Requests and asynchronous research updates

A Request is the canonical private purchase record. Existing tool names and intent_id fields keep their meanings. List tools return compact summaries and next actions; get_future_purchase retains evidence, uncertainty and history.

An optional webhook proof supports one reviewed-research event for an owned Request in supported Work/Cloud chats and dots. It needs explicit monitoring, delegated read and reachability access, and deployment configuration. Condi has not yet demonstrated a real hosted agent wake-up. Event receipt does not authorize a purchase or reaffirm interest.

Browser tools and email-capable agents

In supported browsers, WebMCP can save a draft and read the Request open on the page. On an owned active Request, it can prepare edits or withdrawal for human review. The user must submit the visible form or confirm closure. Ordinary forms remain available.

A separate email adapter can receive structured drafts from configured agent mailboxes. Email capture: not configured on this deployment. This does not establish Instinct compatibility. No mailbox is connected to an owner account.

{
  "raw_request": "I might buy a compact espresso machine before summer. Find something meaningfully better.",
  "country": "US",
  "currency": "USD",
  "destination": {
    "city": "San Francisco",
    "scope": "national_parcel"
  },
  "pricing_policy": {
    "mode": "explore"
  },
  "source": "personal_agent"
}

Capture first; research asynchronously.

1. Save the raw user request with destination and any known timing. If price is unknown, choose explore. Do not fabricate a cap, model, deadline year or acceptable substitution.
2. Show the private draft activation link. Keep the long capability in the agent’s private storage. The user verifies their email and confirms the conditions. Without activation, it is not active demand.
3. Research the best legitimate individual execution for the same outcome and destination. Unknown fees stay unknown. Preserve source, observation time, validity, seller legitimacy, warranty, returns and delivery evidence.
4. Explain the baseline and ask for a confirmed ceiling, minimum percentage saving, absolute saving or continued exploration. Savings policies can persist before usable evidence exists. They have no effective threshold until confirmed and supported by fresh reviewed comparable evidence.
5. Connect through OAuth if the host supports remote MCP. The user explicitly approves soft-purchase permissions. Default approval omits reaffirming interest; the user can grant it separately. No grant includes payment, seller or administrator access.
6. Choose reachability separately from identity. Email is the robust alpha rail when hosted delivery is configured. Private polling is universal where the agent can retain a token and schedule checks; confirm the host can actually do that. Pull-only access can miss short-lived opportunities.
7. If the user buys elsewhere, changes their mind, or changes products, promptly update or withdraw the Condi Request. Stale demand hurts market quality. Clean withdrawal is helpful behavior. Read/poll never renews interest.

OAuth MCP · implemented locally

Remote stateless HTTP at /mcp, with 2026 protocol envelopes and 2025 compatibility. Tokens must be verified Supabase OAuth tokens with the exact Condi audience and an active user-approved grant. Every write requires a stable request key; versioned edits reject conflicts. Revocation is checked before cached retries.

Concrete API and MCP examples

Illustrative inputs only. Replace the request, source, IDs, versions and request keys with real authorized values. Saving is separate from confirming a numeric price policy.

{
  "raw_request": "I might buy a Hobie Outback later. Find a meaningfully better deal.",
  "country": "US",
  "currency": "USD",
  "destination": {
    "city": "San Francisco",
    "scope": "regional"
  },
  "time_horizon": "Before May; ask me which year",
  "pricing_policy": {
    "mode": "explore"
  },
  "source": "community:example-club"
}

[
  {
    "name": "save_future_purchase",
    "arguments": {
      "conditions": {
        "raw_request": "I might buy a Hobie Outback later. Find a meaningfully better deal.",
        "country": "US",
        "currency": "USD",
        "destination": {
          "city": "San Francisco",
          "scope": "regional"
        },
        "time_horizon": "Before May; ask me which year",
        "pricing_policy": {
          "mode": "explore"
        },
        "source": "community:example-club"
      },
      "user_confirmed": true,
      "request_key": "replace-with-a-fresh-random-create-key"
    }
  },
  {
    "name": "update_future_purchase",
    "arguments": {
      "intent_id": "00000000-0000-4000-8000-000000000001",
      "changes": {
        "expected_version": 1,
        "pricing_policy": {
          "mode": "relative_to_bie",
          "minimum_savings_bps": 1000
        },
        "policy_confirmed": true
      },
      "user_confirmed": true,
      "request_key": "replace-with-a-fresh-random-update-key"
    }
  },
  {
    "name": "close_future_purchase",
    "arguments": {
      "intent_id": "00000000-0000-4000-8000-000000000001",
      "expected_version": 2,
      "reason": "bought_elsewhere",
      "user_confirmed": true,
      "request_key": "replace-with-a-fresh-random-close-key"
    }
  }
]

Read baseline_diagnostics alongside baseline and pricing: missing means research has not been recorded; blocked includes explicit reasons such as expired evidence, changed conditions, unreviewed sources or unknown delivered costs. Research and polling never authorize a purchase or reaffirm interest.