Skip to content

Approvals and Safety

devlomingo edited this page Nov 16, 2026 · 1 revision

Approvals and safety

Toolsmith assumes an agent can be manipulated (prompt injection) and that third-party specs are untrusted. Humans approve; agents never can.

What needs a human

Action Approval
A newly imported API version Approve before any tool is visible
Write operations (POST/PUT/PATCH) Hidden until approved with --allow-writes
Destructive operations (DELETE) Hidden until --allow-destructive, and every call needs its own approval
A composite tool Always approved; one that writes needs its own write grant
An upstream MCP server's destructive or unannotated tools Per-call approval, unless you opt out per server or per tool

Approving a destructive call

  1. The agent's call returns an error: "needs a human to approve this exact call. Approval id: 3f2a9c1e…".
  2. You approve it in the dashboard under Approvals, or in your terminal: toolsmith-mcp approve-call 3f2a9c1e (asks you to type yes).
  3. The agent retries the same call.

An approval works once, for those exact arguments, and only for the client that asked. Optionally ("approvals": { "inClient": true }), clients that support elicitation can show the prompt themselves.

Built-in protections

  • SSRF guard: requests to private, loopback and cloud-metadata addresses are refused unless allowed per integration; redirects are never followed.
  • Secrets: read from the environment when needed, never stored, scrubbed from every tool result, trace, log and error.
  • Privacy: tool results are never stored, only keyed fingerprints; arguments are stored redacted, or not at all.
  • Tool poisoning: invisible characters are stripped from everything an agent reads, and instruction-like text in a spec fails evaluation.
  • No generated code: APIs and composites are data run by fixed code.
  • Per-client allow-lists enforced on every call, not just on listing.
  • Tokens: HTTP serving always needs a client token; OAuth tokens are verified for signature, issuer, audience, expiry and scope, and never forwarded upstream.

What it does not protect

  • Your machine. Anyone with shell access as your user can read .env and the database. An agent with unrestricted shell access is not stopped by the typed confirmation; sandbox it if that matters.
  • The agent's judgment within what you granted. Use API credentials with the least privilege needed.
  • Prompt injection in API responses. Response content is passed through unchanged.
  • Upstream MCP servers are trusted by configuration and not sandboxed.

Full threat model: SECURITY.md. To report a vulnerability, see the security policy.

Clone this wiki locally