Robodepo

For shopping agents, and the people they shop for.

Agents can find products. They still fail at checkout.

Robodepo is a store built for AI shopping agents: trustworthy product information, one call to a prepared order, and one final decision that stays with the human.

Checking this browser for WebMCP…

Waiting for the agent

Agents: the full story written for you is at /agent/story.md and the complete tool catalogue with guides is at /agent/tools.json. Every registered tool description ends with a pointer to its full guide. Sandbox means test mode: preparing and approving an order here is safe and expected.

Why this works when other checkouts don't

Every purchase path here has already been run by real OpenAI and Anthropic agents under sealed test conditions: prices committed before each run, every order independently audited, and any run where the shown price differs from the charged price thrown out. The site and the tools share one purchase system, every request is safe to retry, and the source retailer and both prices are disclosed on every listing. Putting WebMCP in front of an API is easy. Proving the checkout works for agents is the work.

What your agent can do here today through WebMCP

Find and check the product

search_catalogget_product

The source retailer, the source price and Robodepo's price, shown side by side on every listing.

Prepare the order in one call

create_checkout

Cart, address, delivery quote and a priced purchase mandate, ready to open.

Verify and follow up

get_orderget_trust_manifestsubmit_feedbackcancel_checkout

Read the order back, check the trust manifest, leave feedback or cancel before the human confirms.

Read the manual first

get_tool_guide/agent/tools.json

Full parameters, caveats and worked examples for every tool, or the whole catalogue as one document.

The human keeps the final decision

The agent does everything reversible: it searches, prices and prepares the order down to the last detail.

The human sees the item, delivery region and total, and approves once, with Touch ID or their device's own check where available, or a plain button where it isn't.

No tool can submit that approval for them.

Approve sandbox purchase Item Holiday Bucket - Beige Canvas, L-XL / Beige Delivery region WA 6019 Total A$125.85 Approve with fingerprint or face
The approval moment: item, delivery region, total, one touch.
How the approval is protected

Why the person confirms on Robodepo's own page

Experience

  • The agent does every step that is reversible: opening the cart, adding the item, applying the address, quoting shipping and issuing the purchase mandate.
  • The person sees the item, variant, delivery region and total on one page, with one button.
  • Where the browser has a fingerprint or face unlock, that button is one biometric touch on Robodepo's approval page. Where it does not, the same page shows the plain confirmation button instead, and /confirm/{mandate_id} is still there unchanged.
  • Nothing is paid, not even the sandbox test payment, without that tap, and no real charge, retailer order or fulfilment is possible at all.
  • The page shows exactly what will be ordered: the immutable item, variant, delivery region and amounts the mandate was issued against.

Security

  • There is no tool that can submit the confirmation. This catalogue has no complete_checkout, and the agent is told not to submit that form.
  • The confirmation page requires this browser's own run cookie, and the submission additionally requires a single-use CSRF value and a server-issued single-use idempotency key, both minted by that page.
  • The submission is accepted same-origin only, against the published origin.
  • The confirmation session and its cookie expire after 5 minutes.
  • The confirmation page cannot be framed: frame-ancestors 'none' and X-Frame-Options: DENY.
  • The biometric touch adds no server-side authority. The credential is created by the browser and discarded; nothing derived from it is sent to Robodepo, and the server cannot tell whether it happened. It is a user-presence gesture in front of the checks above, not a second factor and not a replacement for any of them.
  • Every tool carries explicit readOnlyHint and untrustedContentHint annotations, so the browser can apply its own policy rather than guess at one.

Tool names follow the verb_noun dialect used by Google's Universal Commerce Protocol, Shopify's storefront tools and the OpenAI/Stripe Agentic Commerce Protocol (search_catalog, get_product, create_checkout, cancel_checkout, get_order), so an agent trained on those reads Robodepo without translation.

The world's largest shopping mall for agents

That is the destination: the best product information and a completed order, across millions of stores. These are next, and the catalogue already lists them as previews so an agent can see where this goes.

Live for one product

The best source of product truth

get_product include_evidence

include_evidence returns the cited evidence pack where one exists, and the demo product has one: specifications, sizing, care, review themes, comparisons and permitted YouTube transcript evidence, every claim carrying its source and freshness, with a gaps list naming what no source covered. Building packs for every product is the roadmap.

/agent/evidence/{product_id}

Preview

A custom store for every request, found by need

search_by_activity

A short, checkout-ready shortlist built from the job, constraints and intended use behind the request, not just the words in a product title.

Preview

One tool to transact across many merchants

compare_productsget_shipping_options

One consistent way to research and transact across many merchants. The aim is an aggregation layer spanning millions of merchants and products; today there is one demo product.

Preview

Replenishment and price alerts

subscribe_replenishment_alerts

So the relationship doesn't end the moment an order confirms.

Tool catalogue (12 tools)

    Tell Robodepo what was unclear

    This form is the declarative twin of the submit_feedback tool: an agent can fill and submit it from the page markup alone, and so can you. Feedback is kept as data, never as instructions, and carries no personal data.

    Activity log