How many sandbox keys can I have?

Keys are issued by invitation. Ask if you need more than one, for example one per environment.

Do sandbox questions go through human screening?

No, because they never reach real members. Live questions are screened, and during early access a person reviews each one.

Why are synthetic answers so bland?

They're canned lines, so you can test your flow. Don't read anything into them.

Do rate limits differ between sandbox and live?

No. The sandbox applies the same limits, so what works in testing works live.

Is MCP rate-limited separately from REST?

They share a budget per key: a tool call spends from the same bucket as its REST endpoint. MCP also has its own cap of 120 requests a minute.

API sandbox and rate limits

Q: What happens if I hit a limit?

You get 429 rate-limited with Retry-After (over MCP, a tool error with retry_after). Wait that long and retry.

Sandbox and limits

What the sandbox does, how fast you can call it, and what is not built yet.

Last checked 5 October 2026 against the rate-limit and question settings in the code.

How the sandbox works

Keys starting pfy_test_ are sandbox keys. They are the only keys issued in early access.
Sandbox questions skip human screening, because they never reach real members. Synthetic test members answer them through the same answer path real members use, a few at a time, usually within a few minutes.
Answers carry synthetic: true, questions sandbox: true, and receipts sbx: true and chk: ["synthetic"]. Sandbox receipts are signed with a separate sandbox key (published at /v1/sandbox/jwks.json, never in the production key set), so you can test verification end to end, but the verify endpoint always says valid: false and sandbox: true for them.
Nothing is charged. Quotes show the launch example price alongside charge_minor: 0.
Every sandbox record is flagged as sandbox. Sandbox questions are never shown to real members and never affect their Standing.

Testing tips

Use pools of 5 while testing: they fill fastest.
Synthetic answers are canned lines, so do not read anything into them.
Test your error handling: a link in the prompt returns 422 url-questions-disabled; cancelling after the first answer returns 409 invalid-state.
Rate limits apply in the sandbox exactly as they will in live use.

Question limits

Limits on a question
Item	Limit
Pool size	5 to 50 people
Question text	3 to 280 characters, no links
Options (quick pick)	2 or 3, each up to 80 characters, all different
Image	PNG, JPEG or WebP. Up to 4 MB, which keeps requests under the hosting platform's body limits (about 4.5 MB of binary, 6 MB of JSON including base64)
Deadline	1 to 168 hours (optional)
Long-poll wait	Up to 25 seconds per call
Event stream	Up to 50 seconds per connection, then reconnect with `Last-Event-ID`
Idempotency key	8 to 100 characters, required to ask

Rate limits

Limits are token buckets per key, shared between REST and MCP: a burst up to the bucket size, then a steady refill. Over the limit you get 429 rate-limited with Retry-After (over MCP, a tool error with retry_after).

Per-key rate limits
What	Burst	Sustained
Quotes and receipt verification	60	60 a minute
Asking questions	10	10 a minute
Reading results, receipts and streams	120	120 a minute
Image uploads	10	120 an hour
Opening event streams	5	About 5 open at once
Long-polls (`wait_seconds` above 0)	5	12 a minute, so about 5 waiting at once
Ratings, cancelling, balance, usage and question types	120	120 a minute
MCP requests (any method)	120	120 a minute

Not available yet

Live keys (pfy_live_), payments and credit packs.
Webhooks. Use long-polling or the event stream for now.
URL questions (a page or link shown to members). Upload a screenshot instead.
Short answer, considered feedback and expert view types.
OAuth sign-in for MCP clients. Use an API key header.
Keyless pay-per-call (x402). Every call needs an operator key.
SDKs. Generate a client from the OpenAPI document.
An uptime commitment. Design for retries: send an Idempotency-Key and resume streams with Last-Event-ID.