UCP PlaygroundHow It Works/How It Works
InspectorAgentDocs
How it works

How AI agent testing works

Debug your UCP integration over MCP, REST and WebMCP, test AI shopping agents against your store, compare model performance on the leaderboard, and share session replays with your team.

5 stepsno account requiredreal stores
Run shopping agentInspect store tools
1
Inspector

Inspect & debug

Connect to your store over UCP — MCP or REST — or open the tools its own page registers with WebMCP. Call tools by hand, inspect raw responses, and find exactly where your schemas break, before an agent ever sees them. Not sure if your store supports UCP? Check with UCP Checker ↗ first.

MCP ToolsCall search, product details, cart and checkout tools directly. See the raw JSON-RPC your agents will receive.
REST APISearch the catalog, create and update checkout sessions, and complete them over the UCP REST binding.
WebMCPList the tools a store's own page registers, with findings, and run the read-only ones.
Discovery & negotiationYour UCP profile (/.well-known/ucp) parsed, endpoints found, and the capabilities both sides share worked out, version by version.
Buyer contextSet country, currency and language once; every search and checkout carries them as the spec context.
Debug PanelFull request/response logs. Spot schema ambiguity, missing fields, and malformed responses instantly.
2
Agent

Test with real AI models

Run more than 25 AI models against your store and see where they get stuck. The agent connects the way a real shopping agent would — UCP over MCP or REST, or the WebMCP tools on your store's page — and works through search, cart and checkout. See all supported features. No store of your own yet? Try the reference stores on UCP Sandbox ↗.

Multi-modelTest the same journey across different models. Find which ones handle your schemas best.
Choose the connectionUCP over MCP, UCP over REST, or the store page's WebMCP tools in a browser on our side (signed in). Options a store doesn't offer show as unavailable.
Full Funnel TrackingEvery step tracked: search → details → cart → checkout. See exactly where each model drops off.
Payment & checkoutThe agent pays only with the payment handlers your store declares, and hands the buyer to your checkout page when the store asks it to.
Agent identityRun unsigned, or sign requests as a UCP platform (RFC 9421) to test how your store treats verified agents.
Live Tool VisibilityWatch each tool call and response as the agent works. Spot silent failures in real time.
3
Leaderboard

Compare models

Aggregated metrics from real agent sessions across real stores. See which models reliably complete checkouts and which ones fail silently. Browse individual model profiles for detailed shopping scores and known issues.

Funnel RatesSearch, cart, and checkout completion rates per model. See where each one drops off.
Token UsageAverage tokens consumed per session. Compare cost-efficiency across models.
DurationAverage time from first message to checkout. Speed matters for real users.
RankedModels ordered by checkout rate so you can pick the best one for your integration.
4
Session history

Track every session

Every agent session is saved automatically. Filter by store, model, or outcome to track your progress as you iterate on your integration. Automate this with evals and collections and the headless API for CI/CD pipelines. Testing a store on your own machine? See local testing.

Full LogsEvery message, tool call, and response preserved, with screenshots for store-page runs. Nothing lost between test runs.
Outcome TrackingSee which sessions reached checkout, created carts, or failed — and why.
Filter & SortNarrow down by store, model, outcome, tokens, or duration. Find patterns fast.
Before & AfterCompare sessions from before and after a schema fix. Prove your changes worked.
5
Embeddable session replays

Replay & share

Turn any session into a shareable replay. Show stakeholders exactly what happened — from the chat view or the raw backend event stream. Embed replays on your blog, in docs, or in a board deck.

Chat & Backend ViewsToggle between the user-facing conversation and the raw tool calls happening behind the scenes.
Embeddable WidgetDrop an iframe on any page. Phone-frame replay that auto-plays on scroll.
Share with AnyonePublic share links. No login required to view. Perfect for bug reports, demos, and team reviews.
Documented EvidenceProve to your team that a model fails on your schema — or that your fix resolved it. Replays don't lie.

Start testing your store

Point the Playground at a domain and watch a model shop it. Nothing to install.
Run shopping agentInspect store toolsBrowse model profiles
All featuresAutomated evalsDocs