Common questions.

What does Watchpost do?

When your AI agent (Claude, ChatGPT, anything MCP-aware) tries to make a purchase, Watchpost runs three checks: merchant trust, listing manipulation scan, and your rules. It either approves, asks you, or blocks. The whole verdict happens in about a second.

How do I connect my agent?

Run `npx @watchpost/mcp install` once. The CLI walks you through a browser flow to grant access, then writes the right entry into your agent's MCP config (Claude Desktop, for example). Restart the agent and you're done.

What's a 'verdict'?

Watchpost's three-part decision on a purchase. Approve means the agent can proceed. Review means it should pause and wait for your tap. Block means it shouldn't proceed at all.

Why does Watchpost ask for review?

By default, anything above $25 needs your tap. You can change that threshold, or any rule, from the dashboard.

How does the listing scan work?

Mechanical pre-checks run first: zero-width characters, known prompt-injection patterns, recurring-billing keywords, drip pricing. If the listing has substantial text and the pre-checks didn't already flag it, Claude scans the rest looking for AI-targeted manipulation.

Where is my data stored?

Postgres in the EU (Neon). We store listings, verdicts, and your rules. We don't sell or share. Email us and we'll export or delete your account.

What if I miss a notification?

Review verdicts stay open in your dashboard and your agent waits for your tap. If nobody responds within 24 hours the review expires and is treated as a decline (fail-closed), so a missed notification never leaves a purchase hanging. Tap Allow or Block before then and the agent proceeds (or doesn't) the moment you do.

Can I block specific merchants?

Yes. Add the domain to your rules' forbidden list. Future agent calls to that merchant block immediately, with the reason shown.