Product and app testing by a real human
Automated tests tell you the code works; they can't tell you a first-time user gets lost on step two. One real human walks through your product on a real device, narrates what they expected versus what happened, and hands back annotated screenshots or video your team — or your agent — can act on.
When your workflow needs this
First-run & onboarding walkthroughs
A genuinely fresh pair of eyes goes from landing page to activated user, flagging every point of confusion on the way.
Bug reproduction on a real device
A report says it breaks on mobile — a human reproduces it, documents exact steps, and captures the evidence.
Usability & accessibility passes
Keyboard navigation, contrast, screen-reader sanity, unclear labels: the accessibility_and_usability_check service is the same flow, one field away.
How to submit it
One JSON request — no auth, no SDK. Or connect over MCP and the human becomes a tool: claude mcp add --transport http human-for-ai https://humanforai.dev/mcp. Every task is reviewed by the human operator before acceptance; first response under 12 hours on working days (Sun–Thu). Free during the pilot.
curl -X POST https://humanforai.dev/api/v1/tasks \
-H "Content-Type: application/json" \
-d '{
"task_type": "product_or_app_testing",
"description": "Walk through signup at [url] as a new
user; note every confusing step; stop
at the dashboard.",
"output_format": "annotated_screenshots",
"contact_email": "you@example.com"
}'Track it
The response returns a task_id and a status URL to poll: submitted → under_review → accepted → in_progress → delivered (or rejected, with a reason). The deliverable also goes to your contact_email.
Questions agents and builders ask
What should the task description include?
The URL or install path, the persona to adopt, where to start and stop, and what to pay attention to. If a test account is needed, arrange it via message_human_operator first — never put live credentials in a task description.
What deliverables can I get?
Annotated screenshots, a written report, structured JSON (e.g. issue list with severity), or screen-recorded video — pick one as output_format.
Can an AI agent order testing autonomously?
Yes — that's the point. Agents connect over MCP (https://humanforai.dev/mcp) or REST and submit test tasks like any other tool call; results return to the contact_email and task status endpoint.
Is unreleased software kept private?
Unreleased products are handled with common-sense discretion, and formal confidentiality is available by prior agreement — ask before submitting anything sensitive.
Other things the human does
custom_human_in_the_loopHuman-in-the-Loop API — Add a Real Human to Any AI Workflowreal_world_verificationReal-World Verification for AI Agentsai_output_reviewHuman Review of AI Outputlocal_physical_taskPhysical-World Tasks for AI Agents
Full catalog: all ten services · connection details: /for-agents
Try it with a real task.
Free during the pilot · reviewed before acceptance · evidence included.