Playground

The playground at /playground issues real requests against the live API. There is no sandbox: a generation you run there is queued, costs credits, and appears in your usage and request logs like any other call.

The layout#

The left side is the work: pick Image, Video or Edit, choose a model, write a prompt, press Run. The result appears below it.

The right side is Settings & code: aspect, quality or resolution, duration, sound, reference images, how you authenticate — and a live code panel showing exactly what your settings mean on the wire.

Task, not endpoint#

You choose what you want to make; the endpoint follows from that plus your auth mode.

Task	My session	API key
Image	POST /api/generations/image	POST /v1/images/generate
Video	POST /api/generations/video	POST /v1/videos/generate
Edit	—	POST /v1/images/edit

Edit has no session endpoint, so selecting it switches you to API key mode automatically. The status strip at the bottom always names the exact endpoint the current combination resolves to.

Two ways to authenticate#

My session is the default. It uses the cookie you are already signed in with, and the server picks one of your active keys. No secret enters the browser. Use this to answer "does generation work at all?"

API key sends Authorization: Bearer pxsk_… to /v1/* — byte-for-byte what an external client does. Use this to answer "does this key work, and are its scopes right?"

Keys are stored as SHA-256 hashes, so a raw secret cannot be listed — it exists once, at creation. The panel shows your key prefixes for reference; paste the value you saved, or make a new key on the API page. A pasted secret is held in memory for that page only and cleared on reload.

Controls follow the model#

Options are read from the live model config, not hardcoded, so they always match what the selected model accepts:

• Aspect comes from that model's allow-list — some offer 5 ratios, others 15.
• Quality or Resolution uses the model's own control, including its labels. One model calls 1K/2K/4K "Low / Medium / High"; another exposes only 2K/4K as "2K / 4K".
• Duration is the model's supported set of seconds.
• Credits are the real per-resolution cost, and for video are multiplied by duration and adjusted for sound.

Switching model re-clamps any setting the new one does not support, so the form can never send a value the API would reject.

Costs are shown in credits, not dollars — the API bills in credits and never returns a currency amount.

The code panel#

Five tabs, all live:

• request — the exact JSON body
• response — the full body from your last run
• cURL — the same call for a terminal
• CLI — the equivalent picx command
• MCP — the same call as an MCP tool invocation

Async jobs#

Video and both session endpoints answer 202 and are polled every two seconds for up to two minutes. Stop ends the polling; it does not cancel the job, which continues server-side and will still consume credits.