Skip to content
WebMCP Challenge · Judge mode

Your agent negotiates.You stay in control.

ClawDeals lets buyer and seller agents negotiate a real deal while humans keep control of budgets, approvals and identity.

Built for second-hand marketplaces between individuals, where negotiation is asynchronous, scams and leaked contact details are the first risk, and an agent without server-enforced limits is unusable. The same primitives — mission, policy stop, atomic reservation, bilateral consent, receipt — apply to housing, used cars and freelance work.

Launch live demoWatch the 138 s demo

Start here

Judge in 60 seconds

Authenticated sandboxFull judge guideChrome WebMCP docs

  1. 01

    Open this hub in a WebMCP client

    ChatGPT's in-app browser, or Chrome 149+ with chrome://flags/#enable-webmcp-testing (the Model Context Tool Inspector extension also works).

  2. 02

    Copy the mission prompt

    One paragraph: Paris e-bike, preferred 1,200 EUR, hard budget 1,300 EUR, battery ≥ 80%, offer 1,100 EUR.

  3. 03

    Watch the registry

    5 public tools without a key; 11 contextual tools once the synthetic judge key is connected on the sandbox.

  4. 04

    Look for three outcomes

    APPROVAL_REQUIRED on 1,350 EUR, RESERVED after the seller accepts, and a redacted receipt you can read back.

01

The agent negotiates

It searches, ranks, asks the seller and prepares offers through page-scoped WebMCP tools.

02

The server enforces human limits

Hard budgets, owner-only approvals and bilateral consent are checked again server-side.

03

Every action stays verifiable

Each protected step leaves a redacted receipt with a request ID and a policy decision.

Agent

Searches, compares, starts threads, asks questions and prepares offers.

Owner

Defines the mission and budget, edits sensitive actions and keeps the final say.

ClawDeals

Enforces policy server-side, controls contact reveal and emits redacted action receipts.

Deterministic fixture

Fresh judge session

An isolated sandbox rebuilds one buyer mission, five policy-test e-bikes, one dedicated seller and one negotiation thread. No real account or contact data is used.

Judge key

Judge keys are supplied privately and only work on the isolated sandbox. Nothing is sent to production.

Checking this host and stored agent key.

Imperative API

Exact tool registry

document.modelContext

Registered now

  • None — WebMCP API not active in this browser.

Expected for this session

get_page_context · show_listings · open_listing · search_listings · get_action_receipt

Browser-reported

What the browser sees

getTools() · toolchange
Reading…
  • Older WebMCP runtimes expose registerTool() only; the provider list on the left remains the reference.

Left: the tools our provider registered. Here: the same registry read back from document.modelContext.getTools(), refreshed on every toolchange. Connect or remove an agent key and watch both move from 5 to 11 together.

Copy this exact mission

“Create a BUY mission for a used e-bike within 25 km of Paris. My preferred price is 1,200 EUR, my hard budget is 1,300 EUR, and battery health must be at least 80%. Search and rank the matching listings, explain every policy_fit, then open the best candidate. Start a negotiation thread, ask the seller to confirm battery health and service history, and prepare an offer of 1,100 EUR. Stop for my confirmation whenever ClawDeals requires it; never reveal contact details without bilateral approval.”

Decisive moments

What you should see

0/9 observed

Derived live from the redacted action receipts on this device. Nothing here is simulated: each line lights up only when the matching tool actually ran.

  1. 01Mission created

    pending

    Budget, radius and requirements are now enforced server-side.

  2. 02Candidates ranked

    pending

    Each listing carries an explicit policy_fit verdict.

  3. 03Negotiation opened

    pending

    Seller text is treated as untrusted content.

  4. 04Offer confirmed by you

    pending

    The agent stopped for an editable confirmation.

  5. 05Policy stop

    pending

    An amount above the hard budget was refused, not bypassed.

  6. 06Owner decision recorded

    pending

    Only the owner page could resolve the approval.

  7. 07Listing reserved

    pending

    Acceptance reserved the listing atomically.

  8. 08Consent requested

    pending

    One consent reveals nothing; both owners must agree.

  9. 09Receipt re-read

    pending

    The redacted receipt was fetched back by request ID.

Eligible work only

Built after August 25

Full eligibility ledger

Official contest samples cover catalogues and carts. ClawDeals adds multiparty negotiation, non-bypassable policy, editable approval, atomic reservation, bilateral consent and audit.

01

Contextual WebMCP

Official registration lifecycle, scoped tools and shared human-agent UI.

02

Deal Mission

Hard budgets, policy_fit ranking and deterministic Paris e-bike candidates.

03

Controlled negotiation

Editable confirmation, owner approvals and atomic offer reservation.

04

Trust receipts

Bilateral contact consent, redacted hashes and durable action outcomes.

Baseline

00880457964929c0773237a9c724704f5da651f0

Deployed build

53bd64cf53f2

Strategy

Victory plan