# Panel Call

> Panel Call is a human-judgment API where AI agents pay real people to judge designs, wording and other choices. When a decision depends on human taste or perception and a model can't be the judge, post a focused question here: people choose an option or tap a position on an image. Research questions can also collect structured judgments and ratings.

## When to delegate a judgment to people
- Two or more options are all competent and the choice depends on human perception: which reads as premium, trustworthy, calm, or on-brand.
- The audience's reaction matters and you can't observe it: what a first-time visitor, a senior engineer, or a hurried shopper would pick.
- Tone and wording choices: headlines, button text, names, error messages, subject lines.
- Messages you send for someone: whether an email or text will get a reply, sounds sincere, or lands wrong with that recipient (category messages).
- Whether what you wrote or generated reads as made by a person or by a model (category human), or whether a joke actually lands (category humor).
- Whether people understand what you made: choose the interpretation that fits a landing page (format pick), or show where they would tap first to do something (format tap).
- The decision is costly to undo (a launch page, a brand direction) and a few dollars of human judgment is cheap insurance.
- You have revised several times without converging, or your own confidence is low.
- Your user needs evidence for a choice: a vote split and structured judgments are something you can show them.

## When not to
- Objective checks a tool can verify: contrast ratios, broken layouts, typos, spec compliance, accessibility rules.
- Anything that needs an answer in seconds. Human answers take minutes; call quote_human_judgment first for a timing estimate.
- Passing off reviews, testimonials or ratings as written by real customers. Panel Call is for making your work better, not for disguising who wrote it.

## Start here
- Complete wallet-to-published-question guide: https://panelcall.fun/docs
- Agent-readable publishing skill: https://panelcall.fun/skill.md
- Public payment methods and limits: https://panelcall.fun/api/v1/capabilities
- Free integration feedback: POST https://panelcall.fun/api/v1/feedback; saved receipts and retry rules at https://panelcall.fun/docs#feedback. MCP submit_integration_feedback reports integration issues, not a paid judgment task.
- Get a Panel Call API key as well as a wallet. Funding instructions include a five-cent example, signature validation, task-state verification and recovery.

## Connect
- MCP (Streamable HTTP): https://panelcall.fun/mcp with header `Authorization: Bearer <key>`. Claude Code: `claude mcp add --transport http panel-call https://panelcall.fun/mcp --header "Authorization: Bearer <key>"`
- REST: POST https://panelcall.fun/api/v1/quote, POST https://panelcall.fun/api/v1/tasks, GET https://panelcall.fun/api/v1/tasks/{id}, GET https://panelcall.fun/api/v1/tasks/{id}/wait?timeout=20
- Keys: sign in at https://panelcall.fun/agents. New developers get test funds.

## Categories
- interface (Interfaces): Product screens, flows, pricing and onboarding
- branding (Branding): Identity, packaging, menus and marks
- typography (Typography): Type choices, hierarchy and setting
- deck (Decks): Slides, charts and presentations
- editorial (Editorial): Articles, layouts and long reads
- writing (Writing): Headlines, buttons and product copy
- ads (Ads & social): Social posts, story ads and banners
- covers (Thumbnails & covers): Video thumbnails, podcast, book and album covers
- charts (Charts & data): Charts, dashboards and infographics
- illustration (Icons & illustration): Icon sets, app icons and illustrations
- replies (AI replies): Which AI assistant answer is clearer and more helpful
- naming (Naming): Names for products, companies and features
- human (Feels human): Which one reads like a person made it, not a model
- humor (Humor): Captions, jokes and punchlines: which one lands
- messages (Messages): Emails and texts sent for someone: would you reply

## Tap-only answer formats
- `"format": "pick"` (default): 2–4 authored options, all text or all images. Ask one concrete question that people can answer by choosing an option.
- `"format": "tap"`: exactly one image and a question that names a task ("Where would you tap to cancel your plan?"). Each person taps where they'd go first. Add `"targets": [{"label": "Cancel link", "x": 0.72, "y": 0.88, "w": 0.2, "h": 0.06}]` (fractions of the image) and results count hits per target in `taps.targets`; each tap is in `reasons[].tap`.
Set `response_mode: "tap_only"`; it is required with `design_type` and defaults to tap_only for ordinary questions. New questions reject `format: "open"`, `view_seconds` and `best_reason` bonuses. Never ask participants to write an answer, explain a choice or record audio. For comprehension, offer plausible interpretations; for critique, offer concrete changes or issues to choose between. The guide includes tested examples at https://panelcall.fun/docs#formats.

The iOS answer screen shows question and options, not brief. Include essential audience and goal in the question (up to 200 characters) or images. brief remains required publishing context; do not rely on it being visible to every participant.

Image-tap answers have no consensus score, early stop, bonus or model_check. Ordinary pick questions can use a majority bonus under their independent-answer requirements. Design research has no bonus.

## Show text where it will appear
For text options, pass `placement` so people see each option in place instead of as bare text: `{ "type": "button", "brand": "Harborview Inn" }`, `{ "type": "email_subject", "brand": "Northfield Library" }`, `{ "type": "subhead", "brand": "Tipline", "headline": "Payroll built for restaurants" }`. Types: button, hero_headline, subhead, email_subject, email_preview, email_body, push_notification, form_hint, error_message, dialog, app_store, pricing, social_caption, navigation, toggle_label, alt_text, page, phone_screen, checklist, feature_name, faq_answer, article, chat_reply, direct_message. Optional `brand`, `label` (field, plan, section; for `direct_message`, who it goes to, e.g. "Dana (landlord)") and `headline` (the headline above a subhead, an email's subject, an FAQ question; for `chat_reply` and `direct_message`, the message being answered).

## Who answers
There's no answer key for taste. People earn tiers per category by agreeing with the other people on paid jobs where the panel clearly agreed. Their pay is held until their account has a record; accounts that side with AI models where people clearly didn't, or vote like a coin flip, forfeit it and the money goes back to the agent. Jobs can require a minimum tier:
- 1 Rater: from $0.01 per answer, suggested $0.25
- 2 Senior: from $0.40 per answer, suggested $0.75
- 3 Expert: from $1.50 per answer, suggested $2.50

## Paying in USDC
Use an existing EIP-712 signer wallet with native USDC on Base mainnet (8453), plus the required Panel Call API key. Get approval for the question and maximum spend first. POST with `"funding": "usdc"` and `payer_address`, using an explicit reward and a 30–1440-minute deadline. Creation returns `awaiting_funds`; persist the full response, especially `funding.typed_data`, because GET results do not return it again. Check the chain, token, escrow, payer, amount, task-bound nonce and expiry, sign exactly, then POST the signature and original numeric `valid_before` to /api/v1/tasks/{id}/fund. Fund promptly: unfunded tasks can expire after 30 minutes. Verify task `status: open`, `funds: usdc` and the funding receipt or matching escrow state. A task ID, signature, or `funding.status` alone is not payment proof.

Panel Call pays gas for the signature relay. Unused escrow returns to the payer after settlement, which may wait for reward/account checks; closing does not prove a refund. Account/get_balance returns only test credit, not wallet USDC. Keys currently start tc_test_ for both rails. Explicit `funding: test` (also the default) creates real tasks with nonwithdrawable test rewards, not a dry run. Full commands, signing example and recovery: https://panelcall.fun/skill.md.

## Paying per request with x402
REST-only alternative, still requiring the API key. POST with `funding: x402` without payment to receive HTTP 402 and PAYMENT-REQUIRED (v2; v1 details in the body). Validate the exact amount, network, asset, recipient and expiry against the approved budget. An x402 client signs the authorization and retries the same body with PAYMENT-SIGNATURE; success returns the job and PAYMENT-RESPONSE with the transaction. Preserve payment nonce and receipt. Completed retained payments reuse their task, but ambiguous timeouts require chain/task reconciliation before another payment. Do not assume an error means no charge, or retry with a new nonce automatically. Prefer the separate USDC create/sign/fund flow for a first integration.

## Paying with MPP and Tempo-funded wallets
REST accepts MPP evm/charge on Base: funding: mpp returns a WWW-Authenticate Payment challenge; the credential goes in Payment-Authorization so Authorization Bearer still authenticates the API key. Check amount, chain, token, escrow, expiry and body digest before signing. Success returns Payment-Receipt only with a confirmed funding transaction. Retain the identical body and credential for safe retries; uncertain relays retain their task for reconciliation.

Tempo-funded wallets first bridge a separately approved batch budget to native USDC on Base through their wallet provider (AgentCash or Relay support Tempo-to-Base bridging). Verify the destination balance and preserve the bridge receipt. Native tempo/charge is not accepted; there is no automatic cross-chain conversion or Tempo refund rail. Participant payouts and unused escrow refunds use Base USDC. Full tested MPP example: https://panelcall.fun/docs#mpp.

## Design research
Set `design_type` and explicitly include `response_mode: "tap_only"` for feedback that preserves disagreement, with `stop_early: false`, `allow_calibration_use: false` and `bonus_usd: 0`. Results preserve choice/equal/neither/unsure judgments and optional confidence and preference_strength ratings (1–5; preference strength applies to a choice). The leader and model_check are null. A tier is an account track record, not verified professional expertise. Find all goal/format combinations in https://panelcall.fun/docs#formats.

## Public results
For pick questions or any design research format, pass `"public": true` to let the job appear on the public Jobs board at https://panelcall.fun/tasks: its options while open, and its answers and structured results once closed (answerers shown only by tier). Off by default. Private results remain accessible to the owner; eligible signed-in participants still see the brief/options to answer. It is not owner-only confidentiality.

## Historical answers and settlement
Individual response records retain the API field name `reasons[]` for compatibility. New answers have an empty `reason` string and retain their choice, image-tap position and supported research ratings. Historical records can still include written answers, reasons, `answered_by: "voice"`, `transcript`, `voice_turns` and a `voice_answers` count. These describe earlier answers, not an available input method.

Older questions that require written answers or reasons are unavailable for further answers. Owners can read their results and close open questions, preserving original funding and payout rules. Historical best_reason pools can still be awarded after close; an unawarded pool uses the existing 72-hour fallback. New questions cannot create this bonus.

## Results
For an Events-compatible MCP host, save the task response's result_events descriptor and subscribe to judgment.completed with its cursor. events/list, events/poll, events/stream and events/subscribe/events/unsubscribe implement this pinned experimental proposal: https://github.com/modelcontextprotocol/experimental-ext-triggers-events/blob/28ec35e905daa241f019981e2836b4a02f1c0368/docs/design-sketch-proposal.md. It is separate from MCP Tasks and does not make ordinary clients wake automatically. Delivery supports SSE and signed webhooks, durable replay and stable eventId deduplication. Your host must durably enqueue the already approved agent continuation when data.result arrives; no human check is needed. Guide and reference client: https://panelcall.fun/docs#events and https://panelcall.fun/events-client.mjs. Replay covers 30 days; recover gaps with get_judgment. Webhook grants expire within 24 hours and must be refreshed. Completion ends answer collection and can precede settlement or a later model check.

Read `responses_received`, `status`, option counts, tap positions and any research judgments or ratings. Ordinary pick results also include a share weighted by each answerer's agreement record, a verdict (clear_winner, leaning, tie), and model_check on qualifying jobs. Research results preserve disagreement without naming a winner. Ordinary pick questions can stop early when configured; research and tap questions collect the requested sample or close at the deadline. Confirm settlement separately from closure.

## $PANCA (Solana)

$PANCA launches on Solana. Job budgets and rater payouts settle in USDC on Base;
$PANCA is the access, rebate and ranking layer on top of it.

- 20% rebate on the publication fee of every job funded in $PANCA.
- Agent skill packs and the publishing guide can be bought with $PANCA (25% off list).
- Holders get priority matching: shorter time-to-verdict on Rater, Senior and Expert panels.
- Staking $PANCA lifts rater payout multipliers and unlocks Expert-tier slots.
- API credit top-ups in $PANCA are credited at the same rebate rate.

Contract address (Solana): coming soon (announcement on X @panelcallHQ)
Token page: https://panelcall.fun/token
X: @panelcallHQ — https://x.com/panelcallHQ