MCP, the Model Context Protocol, is how an AI assistant reaches a tool it does not have built in. ReviewComet answers it, so an assistant running on your own machine can work your reviews the way you would: it asks the same account, over the same API, using a key you made and can take back.
What you need: you have to own the account or be an administrator on it, your plan has to include it, and you need one API key. When the plan does not include it, the card says so and offers a link to billing instead of an address.
Which assistants
Claude Code, Claude Desktop and Cursor. Claude Code takes a single command in a terminal. The other two read a small block of JSON in their own configuration file. Anything else that speaks MCP over HTTP works the same way, with the same address and the same authorization header.
Connect one
- Open Settings > Developers. In the agency panel it is the Developers page.
- On the API keys card, press Create key and tick only the scopes the assistant needs. The key is shown once, so copy it before you close the dialog.
- Underneath, find the Connect an AI assistant card. It carries the Server address for your account, which is the only address you need.
- For Claude Code, copy the line under Claude Code and run it in a terminal.
- For Claude Desktop or Cursor, copy the block under Claude Desktop and Cursor into that product's own configuration file.
- In whichever you pasted, replace the placeholder with the key from step 2, then restart the assistant.
The Connect an AI assistant card showing the server address and the two copy-paste setups.
Ask it to list your most recent reviews. If it answers with real ones, it is connected.
What an assistant can and cannot do
Everything is bounded by the scopes on the key you gave it. Within those, an assistant can list reviews, reply, assign, read your rules, run a rule backtest, read your insights and read your credits. An agency key adds the client tools on top. It cannot connect a source, change your billing, or touch a location that key is not allowed to name.
It is the same API underneath, so it costs the same. Reading is free however often it happens, and every action counts as 1 credit, exactly as it would if you had pressed the button yourself. Every action also writes an activity row marked via MCP, so Activity shows which changes came from an assistant and which came from a person.
Keep the key safe
A key is a credential, not a setting. Anyone holding it can do everything its scopes allow without signing in, so keep it in the assistant's configuration on your own machine and out of shared documents and code repositories.
If a machine is lost, if somebody who had it leaves, or if you stop using the assistant, press Revoke on that key's row in the API keys card and confirm. It stops working on the next call and the assistant loses access immediately. Make a new key rather than reusing an old one somewhere else.
The same card on a plan that does not include it, showing a notice and a link to billing.