GHOSBC Safety Gate

Use before depending on an unfamiliar A2A agent. Reads only the public Agent Card and optional public registry metadata, checks declared bindings/protocol versions, skill descriptions, security declarations and registry task-verification evidence, then returns a trust/readiness score plus shareable badge metadata. It does not send a task to or execute any skill on the target agent.

Use for a public remote MCP server that an autonomous agent depends on. Remembers the tools/list baseline, returns only added/removed/modified tools, screens new or changed tool definitions, and produces one aggregate ALLOW/REVIEW/BLOCK decision. It does not inspect server source code or authenticated/private MCP endpoints.

Use to inspect a public Shopify store's AI-shopping surface before an autonomous agent trusts it. Reads only public /.well-known/ucp, /agents.md and /api/ucp/mcp metadata, runs initialize/tools/list, screens exposed tool definitions, and returns a readiness/risk packet. It never creates or mutates carts, checkouts, orders, payments, credentials, or store data.

Use only when free usage is exhausted or production volume is needed. Returns Stripe checkout for 10,000 prepaid GHOSBC Safety Gate checks for $19; it does not charge or receive payment credentials.

Use before sending context to an external model, tool or agent when the payload may contain credentials or private material. Redacts common secret patterns and flags policy-extraction language. Not a complete DLP/compliance system.

Use before an agent follows untrusted instructions or requests capabilities. Returns ALLOW, REVIEW, or BLOCK plus an audit digest. Best for prompt/policy routing; use screen_consequential_action for a concrete purchase, write, deployment, deletion or other bounded action.

Use immediately before a consequential agent action such as purchase, payment, transfer, send, deploy, publish, execute, cart mutation, checkout mutation, or delete. Compares the proposed action with caller-declared allowed actions/targets, amount ceiling, currency and expiry. ALLOW only when explicit bounds fit; REVIEW when bounds are missing; BLOCK when limits are violated. Does not execute the action.

Use before exposing a third-party MCP tool to an autonomous agent, or after a tool definition changed. Screens name, description, schema and annotations for injection-like language, credential/private-context surfaces, side-effect risk and weak contracts. Advisory metadata gate, not source-code verification.

Use immediately before an agent delivers a response outside its trust boundary. Flags likely credential leakage or policy-extraction content and returns a sanitized response when review is needed. Not factuality verification.