How it works
Scan, build, pre-flight, file, host
Five steps. The first is free and takes two minutes; the middle three take about five business days; the last one runs for as long as you want it to.
Scan Build Pre-flight File Host
1. Scan
You give us a domain, an OpenAPI URL, or an MCP URL. We probe it from the public internet the way a directory reviewer would and write down every reason each of the three would reject it today. No account and no card. Most of the findings are things you could fix yourself, and the report says so plainly, because a report that pretends everything is hard is not worth reading.
2. Build
We write one definition of your business as a set of tools: what each one does, what it takes, whether it only reads or actually changes something, and who has to say yes before it runs. From that definition we serve an MCP endpoint, which is what Claude and ChatGPT speak, and an OpenAPI document, which is what Muse reads. Both describe the same tools, so there is one thing to keep correct instead of three.
Every write takes a key the caller generates. Send the same key twice and you get the first answer back, not a second booking. Reads never wait for anybody.
3. Pre-flight
We drive real assistants against your connector through a proxy that breaks things on purpose: it drops a response so the agent retries, expires a session mid-task, returns an error where the agent expected a result, plants hostile text in a field the agent will read back, and sends two agents at the same slot at once.
Then we check what actually happened in your data, not what the transcript claims: how many rows were written, whether the second attempt created a second booking, whether a write ran before the owner approved it, what status the call ended in. Only one question goes to a model, and that is whether the assistant told the user it had done something it had not. You get the trace and a failing test for anything we find.
4. File
Each directory wants different things and two of them want them in a particular format. We generate the listing text inside each field limit, the test cases ChatGPT wants run on web and on mobile, and the documentation and privacy URLs the reviewer will open. Then we submit, track, and answer whatever comes back. If a directory rejects us, we fix and re-file at no charge within the first 60 days.
5. Host
A live connector needs someone watching it. We run yours, and we put a human in front of anything that commits you.
Fig. 1 — the approval gate: reads run, writes queue, one waits for a humanFig. 1 — the approval gate
When an agent calls a write, the connector does not do it. It records the request, sends you a WhatsApp message with a six-character code, and answers the agent with a pending status and a request ID. You reply YES, NO, or ALWAYS, and ALWAYS remembers that answer for that one agent. The agent polls until there is a real answer: confirmed, denied, expired, or a structured error that says what went wrong.
Alongside that you get attribution, so you can see that Claude users asked for quotes fourteen times this month and three of them booked, uptime monitoring, and re-filing when a directory moves the goalposts.