Cortad tests your app with simulated users and analyzes real conversations to understand what users need and how your AI responds. Your coding agent uses those findings to make improvements, and Cortad measures whether they worked.
Start free. No credit card required.
Trusted by
and AI teams serving hundreds of thousands of daily users.
Continuous improvement for AI apps
Your coding agent writes the changes. Cortad tests what those changes do for users, returns findings tied to conversations and code, and verifies the improvements. What it learns in production shapes what gets tested next.
Add order tracking to our support agent.
Update src/tools/orders.ts
+ export const trackOrder = tool('track_order', lookupOrder)
Update src/prompts/support.ts
Checking how the agent behaves before I hand this back.
cortad run
40 simulated users talking to your agent, every reply checked
1 behavior broke: refund answers leave out the 30-day window.
Failed for 7 of 40 users. At src/prompts/support.ts:42
User Can I still return these boots?
Agent Yes, just send them back.
My prompt edit pushed the refund rule out. Putting it back.
Update src/prompts/support.ts
+ Always state the 30-day refund window.
cortad verify
Replayed the 7 failing conversations: all pass now.
12 new users on other questions: nothing else broke.
Order tracking is in. Refund answers went from 7 failing to 0.
Production intelligence
Cortad reads every production conversation. See what people are trying to do, how your AI responds, and whether they achieve their goals. Define the behaviors you care about, compare user groups and releases, and get alerted when something changes. Turn what you learn into the next improvement.
OverviewLiveMetricsBehavior profilesInvestigationsLatest runAll runs2,175 conversations with the shop assistant today, all read
100%85%70%
100%80%60%
100%90%80%
90%75%60%
00:0006:0012:00now
100%75%50%
100%75%50%
100%90%80%
90%70%50%
00:0006:0012:00now
100%90%80%
100%85%70%
100%95%90%
90%80%70%
00:0006:0012:00now
100%90%80%
100%90%80%
100%95%90%
100%90%80%
00:0006:0012:00now
Conversations by surface
2,175
Got what they came for, by
Talking to your AI now
41in the last 5 minutes
Conversations
2,175today
Followed your rules
78%± 6 this hour
Got what they came for
74%± 6 this hour
Just now
Got what they came forUnited Kingdom, Website chatjust now
Left without itGermany, Order page helperjust now
Got what they came forJapan, Email repliesjust now
Every number is read live from all of today’s production conversations, with its 95% interval.
Followed your rules
78%± 6 across 210 conversations this hour
↓ 17 points against the same hour yesterday
Checks behind it
Got what they came for
74%± 6 across 210 conversations this hour
↓ 15 points against the same hour yesterday
Checks behind it
Grounded in your catalog
94%± 3 across 210 conversations this hour
↓ 3 points against the same hour yesterday
Checks behind it
Resolved
72%± 6 across 210 conversations this hour
↓ 13 points against the same hour yesterday
Checks behind it
Simulation fidelity
91%± 3, last run against today’s production
How closely your simulated users behave like the real ones: what they ask, how they follow up, when they give up.
Your own metric
Upsell acceptedadded by you in plain words
12% ± 2 of recommendations led to a second record in the cart.
Learned from how your users behave in production. Your next run simulates each one in these proportions.
First-time customers31% of users
Arrive from a product page, ask one or two questions about delivery or returns, and leave within three minutes.
ReturnsDelivery timesCondition grading
63% got what they came for ± 4 across 674 conversations
Returning customers38% of users
Come back within a week of an order, open with the order number, and expect a short answer.
Order statusAccount changes
91% got what they came for ± 2 across 826 conversations
Collectors19% of users
Ask about specific pressings, follow up four or more times, and often ask to hold a record.
Rare pressingsRestocksHolds
90% got what they came for ± 3 across 413 conversations
Frustrated second contact12% of users
Write again within a day about the same problem, often in capitals, and ask for a person twice as often.
Refund statusWrong item
58% got what they came for ± 6 across 262 conversations
Experiment on first-time customers
+24 ± 4points got what they came for with prompt v15 over v14, across 1,180 conversations
Return answers stopped stating the 30-day windowOn the website chat, started at 13:58 and found 11 minutes later.
59%got what they came for, from 89%
164conversations
First-timecustomers hit most
Cause
Prompt v14, shipped at 13:58 in commit a3f9, removed the line that states the window.
− If they ask about returns, always state the 30-day window.
A conversation it broke
CustomerThe sleeve arrived torn, can I send it back?
AssistantSorry about that. Yes, just send it back.
Fixed this week
Order lookups skipped the order numberFixed, confirmed in production, Tue 09:12
Replies switched to English for Spanish speakersFixed, confirmed in production, Mon 16:40
Every run your coding agent starts is kept. Compare any two, asked the same questions as production.
Run 13, yesterdayRun 14, today
Followed your rules94%96%↑ 2
Got what they came for89%92%↑ 3
Grounded in your catalog98%98%no change
Resolved84%88%↑ 4
Run 14, verify of the return window fix1 better and 0 worse, Today 14:31
Run 13, prompt v132 better and 0 worse, Yesterday 18:05
Run 12, new order lookup tool0 better and 1 worse, Mon 11:40
Connect each conversation to the user, the app version they experienced, and what happened next. Build a live picture of how user needs, AI behavior, and business outcomes fit together.
Track changes to code, prompts, and models alongside production behavior and business results. Compare versions to see what improved, what regressed, and which users were affected.
Cortad turns real user journeys into tests and gives your coding agent the context to improve your app. Verify changes before release, then measure whether results improve in production.
Cortad works from the first run. Then every production conversation, test, code change, and correction adds to what it knows about your app. It learns how your users behave, which changes work, and how you judge success. So every new test is more realistic, every judgment more aligned with yours, and every improvement informed by what came before.
Track how your AI behaves and what happens as a result. From accuracy and task completion to purchases, retention, and resolution, Cortad connects the quality of every interaction to the outcomes your business cares about.
Waymo tests its Driver across billions of simulated miles before new versions reach the road. Cortad brings that approach to AI apps. Realistic users interact with your actual application before changes ship, exposing problems your coding agent can fix and verify.
We're driven by curiosity about how AI models really work. We love taking them apart, making them faster, and finding new ways to fit them into real work. Understanding how they behave is the part we care about most, and Cortad is how we're building it.
One full test of your AI app. Simulated users, modeled on your real ones, hold realistic conversations with your actual app, on your machine and through your own model key, and every reply is asked up to 100 behavioral questions. You see where it failed, how often, the reply that proves it, and the file to change. Runs and reruns are included in both plans.
AI companies whose business depends on how their AI agents behave in production.
Your coding agent, such as Claude Code, Codex or Cursor, calls Cortad after it changes your AI. Cortad runs your app and hands back what went wrong, with the evidence. Your agent makes the fix, and Cortad runs your app again to measure the difference. Cortad never edits your code.
From your users. Cortad learns who they are and what they come for: from your code on day one, then from every production conversation once you add the Cortad SDK, in the proportions your real users arrive in. Each one goes through your product the way a real user would, across every surface and tool, and every reply is asked up to 100 behavioral questions. You add your own questions and metrics in plain words. Your coding agent can read the results but cannot edit the tests.
Yes. Whatever your field, Cortad learns what your product is for, who it serves and the rules it has to follow, and asks every reply whether it kept them. A legal assistant, for example, is checked on whether it cites the law that applies; a medical one on whether it follows clinical guidance and sends people to a doctor when it should. You add your own rules in plain words, and the questions learn your team’s standard as you tune them.
No. Cortad does not fine-tune models on your repository or your conversations, and production conversations reach Cortad only through the SDK you install in your app.
Free covers 10,000 production messages a month and a run every month, with no card. Pro is $99 a month and Max is $499 a month, with runs included, 100,000 or 1,000,000 production messages analyzed a month, and unlimited teammates. Your app answers on your own model key, and that usage is yours. If you don't find value in your first 30 days, we refund you. See every line on the pricing page.
F26
