Docs

Watchpost does one thing: it reads an intended purchase and returns approve / review / block before the money moves. There are four ways to get that verdict in front of your agent. They all reach the same checks — merchant trust, listing risk, your rules — so pick by where your agent actually runs.

Connector install

For desktop MCP hosts — Claude Desktop and Cursor. Run npx @watchpost/mcp install, confirm in the browser tab it opens, restart the agent: the CLI writes the MCP config for you. Install guide.

API reference

For anyone building their own agent or backend — one HTTP endpoint, one connection token, plus native ACP and AP2 adapters if your checkout is already protocol-authorized. API reference.

Skill

For OpenClaw, Hermes, and any other Agent Skills runtime — a Markdown instruction file and a helper script your agent loads, no MCP host involved. How to add it.

Browser extension

For the ungoverned path — an agent (or you) clicking through an ordinary web checkout where no payment protocol is involved. It watches for the place-order click and gates it. Beta, side-load for now. Details.

Not sure which

If your agent runs in a desktop app you did not write, start with the connector. If you wrote the agent, go straight to the API. The skill covers runtimes that load Markdown skills instead of speaking MCP, and the extension covers the case where there is no agent API at all — just a checkout page.