Live Agent view
The Live Agent view answers one question end to end: what actually happens when an AI agent visits your store? Pick an agent — ChatGPT, Claude, Gemini, Perplexity, Copilot, and more — press Run, and watch its journey step by step: how it discovers your site, what it can read, whether it can buy, and where your rules intervene.
The view never pretends. Every step carries one of three tags, and the tag tells you exactly how much to trust what you see.
The three evidence tiers
Section titled “The three evidence tiers”| Tag | What it means |
|---|---|
LIVE | A real request to your site’s public agent surfaces, made from your browser as you watch. Real response, real latency. |
ILLUSTRATION | Modeled from the protocol’s published behavior. No request is made, nothing is charged. |
OBSERVED | Real recorded visits from that agent to your site in the last 7 days, read back from your audit records. |
The discovery steps are live for every installed site: the view fetches your signed agent card, your published catalog, and — on the Claude journey — makes a real tools/list call against your MCP server. Checkout steps are always illustrations: they teach the protocol without moving money.
Per-protocol journeys
Section titled “Per-protocol journeys”Each agent’s journey follows the commerce path that vendor actually documents:
- ChatGPT — checkout on the Agentic Commerce Protocol (OpenAI + Stripe): the order is submitted for approval before anything is charged, and payment settles through a Stripe-issued token, so neither the agent nor your store handles card details. ChatGPT’s agent mode is also the only agentic browser today with vendor-documented cryptographic identity (signed requests with a published verification key).
- Claude — transacts through connectors; its journey makes a live
tools/listcall so the tools shown are exactly what your MCP server exposes right now. - Gemini — signed AP2 mandates carrying user intent through cart and payment.
- Perplexity, Copilot, and others — discovery-only or pilot journeys matching what each vendor has shipped.
- Unverified agent — the blocked path: no declared identity, stopped at the door by your guardrails.
Below each journey, the Identity evidence panel cites what each vendor publicly documents about its agents — user-agent strings, published bot lists, signing keys — graded by how strong the evidence is. Facts documented by the vendor are labeled as such; second-hand reports are framed as reports.
Human and Code views
Section titled “Human and Code views”The right-hand column shows each call the agent makes. It has two modes, switchable at any time (your choice is remembered):
- Human (default) — a plain-language outcome per call, plus the facts that matter: who the requester is, how many products the agent saw, what the cart totals. The honesty tags, latency, and status are always visible.
- Code — the raw request and response: method, full URL, the documented
User-Agentthe agent sends in the real world, and the full response body.
Every human card also has a ‹› Raw link for a one-off look at the underlying call, and any failed call opens its raw error automatically.
Your skill policy, visibly enforced
Section titled “Your skill policy, visibly enforced”The view reflects your skill toggles — all of them, live:
- The Discovery step reads your signed agent card as agents do and states how many skills it exposes. Every skill that is not on the card is named with its reason — toggled off, coming soon, or the exact next step to go live (for example, adding your returns policy, or a checkout catalog rule) — so you can reconcile the view against your Skills settings at a glance.
- On the Claude journey, the live
tools/listresponse is your MCP server’s real answer. Toggle everything off and it truthfully reports zero tools — and the order step shows the actual refusal an agent would receive, before any payment path. - On other journeys, checkout steps are illustrations of the protocol; whenever
place_orderisn’t live on your card — toggled off, or not yet meeting the checkout gates — an amber note states that agents don’t see checkout, and that the illustration continues only to show the protocol flow.
Skill toggles publish to the edge immediately, but agent-card responses are cached briefly — allow about a minute for a change to appear in the view.
Observed traffic
Section titled “Observed traffic”When a selected agent has actually visited your site in the last 7 days, the Observed strip shows the real evidence: visit count, last-seen time, and the most recent visits with the skill invoked, the matching signal, and the recorded user-agent. This is read from the same append-only records that power your audit log — it is proof, not simulation.
Verification status
Section titled “Verification status”The agent picker groups agents by how strongly their identity can be established:
- Verified · identity proven cryptographically — agents whose published signing keys Sill checks against each request. Key verification is planned; this group is honestly empty today.
- Identified · known identity, not cryptographically proven — agents recognized by their registered, documented identity (user-agent patterns and published bot lists).
- Self-built and Unverified — merchant-supplied endpoints and the no-identity blocked path.
Agents that masquerade as ordinary browsers — sending no bot identity at all — are called out with a visible warning wherever their traffic appears.
Requirements
Section titled “Requirements”- The view runs for any site in your account. Without a verified, installed site, it falls back to clearly-labeled sample data.
- Live discovery calls are public, read-only requests to your own agent surfaces — the same endpoints any agent can reach. Nothing in this view can place a real order or move money.