How to audit your site from an AI assistant

Docket speaks the Model Context Protocol, the standard AI assistants use to reach tools on your own machine. Point an MCP client at the copy of Docket you already own and you can ask it to audit a site in conversation, and get the findings back as structured data it can reason about — rather than a screenshot you retype.

Point your assistant at it

Docket serves MCP over stdio, so there is no server to run and no port to open. The client launches Docket itself. In Claude Desktop, open ~/Library/Application Support/Claude/claude_desktop_config.json and add:

{
  "mcpServers": {
    "docket": {
      "command": "/Applications/Docket.app/Contents/Resources/docket/docket",
      "args": ["mcp"]
    }
  }
}

That path is the audit engine inside the installed app, not the app icon you click. Restart the client and Docket appears as a tool. Any MCP client works the same way — the command and the argument are all it needs.

The two tools

audit_site crawls a site and returns the score and every finding: how bad it is, which pages it affects, how long the fix takes and the change to make. It takes a url, and optionally max_pages, max_seconds and offline.

list_checks takes no arguments and answers “what does Docket actually check?” without crawling anything. On this copy it returns all 97 checks across 13 areas.

Two details worth knowing, because they are choices rather than accidents. The page budget you ask for is the page budget you get — Docket will not quietly crawl five times more because the site answered quickly, which is what its own --adaptive default does when a person is watching. And offline keeps a run entirely local: no PageSpeed call, no DNS lookup, nothing leaving the machine.

It is the same engine, and the same licence

Nothing about this is a cloud service. The crawl runs on your Mac exactly as it does when you click the button, there is no API key, and there is no per-call charge — your assistant is talking to the app you bought, not to us.

Which also means an unactivated copy will not audit. Ask it to and it answers:

This copy of Docket is not activated. Enter your licence key to run an
audit — you can find it in the email from your purchase.

list_checks still works, because it crawls nothing. Activate once with docket licence --key YOUR-KEY, or in the app, and the audit tool works from then on. Docket is $349 once, so there is no seat to add for the assistant.

What it does not do

The MCP surface is deliberately small: two tools, not a wrapper around every command. It does not schedule audits, compare two crawls the way docket diff does, or expose the site-history features. If you want those, they are in the app and the CLI, and the honest answer is that nobody has yet asked for them through an assistant.

Common questions

Does Docket work with Claude Desktop?

Yes. Add Docket to mcpServers in claude_desktop_config.json with the command /Applications/Docket.app/Contents/Resources/docket/docket and the argument mcp, then restart the client. Any MCP client works the same way, because Docket speaks the protocol over stdio.

Does using Docket through MCP cost extra?

No. There is no API key and no per-call charge. The crawl runs on your own Mac, the assistant is talking to the copy of Docket you bought, and the licence is the same one-time purchase.

Can an assistant audit a site if Docket is not activated?

No. Asking it to audit returns 'This copy of Docket is not activated'. Listing the checks still works, because that crawls nothing. The licence gate is the same one the app and the command line use, so MCP is not a way around it.