Audit a website from an AI agent over MCP
Updated 2026-11-17
An agent is good at turning a list of problems into tickets, pull requests or a client email. It is bad at crawling politely and remembering which check found what. This guide splits the work along that line: the API does the crawl and ranking, the agent does the writing.
1. Connect the server
Add the MCP endpoint /mcp to your client with your key as a Bearer token. The catalog tools (list_endpoints, search_endpoints, get_endpoint, quote_endpoint) work without a key, so an agent can discover prices before anyone pays. Exact client settings are on the MCP page.
No key yet? The agent can create a capped trial key itself with POST /v1/agent/signup; see agent signup.
2. Quote before running
Ask the agent to call quote_endpoint for seo.audit with the domain. The quote validates the input, returns the price and says whether the wallet can pay. Nothing is charged. This catches typos in the domain before they cost money.
3. Run the audit
Then run_endpoint with the same input. The call returns the run envelope: cost_usd, source and the report in output. A full audit is large; when you only need the plan, ask for only those parts with the fields response control (for example fields=summary,actions,plan) to keep the agent's context small (see response controls).
4. What to ask the agent to do with it
- Tickets: one ticket per item in
plan.now, with the evidence URLs pasted in and a link to the matching page in the check reference. - Client summary: the diagnosis paragraph, the top three actions and the
not_coveredlist, so the client knows what the report does not measure. - Code fixes: for sites you build, many fixes (missing titles, canonicals, viewport, lang, alt text) are template changes an agent can propose directly.
Limits worth telling the agent
The crawler reads served HTML and does not run JavaScript, so single-page apps can show false "thin content" or "missing H1" findings. Topic gaps come from competitor sitemaps and are candidates, not search-volume data. Both caveats are in the report itself; instruct the agent to repeat them rather than smooth them over.
Cost
One audit is one fixed price per report, shown on the endpoint page. Failed calls (unreachable site, invalid input) cost nothing and say billed: false.
Run the full audit on your site. One call checks every item in the check reference and ranks the fixes.
Open the playground Endpoint details