You added your site, the first report came back, you connected Claude Code or Cursor in one command, and then nothing happened. The dot next to your agent is green. The report is sitting there. That is the most common state we see, and it is not a bug: a coding agent does not start work by itself. It waits for you to ask.
Why does a connected agent do nothing?
Answer. Because connecting only gives the agent a door. It walks through it when you send a message that needs the door.
An MCP connection is a set of tools your agent may call. Until a prompt makes one of them useful, the agent has no reason to call any. It does not poll, it does not read your report in the background, and it will not surprise you with changes. This is by design: every change to your site should trace back to something you asked for.
What is the one sentence?
Answer. "use moonleap: get_overview for <your site> and tell me what to do first, with the evidence."
Say it in the terminal where the agent runs, with your project's name. The agent calls one tool and gets the whole picture in one reply: every open finding with the pages and values it was computed from, the recipe for each fix, your Search Console and Analytics numbers if you connected Google, whether ChatGPT and Gemini name your site when asked the questions your customers ask, the security surface, and a short list of what to do first. It reads all of that before it says anything to you.
What happens after I say it?
Answer. The agent checks each finding against your code, shows you a plan, and changes nothing until you approve.
The reply you get is a plan, not a diff. For each item the agent says what it verified in the repository, what it would change, and in which files. You approve some, all, or none. When it ships a fix it reports back to Moonleap with a note; that note becomes a line in your dated ledger, and the next measurement tells both of you whether the fix worked. Findings the agent disagrees with come back as disputed, with its evidence, and stay closed. Findings you decline stay closed for ninety days.
How do I know the loop closed?
Answer. The "hand this report to your agent" card on your Overview disappears, and the agent's card starts counting.
On the Overview page a highlighted card sits above your to-do list until the first call arrives. After that it is gone and the small "Your agent" card shows what was fetched, what was done, and what is still waiting for a report. If you want the numbers in your pocket instead, the same summary arrives in Telegram when you send /overview in a connected chat.
FAQ
Does the agent need my repository?
No. It reads the public site the way a browser does. The repository work happens on your machine, by your own agent, under your approval; Moonleap never sees the code.
What if I use a different prompt?
Any plain request that mentions Moonleap works; the agent picks the tools. The sentence above is the one that brings everything in a single call, which is why we suggest it first. "check Moonleap for <site> and fix what we agree on" is the second one, for the day-to-day loop.
I said it and the agent says the tool does not exist.
Your session started before the server was added. Coding agents load their tool list when a session starts; open a new session and say it again.
Sources
- Moonleap docs, the tool reference and the working protocol https://moonleap.io/docs
- Moonleap connection state on our own projects, 9 September 2026 (four sites; each one showed "connected, nothing fetched" until the first prompt)