-ready exit codes.

Install From Source

Requirements:

  • Go 1.24 or newer.
mkdir -p bin
go build -o ./bin/argusgate ./cmd/argusgate

During development:

go run ./cmd/argusgate --help

Quick Start

Validate the example policy:

go run ./cmd/argusgate policy validate --policy examples/policies/default.yaml

Scan the safe fixture:

go run ./cmd/argusgate fixtures scan --path examples/fixtures/safe-tools.yaml --policy examples/policies/default.yaml --report safe-report.json

Scan the malicious fixture:

go run ./cmd/argusgate fixtures scan --path examples/fixtures/malicious-tools.yaml --policy examples/policies/default.yaml --report malicious-report.json

The malicious fixture is expected to exit 1 because it contains high-severity findings.

Emit JSON to stdout:

go run ./cmd/argusgate fixtures scan --path examples/fixtures/safe-tools.yaml --format json

Override the policy fail threshold:

go run ./cmd/argusgate fixtures scan --path examples/fixtures/safe-tools.yaml --policy examples/policies/default.yaml --fail-on medium

Scan a local MCP config:

go run ./cmd/argusgate scan --config examples/configs/mcp-config.yaml --policy examples/policies/default.yaml --report config-report.json

CLI

argusgate --help
argusgate --version
argusgate scan --config <path> [--policy <path>] [--report <path>] [--fail-on high] [--format text|json] [--quiet]
argusgate policy validate --policy <path>
argusgate fixtures scan --path <path> [--policy <path>] [--report <path>] [--fail-on high] [--format text|json] [--quiet]

Exit codes:

  • 0: no findings at or above the configured fail level
  • 1: findings at or above the configured fail level
  • 2: invalid config, invalid policy, parser error, report write error, or internal error

Policy Example

version: "0.1"
project:
  name: "argusgate-example"
defaults:
  fail_on: "high"
  allow_unknown_tools: true
rules:
  deny_tools:
    - "shell_exec"
    - "run_command"
    - "file_write"
  deny_keywords:
    - "ignore previous instructions"
    - "do not tell the user"
    - "private key"
    - "exfiltrate"
  paths:
    deny:
      - "~/.ssh"
      - "/etc"
      - ".env"
      - "kubeconfig"
    allow:
      - "./examples"
      - "./testdata"
servers:
  local-filesystem:
    allow_tools:
      - "read_file"
      - "list_directory"
    deny_tools:
      - "write_file"
      - "delete_file"

Policy precedence is documented in docs/policy-format.md.

JSON Report

Reports include:

  • scanned_at
  • argusgate_version
  • source_type
  • source_path
  • servers
  • tools
  • findings
  • severity_summary
  • policy_summary
  • exit_decision

Example shape:

{
  "scanned_at": "2026-05-22T12:00:00Z",
  "argusgate_version": "0.1.3",
  "source_type": "fixtures",
  "source_path": "examples/fixtures/malicious-tools.yaml",
  "servers": [],
  "tools": [],
  "findings": [],
  "severity_summary": {
    "critical": 0,
    "high": 0,
    "info": 0,
    "low": 0,
    "medium": 0
  },
  "policy_summary": {},
  "exit_decision": {}
}

Finding evidence is redacted when it looks secret-like.

CI Usage

Example GitHub Actions step:

- name: ArgusGate fixture scan
  run: |
    go run ./cmd/argusgate fixtures scan \
      --path examples/fixtures/malicious-tools.yaml \
      --policy examples/policies/default.yaml \
      --report argusgate-report.json

The repository CI runs:

go test ./...
go vet ./...
mkdir -p bin && go build -o ./bin/argusgate ./cmd/argusgate

Current Limitations

  • Static analysis is heuristic and can miss issues.
  • Findings can include false positives.
  • The scanner does not connect to live MCP servers.
  • The scanner does not execute tool commands or invoke MCP tools.
  • No runtime proxy/gateway is implemented in v0.1.x.
  • No database, web UI, OAuth/RBAC, Kubernetes deployment, or SaaS workflow is included.

Roadmap

  • More fixture coverage and detector tuning.
  • Tool pinning and rug-pull detection.
  • Optional live MCP metadata inspection.
  • Runtime MCP gateway/proxy after the scanner and policy engine stabilize.
  • Audit logging and telemetry for future runtime enforcement.

Contributing

See CONTRIBUTING.md. Keep examples safe and local. Never include real credentials in issues, docs, fixtures, tests, or reports.

License

Apache-2.0. See LICENSE.

Ver o original no GitHub

Bom para

Relacionados