Tool safety

Grok Bot API tools safety checklist.

Before Grok can call your API through MCP, decide exactly which tools are safe, how auth is protected, and what should remain blocked. Sourced generates this report automatically from OpenAPI.

01Details

Allowlist first

01

Start with non-sensitive read operations only.

02

Do not treat an empty allowlist as safe; for Grok, omitted or empty allowed_tools can mean broader exposure than intended.

03

Name tools from stable OpenAPI operationId values.

04

Keep internal/admin endpoints excluded unless the owner explicitly approves them.

02Details

Auth boundary

01

Keep API secrets in the MCP server environment or protected connector headers.

02

Never accept an API key as a tool parameter that the model can see in conversation.

03

Require HTTPS for authenticated remote MCP servers.

04

Block OAuth/OpenID Connect until callback URLs, token storage, refresh, scopes, and revocation are reviewed.

03Details

Operation risk

01

Sensitive reads include users, API keys, billing, tokens, secrets, and admin inventory.

02

Write operations create or update state and should be reviewed one by one.

03

Destructive operations delete, refund, cancel, rotate, revoke, transfer, or purge data and should be excluded by default.

04

Rate limits and idempotency behavior should be documented before enabling mutating tools.

04Outputs

Sourced-generated proof

01

agent-readiness.md summarizes the safety counts and blocking issues.

02

adapter.ir.json stores normalized operations, auth, schemas, and safety labels.

03

harness.tests/safety-assertions.mjs checks the manifest and default-enabled tools.

04

PR_PLAN.md records the exact files and human approval gates before any customer-owned repo write.