Ask your users why, while they can still tell you.
Code got cheap. Knowing what to build didn’t.
Ask your users why, while they can still tell you.
UserTold.ai runs a GPT-Live-1 voice interview about the moment a real user gets stuck, in the same session — the quote, the screen, and the page path stay together. You and your coding agent review the Evidence before it supports a Finding.
Start recordingWatch it in 46 seconds
GPT-Live-1 voice interviewsScreen + navigation contextMCP · CLI · REST$0.25 / recorded minute
Interview — trial user, day 3study: onboarding-friction
00:0001:0502:1003:1504:20
navigation/pricing00:03Arrived from the pricing page./settings/account01:12Opened account settings looking for billing.← back ×201:57Left and came back twice inside forty seconds./settings/account02:46Went back to show where they had been looking.screen01:12Scanning the account settings list.01:57A second pass over the same screen.voice02:22“I expected billing under account settings. I went back twice and still couldn’t find it.”03:02Answering the debrief: where they looked first, and what they expected to find.
01:12 observed02:22 said02:41 asked why
Observed01:12
Opens account settings. Scans. Goes back. Returns and scans again.
/settings/account → back ×2
Said02:22
“I expected billing under account settings. I went back twice and still couldn’t find it.”
Asked why — 19 seconds later02:41
The debrief asks about the exact moment it just watched — while the user can still explain what they expected and where they looked.
Evidenceextraction confidence 91% Quote, replay, and page path stay attached.
Draft FindingUsers look for billing in account settings — 3 linked moments from 3 users.
✓ Pushed to LinearAfter source review, send the Finding to Linear. Your team decides whether to take it into delivery.
● WatchingWhen the Linear issue completes, later Evidence can prompt a review for possible recurrence.
The gap
You have the users. You keep missing the moment.
Nobody books the call
You email ten users a calendar link. One replies. Zero show. The personalized feedback email? Crickets — even from the power users.
Analytics say what, never why
The funnel shows where they left. Session replays pile up that nobody has time to watch. The reason walks away untold.
Feedback arrives without context
A churn survey says “too complicated.” Which screen? Which task? By the time you can ask, they’re gone.
UserTold asks at the only moment users can actually answer: while it’s happening, inside your product.
How it works
From a real moment to a merged fix
Choose a research question, collect interviews, and review the Findings before sending them to your tracker.
01 · Study
Choose what to learn
Install the project script once. Set the study goal, participant task or questions, and pages where users can join.
02 · Interview
Interview users in your product
With consent, record conversation and product use. Screen recording is available on supported desktop browsers.
03 · Evidence
Review the source moments
Open evidence cards linked to what participants said or did, with the recording timestamp and available page context.
04 · Finding
Understand the pattern
UserTold groups related evidence into draft Findings. Check the sources and the summary, then mark a Finding reviewed.
05 · Delivery
Send a reviewed Finding
Choose which Findings to send to Linear for product triage or to GitHub for delivery. Each issue keeps its evidence links.
06 · Follow-up
Check after completion
Linear completion resolves linked evidence. Review similar evidence from later interviews to see whether the problem may have returned.
Evidence-first
Open the evidence. Don’t trust the summary.
AI research tools hand you conclusions. UserTold hands you conclusions with the receipts attached — what a person said, what was observed, and what the model inferred stay in three separate fields, next to the second of the recording they came from.
Evidence · struggling\_momentses\_xyz789 · 02:22
“I tried this flow three times and still cannot find where to change billing.”
Replay 02:22/checkout/step-3Extraction confidence 91%2 similar moments
The same evidence, as your agent sees it
{ "signal_type": "struggling_moment", "quote": "I tried this flow three times and still cannot find where to change billing.", "observed_facts": ["Returned to /checkout/step-3 three times"], "interpretation": "Billing settings are not where users look for them during checkout.", "confidence": 0.91, "intensity": 0.8, "interviewRef": "ses_xyz789", "timestamp_ms": 142300, "page_url": "/checkout/step-3" }
For coding agents
Your agent can ship. Now it can find out what’s worth shipping.
UserTold puts MCP, the CLI, and the API first. A coding agent can design a study, wait for real users to take part, and read the evidence as JSON, where every tool declares the exact shape of its answer up front, before it proposes a fix.
MCPmcp.usertold.ai/mcp — OAuth, sourced setup guidance, tools for projects.\ studies.\ interviews.\ evidence.\ findings.\*, and product feedback
CLIusertold study create --title "My Product Study" --activate --format json
HandoffReviewed Findings can be sent to Linear intake or GitHub delivery with quotes and source moments attached.
an agent-run study, end to end
read usertold://projects call studies.create … real users take part, inside your product … call evidence.list call findings.create_from_evidence … human or agent verifies the grouping … call findings.send → Linear UT-214On purpose
Three things UserTold refuses to do
Recruit strangers
No panels, no rented respondents. It interviews people already using your product — and a Study with no takers stays at zero Interviews instead of producing Evidence from nowhere.
Interrupt a struggle
During observation, stuckness is evidence. The interviewer doesn’t jump in to rescue, hint, or steer — it asks why afterward, when the moment is on record.
Fake certainty
What a participant said, what was observed, and what the model inferred are never blended. Every Evidence card carries its source and a confidence score.
Pricing
Priced like an API, not a research department
Pay for recorded interview minutes. No subscription, no seats, no per-interview minimum — billed on exact recorded seconds, itemized in Billing.
$0.25 / recorded minute
One predictable UserTold charge. We operate and pay for the interview AI.
5 min $1.2520 min $5.0060 min $15.00
$0.15 / recorded minute
UserTold charges the platform fee. OpenAI bills inference directly to your provider account.
5 min $0.7520 min $3.0060 min $9.00
Start with Managed AIPricing details
Questions
Before you ask
Where do the participants come from?−
From users your team can already reach. Invite them through your product or a direct study link. UserTold does not recruit participants or guarantee responses.
Who owns the data?+
Do I need Linear or GitHub?+
What kinds of studies can I run?+
GPT-Live-1 voice interviews inside your product
An interviewer that asks about what just happened.
Let a user try a real task. UserTold observes, prepares a running evidence summary when enabled, then GPT-Live-1 leads a natural voice debrief. Ask about their expectation while the experience is fresh.
Copy a complete interview script
Choose a SaaS template for discovery, onboarding, pricing, churn, concept testing, or feature requests.
Understand onboarding drop-off
Observe setup, ask about the point of hesitation, and review what prevented first value.
Understand pricing confusion
Learn how a buyer interprets your plans, usage limits, and charges before choosing.
Discover the need behind a request
Connect the requested feature to the task, expected outcome, and current workaround.
The next thing you ship should be something a user told you.
Add the script to your product, verify it on one page, and take the starter interview yourself. Read the evidence it produces before you point it at real users.