Don't ship MCP servers that fail

Simulate many user scenarios across ChatGPT, Claude, and more in parallel. Catch the edge cases your tests miss.

1. 1Describe
2. 2Confirm details
3. 3Run swarm

Leave

Create a swarm of your users

Simulated users run through your server so you can see what breaks.

Choose the clients and servers your users will interact with. *

This will shape the swarm's behavior. You will be able to compare behavior across clients side by side.

ChatGPTClaudeCursor+ Add client

acme-payments· v2.15.0-rc.1 · 5 tools

Describe your users and their behavior. *

We build the user goals from your input. Your inputs are not final. You can edit personas and goals on the next screen.

Support agents refunding duplicate charges, finance ops reconciling Tuesday's payouts, and developers wiring a failed-payment webhook.

Add existing persona

Generated personas

• MCMaya Chen · Support agentFind the duplicate charge on order A-2291 and refund it.8 × 3 clients
• JHJordan Hale · Finance opsMatch payout po_7Hd2 to its charges and flag anything missing.8 × 3 clients
• PSPriya Shah · DeveloperCreate an invoice, fail its payment, and confirm the webhook fires.8 × 3 clients

CancelContinueRun swarm

See how real users interact with your product. Measure sentiment and uncover usability gaps.

Refund & reconcile · beta

Findings2InsightsSessions15

EditOpen previewShare

Access settings

Anyone with the link (guests included)

Has access

• MMara Okafor (you)[email protected]Owner
• Dana Reyes[email protected]Tester
• Luis Moreno[email protected]Tester
• Sam Kaur[email protected]Tester

Ratings

Per-turn ratings

starsthumbs

Preview

ChatGPTMCPJamCopy link

Refund the duplicate charge on order A-2291

list_charges· refund_charge

I found two charges for order A-2291: ch_3RkA9 for $129.00 and ch_3RkQ9 for $129.00, 41 seconds apart. I refunded ch_3RkQ9 as a duplicate. The customer will see it in 5 to 10 business days.

Rate this response

Send

Message…

AI can make mistakes. Please double-check responses.

Previewing as you, signed in. A tester opening this link gets the guest experience, which can differ.

Per-turn rating · zoomed

How did the last few turns go?

Picked the right tool. Could've grouped by service.

Send →

Turn user workflows into repeatable, durable test suites. Re-run them automatically across your preferred clients.

Evaluate

OverviewSuites

SetupRun

Suite health

Payments operations

98%average across 12 runs

acme-paymentsPayments operations

100%95%

9/29/39/49/59/69/79/89/99/109/119/129/13

Runs

PlatformSuiteClientServer

Run	Suite	Client	Result	Rate	Platform	Date	Latency	Tokens	Calls
#128	Payments operations		Failed	96%69/72	GitHub	Sep 13, 2:00 AM	18.4s	16.2k	44
#127	Payments operations		Passed	98%71/72	CLI	Sep 12, 2:00 AM	14.2s	12.1k	41
#126	Treasury reconciliation		Passed	100%48/48	SDK	Sep 12, 1:14 AM	16.8s	15.6k	38
#125	Billing exception handling		Failed	94%45/48	CLI	Sep 11, 11:42 PM	21.1s	18.9k	47
#124	Payments operations		Passed	99%71/72	GitHub	Sep 11, 2:00 AM	11.3s	9.4k	29
#123	Dispute resolution		Passed	100%36/36	SDK	Sep 10, 6:18 PM	15.5s	11.8k	33
#122	Treasury reconciliation		Passed	97%70/72	GitHub	Sep 10, 2:00 AM	17.9s	14.7k	40

Block failures before they ship. Make durable cross client evaluation part of every release.