Run opening tests with your own AI assistant

Let Claude, ChatGPT, Cursor or your own code decide what to test on which page, write the openings and read the results, while Darwin shows each visitor one opening and scores them all on visitors who reach booking. You need: an assistant connected to Darwin with Full access (see MCP server), or a key with experiments:write and agent:write.

Leave new tests to your assistant

Ask your assistant to “stop Darwin starting tests on its own” (set_test_automation), or switch off Darwin starts new tests in Experiments → Settings. Keep Run tests automatically on: it is what scores the tests. With the API: PATCH /v1/agent with {"automation": {"auto_launch": false}}.

Look before you test

“What does our pricing page say first?” uses get_page_opening, which answers with the opening, whether a test is already running there, and whether it is the page’s own opening or the site’s. GET /v1/analytics has each page’s visitors and meetings, for choosing where to test.

Start a test

“Test an opening on /pricing that asks for their team size first.” The assistant writes the openings and calls create_experiment. With the API:

curl https://api.meetdarwin.ai/v1/experiments \
  -H "Authorization: Bearer $DARWIN_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "name": "Pricing: ask the team size",
    "page": "/pricing",
    "openings": [
      { "headline": "How many people would use it? I will show you the plan that fits.",
        "buttons": [{ "label": "Under 10" }, { "label": "10 to 50" }, { "label": "Over 50" }] }
    ]
  }'

Darwin adds the opening the page shows now as the current one, so your openings are compared with what visitors see today. Test a whole page type with "unit": "type:pricing", a page group with "unit": "group:…", or every page on the site’s opening with "unit": "site". One test per page at a time.

Read the results

“How is the pricing test doing?” uses get_experiment: each opening’s views, conversations, visitors who reached booking, meetings, chance to beat the current opening and share of visitors, a sentence on where the test stands, and what a new opening still needs before it can be applied. GET /v1/experiments/exp_… returns the same under live. Reading results saves nothing.

Change, apply or stop

edit_opening and remove_opening change a running test; new words start as a fresh opening. apply_winner makes the winner the page’s opening once it has the evidence Darwin requires; stop_experiment ends a test without one. To change a page’s opening without a test, use set_page_opening (a page, or a section such as /blog/*), which is refused while a test runs on that page.

Tests your assistant starts appear in Experiments like any other, marked as written through the API or by an AI assistant, and the experiment.started and experiment.winner_found webhooks fire for them. It worked if get_page_opening for the page shows source: test with your test’s id, and the test appears in Experiments.

Get started

Guides

Run opening tests with your own AI assistant