Repository navigation
Approvals and Safety
devlomingo edited this page Nov 16, 2026
·
1 revision
Toolsmith assumes an agent can be manipulated (prompt injection) and that third-party specs are untrusted. Humans approve; agents never can.
| 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 |
- The agent's call returns an error: "needs a human to approve this exact call. Approval id: 3f2a9c1e…".
- You approve it in the dashboard under Approvals, or in your terminal:
toolsmith-mcp approve-call 3f2a9c1e(asks you to typeyes). - 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.
- 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.
-
Your machine. Anyone with shell access as your user can read
.envand 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.
toolsmith-mcp · MIT · Repository · Report an issue · Changelog
Getting started
Using it
Operating it
Links