Evaluat is agent-ready.
Evaluat runs real-browser performance tests. The same free website speed test the browser tool uses is available to AI agents over a Model Context Protocol server, and to any client over a small HTTP API. Give it a public URL and it loads the page once in a real browser, then returns LCP and CLS, plus FCP and TTFB, an A to F grade, and a shareable video report. A single cold load does not produce a representative INP.
- One honest tool, run_website_speed_test
- Real browser from the UK or the EU
- No API key. Open and rate limited.
Connect an AI agent.
The MCP endpoint speaks JSON-RPC 2.0 over Streamable HTTP. Point your MCP client at it and the
run_website_speed_test
tool appears. It takes a url,
an optional region
(london or frankfurt), and an optional unlisted
flag. Every report is public and permanent and shows the full URL that was tested. Set unlisted to ask search
engines not to index it. It stays public either way and still appears in the test history of that page.
- Endpoint
- https://www.evaluat.com/api/mcp
- Transport
- Streamable HTTP, JSON-RPC 2.0
- Tool
- run_website_speed_test(url, region, unlisted)
Add it to an MCP client config:
{
"mcpServers": {
"evaluat": {
"type": "http",
"url": "https://www.evaluat.com/api/mcp"
}
}
}Clients that only speak stdio can bridge over HTTP:
npx mcp-remote https://www.evaluat.com/api/mcp
It is also listed on
Smithery (opens in a new tab)
as evaluat/pulse, if you would rather add it to your client from there.
The server card is published at /.well-known/mcp.json.
Only run tests against URLs you own or are authorised to test. Use of the MCP endpoint and API is subject to our Terms.
Or call it over HTTP.
The full surface is described in the OpenAPI 3.1 specification. Start a run, then poll for the result.
- POST /api/pulse
- Start a test. Returns a token and the address of the report.
- GET /api/pulse/{token}
- Poll status and metrics until completed.
The browser tool guards POST /api/pulse
with a Cloudflare Turnstile token. Agents should use the MCP endpoint, which performs that step server-side. The
MCP endpoint accepts a JSON-RPC call directly:
curl -s https://www.evaluat.com/api/mcp \
-H 'content-type: application/json' \
-d '{
"jsonrpc": "2.0",
"id": 1,
"method": "tools/call",
"params": {
"name": "run_website_speed_test",
"arguments": { "url": "https://example.com", "region": "london" }
}
}'Everything an agent needs to discover Evaluat.
Agent and API questions
Can an AI agent run an Evaluat speed test?
Yes. Evaluat runs a Model Context Protocol (MCP) server with one tool, run_website_speed_test. An AI agent calls it with a public URL and gets back LCP and CLS, plus FCP and TTFB, an A to F grade, and a shareable video report. A single cold load does not produce a representative INP. Reports are public and permanent, so an agent should only test a URL its user is happy to publish.
How do I connect the MCP server?
Point your MCP client at https://www.evaluat.com/api/mcp over Streamable HTTP. Clients that only speak stdio can bridge with npx mcp-remote. There is no API key.
Does the API need an API key?
No. The MCP endpoint and the website speed test are open and rate limited. Each call runs one real browser, so heavy use is throttled.
What does a speed test return?
Lab metrics from a single real-browser page load: LCP and CLS (Core Web Vitals), plus FCP and TTFB. Because there is no user interaction in a single load, it does not produce a representative INP. You also get an A to F grade (Evaluat's composite, scored against Google's published thresholds) and a shareable report link with a video of the load. The test runs from the UK or the EU.
Need the whole journey, under real load?
The speed test measures one page load. When you need to script a click-through journey and run it at your forecast peak in many concurrent real browsers, with Core Web Vitals, video, and logs per session, that is what the Evaluat platform does.