What is UCP Playground?

An independent test harness for agentic commerce. Inspect any UCP store’s tools over MCP, REST or WebMCP, run AI shopping agents against it, and replay exactly what happened, call by call.

Do I need an account?

No. You can inspect a store and run an agent without signing in. A free account raises your limits and keeps your runs, and a team shares them.

Does it charge real cards?

Not on our test stores: payments there use test cards and nothing is charged. On a live store the agent stops before it pays.

Which stores can I test?

Any store that publishes a UCP profile at /.well-known/ucp, your own staging or pre-production endpoints, and our Retail, Stays and Food test stores.

How is it different from a UCP validator?

A validator checks that a store’s UCP profile is well formed. The Playground calls the store’s tools and has real AI models shop there, so you see what works end to end and, when a run fails, whether the store or the agent was at fault.

UCP PlaygroundThe universal test harness for agentic commerce

Test AI shopping agents on any UCP store. Point a real agent at a real store, debug your MCP, REST, A2A, WebMCP and Embedded Checkout Protocol integrations, compare models like Claude, GPT and Gemini, and share every run with your team.

Inspect tools without running an agent

A recorded run at the checkout: the store declined the card twice, and the Agent told the buyer. — A real run on our test store: the card is declined twice, and the agent tells the buyer nothing was charged.

Then prove it, call by call

Two views of the same store. Inspect is deterministic: send the store’s tools your own calls and the same call gets the same answer, so you can prove what each tool does. A run is probabilistic: the model chooses what to call and when to stop, so you see what agents actually do there.

Inspect flowers.ucpsandbox.com

Retail · UCP 2026-08-25

Recorded on our test store. Nothing is charged.

Loading the recorded calls…

Any UCP store, any vertical

Test a live store, your staging build or one of our test stores. Retail, Stays and Food each have a test store you can run an agent on today, and nothing is charged there. Hover a store to see it, pick one to start a run.

Any store that publishes a UCP profile at /.well-known/ucp works the same way, on any platform.

Every run, recorded and replayable

Replay any run step by step. The chat, the trace and the timeline move together.

Loading the recorded run…

Buyer view Exactly what the buyer saw, message by message.
Trace Wire by wire: every request and response the agent and store exchanged.
Timeline Each step on one clock, so you see where the time went.

Pinpoint what to fix in every agent flow

Every error is labelled with where it came from, so you know whether to fix the store or improve the agent, and how far each run got before it stopped.

Where it broke

mcpupdate_checkout → 400 · postal_code is requiredStore
mcpget_product → unknown product idAgent
mcpcomplete_checkout → declined by the issuerExpected

Declined cards, price changes and sold-out items show as expected answers, not errors.

How far it got

Searched only
Built a cart
Reached checkout
Tried to pay
Paid

The run above

Got to: Tried to pay
Calls: 6
Errors: 0

The three declined payments are expected answers, not errors.

Share any run as a link, an embed or a report.

Built for the teams shipping agentic commerce

Store and platform engineers

“We’re shipping a UCP profile.”

See your store the way an agent does. Call every tool over MCP, REST or the website, and test a staging build before it goes live.

Inspect and pre-production endpoints

Agent builders

“We need to test against real stores.”

Run the same task across 26 models from 8 makers, under the same conditions, and see how far each got and why it stopped.

Runs

Product and commerce teams

“We need proof of what an agent can do at our store.”

A replayable record of every run, shareable as a link or an embed, with each error labelled store or agent.

The run record and sharing

Run an agent on your store Read the docs

26 models, 8 providers

Pick the model in the run’s Model pill: 26 models from 8 makers, with Gemini 3.6 Flash as the default. Run the same task on another and see how each one ended.

See the leaderboard Browse the models

26 models · 8 makers

GoogleGemini 3.6 Flash · defaultGemini 3.1 ProGemini 2.5 ProGemini 2.5 FlashGemini 3.5 FlashGemini 3.8 FlashGemini 3.7 Flash

AnthropicClaude Opus 5Claude Sonnet 5Claude Opus 4.8Claude Sonnet 4.6

OpenAIGPT-5.5GPT-4oGPT-6 Lunao4-mini (Reasoning)

MetaMuse Spark 1.3Llama 4 Maverick

xAIGrok 4.5

Moonshot AIKimi K3

AlibabaQwen 3.7 MaxQwen 3.8 Flash

DeepSeekDeepSeek V4.1 FlashDeepSeek R1 (Reasoning)DeepSeek V3.2DeepSeek V4 ProDeepSeek V4 Flash

One workspace for the whole team

Create a team and everything moves with it: the runs, the stores you test, the buyer you shop as, the evals and the bill. Switch between your own work and the team’s in one click.

Members and rolesInvite people by email as an admin or a member. Admins manage the team, members run and inspect.
Shared runsEvery run in the workspace is there for the whole team to replay, compare and share.
Saved storesKeep the stores you test, live or pre-production, with their endpoints, so nobody sets them up twice.
One buyer contextSet the country, region, postcode, language and currency the team shops as. Every run sends the same signals.
Evals on a scheduleSave a set of tasks, run it on a schedule and send the results to your webhook.
Usage you can seeRuns, checkouts, tokens, stores and models each month, broken down by member.
Team API tokensTokens that belong to the team, not to one person, for CI and scripts.
Local runner for pre-productionRun the runner on your own machine to test endpoints that aren’t public yet, from localhost to staging. Its runs land in the workspace like any other.

Example

MembersUsageAPI tokens

Sam LeeOwner64 runs
Alex KimAdmin41 runs
Jo ParkMember23 runs

Runs: 128
Reached checkout: 52
Tokens: 2.4M

Why you can trust a result

Every claim on a run page can be checked against what was actually sent and received.

Every result is a recorded runOpen any run and read each request and response as it went over the wire. Nothing is summarised away.

The same conditions for every modelSame task, same store, same buyer context. Only the model changes, so the comparison is fair.

A signed, identifiable agentOur agent declares its UCP platform profile and signs every state-changing request (RFC 9421), so the store knows who is calling.

Independent, on the current releaseNot run by a store, a platform or a model maker. Calls follow the latest UCP release, and our census covers about 25,000 store domains.

Questions

What is UCP Playground?
An independent test harness for agentic commerce. Inspect any UCP store’s tools over MCP, REST or WebMCP, run AI shopping agents against it, and replay exactly what happened, call by call.
Do I need an account?
No. You can inspect a store and run an agent without signing in. A free account raises your limits and keeps your runs, and a team shares them.
Does it charge real cards?
Not on our test stores: payments there use test cards and nothing is charged. On a live store the agent stops before it pays.
Which stores can I test?
Any store that publishes a UCP profile at /.well-known/ucp, your own staging or pre-production endpoints, and our Retail, Stays and Food test stores.
How is it different from a UCP validator?
A validator checks that a store’s UCP profile is well formed. The Playground calls the store’s tools and has real AI models shop there, so you see what works end to end and, when a run fails, whether the store or the agent was at fault.