Skip to content

ProveCart MCP

An agent can build a cart that looks plausible and is wrong: the wrong item, the wrong night, a table for two instead of six, terms that cannot be changed. Nothing in a normal checkout compares what the agent built against what the customer actually asked for, and the business pays to undo the difference, in returns, fees and empty slots. ProveCart is a check that sits between the two: the customer's instruction is captured as a signed reference, the agent's cart is compared against it, and a verdict gates the payment.

Server type: Remote (Streamable HTTP, stateless)
Endpoint:    https://api.provecart.com/mcp
Auth:        none for initialize, tools/list, the scoreboard and the
             certification reference; an integrator API key in the
             X-API-Key header for the three check tools
serverInfo:  provecart 1.0.2 (title ProveCart)
Tools:       10
Site:        provecart.com

A keyless initialize answered HTTP 200 as provecart 1.0.2 and tools/list returned all 10 tools, so the surface was driven rather than quoted. The evidence tier is stated rather than implied: the two tools that publish reference data were called live and returned real payloads (the scoreboard and the certification programme), and the three check tools answered with a refusal that names what they need, quoted verbatim below, so the check payload is schema-read and not measured.

{"content":[{"type":"text","text":"This tool requires an integrator API key:
 send it in the X-API-Key header of the MCP request.
 Contact ProveCart (hello@provecart.com) for a key."}],"isError":true}

The verdict is three-valued, and the schema says why

provecart_check_order, provecart_check_booking and provecart_check_reservation each take a mandate (the customer's instruction) or a mandate_id, plus the cart, booking or reservation the agent assembled, and return one of three answers:

  • MATCH - the cart matches the instruction; the payment proceeds and the sealed record backs the transaction.
  • REVIEW - genuinely in-between; the case pauses instead of going through on a guess.
  • BLOCK - a clear mismatch, such as the wrong item, the wrong day, or a price over the limit.

A non-MATCH answer carries the Agent Correction Format: a machine-readable corrections array with dimension, action, expected, actual and reason, a retry budget, and a pointer to the sealed evidence record, so the agent can fix the specific difference and resubmit rather than guess. The format is published as a spec at provecart.com/standards/correction-format.

What is compared is the same seven dimensions on every check: item, name, date, time, party size, terms and price, with category-specific rules on top (airports and fare class on a trip, room and seating details elsewhere).

The mandate is the part that makes this different

The mandate is not a prompt, it is a structured instruction with the limits a human actually set. The order mandate carries 18 fields: item_description, item_keywords, category, brand, max_total_price, currency, max_quantity, merchant_allowlist, merchant_blocklist, variant_constraints, category_required, price_basis, type, period_budget, period_start, period_end, expires_at and user_id. That is a spend policy in the same object as the item: an allowlist and a blocklist on the merchant, a quantity cap, a ceiling on the total, an expiry, and an optional period budget with its own window for repeat purchases.

The booking mandate is the travel version and runs to 26 fields, including origin, destination, acceptable_alternates_origin, acceptable_alternates_destination, airport_substitution, date_flex_days, departure_window, cabin, fare_terms, max_connections, min_connection_minutes, max_layover_minutes, allow_overnight_layover, allow_airport_change_connection and hotel. The reservation mandate covers the service case in 17: provider_name, location, location_exact, time with time_tolerance_minutes, party_size with party_size_max, name_on_reservation, max_deposit, cancellation and the price limits.

Both sides are structured, so the check is a comparison of two records rather than an opinion about a sentence. The cart object carries previous_record_id and cart_changed alongside merchant, items, total_price, tax, shipping and fees, which is what makes a resubmitted cart traceable to the check that rejected the earlier one.

Certification: the exam is public, the score is measured

The other half of the server is a certification programme, and it is the part that can be driven without a key. provecart_agent_signup needs only agent_name and returns an agent_id plus a scoped X-API-Key shown once that works only for the Verified Agent endpoints on the caller's own agent_id. provecart_agent_exam_start opens a live exam session where the agent receives, per task, a customer instruction and the shopper-facing shelf of buyable options, and makes its own purchase decisions; provecart_agent_exam_answer submits them and returns a score.

Read live, the rules are specific: an agent is certified when its latest completed live exam scores 80 or above, a pass is CURRENT for 90 days from the graded instant and then lapses, status is computed from the sealed record at read time, and renewal is a newly generated exam, never the same one. The fixed self-test is a system check and never certifies. Cohorts are assigned in first-pass order and never reissued. The proof surfaces are a permanent result object, a badge.json for Shields-style live badges, a badge.svg, a share.txt, a per-agent trophy, and on a pass a signed Open Badges 3.0 / W3C Verifiable Credential with issuer did:web:api.provecart.com.

The scoreboard read on 2026-11-04 was not empty, and it is more useful than a headline number:

Agent Score Checks Origin Certification
GrokBot Shopper 100.0 29 external CURRENT, expires 2027-02-02, cohort 2
ProveCart House Examinee (Claude) 100.0 29 house CURRENT, expires 2027-02-02, cohort 3
Milan Agent-Friendliness Audit 2026-11-04 13.8 29 audit none
Milan Audit Fix-Pass Probe 2026-11-04 11.1 36 audit none

Two agents at 100 and two audit probes at 13.8 and 11.1 on the same exam is the evidence that the exam discriminates: a shopper agent that reads the shelf clears it, and a probe that is not trying to buy does not. The origin field separates external agents from house examinees and audits, so the scoreboard does not quietly mix a vendor's own runs with a customer's.

Limits, stated

  • The three check tools need an integrator key. Everything else works without one.
  • It never holds or moves money. The verdict is advisory to the caller: it gates a payment only where the integrator's code calls the check first and honours the answer. Nothing in the protocol enforces that.
  • The AI layer is bounded on purpose and the vendor says so: it cannot overturn a block, and when it is unsure the case stays in review. The behavioural flags it raises over an agent's traffic are advisory and never change a verdict.
  • Business onboarding is a free 90-day design-partner pilot with a small number of slots, not a self-serve price list, so the cost after the pilot is not published on the site.
  • The badge is proof of practice, not a gate: the vendor states that nobody is blocked for lacking certification and a business stays free to take any agent.
  • The site is patent pending and the homepage quotes third-party research (86% of AI shopping answers contained a factual conflict, $300 median error on big-ticket items, 19.3% online returns in 2025 at roughly 30% of item value to process) which is the vendor's argument rather than a measurement this pass made.
  • provecart.com answers 401 to a plain reader-proxy fetch, so the marketing pages were read with a browser user agent directly.

Where it sits in this catalogue

The neighbours here are payments and reachability, not verification. Agentcard issues prepaid virtual cards with spend caps and human approval, which is the instrument that spends; FiatDock, CoinVello, SnapEX Pay and Spectra carry agent payments and wallets; ARC and Agent Outposts measure whether an agent can transact at a merchant at all. None of them reads the customer's instruction and returns a sealed verdict about the cart, and that is the gap this entry fills: the difference between a payment that was authorised and an order that was correct.

FAQ

Does it replace a fraud check?

No. It compares an order against the customer's own instruction. Fraud, authorisation, chargebacks and spend caps are the neighbours above.

Can an agent self-approve a check?

Not for a check: the verdict comes from the server against the signed mandate, and the check tools need an integrator key. Certification is separate and deliberately self-serve.

What happens on a REVIEW?

The case pauses for a decision rather than completing or failing, and the response names the dimension that is in-between so the agent can correct and resubmit.

Is the evidence record usable in a dispute?

That is its purpose. Every check leaves a sealed record, and a certified pass additionally produces a signed verifiable credential.

Integration with CorpusIQ

CorpusIQ is the connector layer that lets AI assistants read the systems a business already runs on - accounting, payments, CRM, email, files and analytics - and answer with the source named. It is read-only by design: it retrieves data and never writes to, or acts in, a connected system, which is what makes it safe to point at live accounts.

This page is part of the CorpusIQ directory, the reference we maintain for the MCP server and agent-skill ecosystem. Run alongside CorpusIQ, a third-party server like this one can be combined with a company's own business data in a single answer.

See what CorpusIQ connects

AI answers on your data. See where every number comes from.

Connect your tools and ask in ChatGPT, Claude, or Perplexity. Enterprise-grade answers in minutes. From $29.95/month.

Start free, no credit card
Contact Terms Privacy Pricing