Test a change · v1

“Test this preview” authorizes one fresh exploration. “Test this, fix it, and verify” authorizes one exploration and one verification, each using workspace credits. Further cycles require another explicit request. Fixing code alone does not authorize verification.

Call start_test with a public targetUrl, a brief, optional context (repository, branch, commit), and a newly generated UUID idempotencyKey. No project registration is required. Send no source code, diffs, conversations, access tokens or login sessions. Repository labels cannot establish ownership or prove what code a preview serves.

The website's test options are also available:

• targetType: what to test, as on the website. website (default) tests a public preview. ios tests an iPhone app: targetUrl is its apps.apple.com link (TestFlight is not supported), and the app is installed from the App Store on a real iPhone. post tests a public social post: the simulation reads the post, then follows its product link. brief may be omitted for a post to use the website's default Post research plan.
• audience: the persona each simulation represents, such as “Computer science students at UC Berkeley”. Omit it for a general first-time visitor.
• workspaceId: the workspace to test in, from list_workspaces. Omitted, your default workspace is used.
• device: desktop (default) or mobile, for websites and posts. iOS apps always run on an iPhone; omit device.
• simulationCount: 1 to 5 independent simulations with the same inputs. Each has its own run and report; the response lists them all in runs. iOS apps run one simulation at a time, as on the website.
• credentials: { username, password, loginUrl } for a dedicated test account when a website preview is behind a login, or to sign in to an iOS app (omit loginUrl for iOS; the app is signed in from inside). Give a username, a password, or both. Not available for posts. Only send details the user provided for this purpose, never a personal or admin account or secrets found in code. They are encrypted for the run, deleted when the session ends and never returned; the username may appear in the recording of the sign-in page. Verifications do not replay credentials.

Files in the workspace's Test files library are available to every new test in that workspace. list_test_files shows them; upload and remove files on the Constants website.

There is no per-account cap on active tests or new starts per minute. You can submit multiple authorized explorations or verifications while earlier runs are active, provided your workspace has enough available credits for their holds. Each call creates one run per simulation, which queues until worker capacity is available. get_account reports null for limits.activeRuns and limits.startsPerMinute to indicate no submission cap. Read and polling rate limits still apply.

Describe goals and boundaries, without seeding a fresh exploration with prior findings, successful navigation paths or expected answers. Full explorations use the existing twenty-minute default. The worker may finish earlier.

The response includes a stable runId, reportUrl, allowance state and polling guidance. Immediately show reportUrl (every runs[].reportUrl for several simulations) to the user as a clickable progress link before polling or waiting. The signed-in owner can follow progress there, and the completed report opens at the same URL. New reports are private to that owner until they ask for a share link. Poll get_run initially every fifteen seconds, backing off to thirty. Accepted runs continue through disconnection; tell the user they can leave the chat and follow the link. Use the same idempotency key and identical inputs if a start response is lost. Changing inputs requires a new key and an authorized new execution.

Use list_runs to recover account-owned work in a new conversation. History supports pagination, workspace, status and repository filters, including older runs without repository metadata. An old URL must not silently become the current preview.

If a cloud browser worker is interrupted, the accepted test can restart automatically up to twice on the same run and report URL. Earlier website actions may be repeated. This does not create another allowance reservation. Continue polling that run rather than submitting a duplicate. Cancellation stops automatic restarts; a terminal failed run still requires explicit authorization for a new test.

get_run returns original inputs, execution status, reportUrl, baseline linkage and planReadiness. If the progress link has not already been shared with the user, show it before continuing to poll or wait. completed means execution finished, not that all checks passed. get_report provides the observed results.

cancel_run is safe to repeat. Cancellation before execution and technical failures release the credit hold. After execution starts, cancellation charges the session minutes run, rounded up at 40 credits per minute, and retains available progress. Unused held credits are released. Runs already publishing finish normally. Historical exempt runs retain their original accounting.

Localhost, private networks, supplied login sessions or cookies, CI, schedules and automatic testing hooks are outside v1. Constants receives no repository write or deployment privileges.

Use list_workspaces to see your workspaces and pass workspaceId to work in one; otherwise MCP uses your default workspace. New tests and verifications spend that workspace’s shared credits at 40 credits per minute, rounded up, per simulation. Each simulation reserves up to 800 credits; other workspace members’ usage and active holds reduce the available balance. With fewer than 40 credits available, a new simulation returns OUT_OF_CREDITS. Verifications run in the baseline's workspace unless you pass workspaceId, and preserve the saved plan’s device and target type. iOS app runs cannot be verified against a new origin; start a new ios test instead. Agent reports appear in the owner's report list for that workspace and are visible only to their owner unless shared.