Start with a clear boundary.

Your first application

1. Create an application in the workspace and give it a name.
2. Select model capabilities and explicitly allow the tools you need.
3. Set a policy-check allowance, a per-request limit, and a daily limit.
4. Review and save your draft. Check its policy from Overview.
5. Inspect completed and blocked checks in Activity, or export the draft as JSON.

What works today

AI inference, in your browser.

The playground runs sentiment classification, sentence similarity, passage ranking, and extractive question answering with Transformers.js and WebAssembly. Each task loads its corresponding model explicitly. Cancel terminates the worker. These tasks process text on your device.

import { pipeline } from '@huggingface/transformers';

const classify = await pipeline(
  'text-classification',
  'Xenova/distilbert-base-uncased-finetuned-sst-2-english',
  { dtype: 'q8', device: 'wasm' }
);

const result = await classify('A promising new idea.');
console.log(result);

Downloads vary by selected model. Browser support, memory, and network availability determine whether loading succeeds. Outputs can be inaccurate and should be reviewed.

Permissions before spending

Every workflow check assesses its tool against the allowlist, then checks the per-request limit, remaining daily limit, and available resources. A failed check blocks the step and spends zero units.

Daily usage is counted by UTC date. A run has four predefined steps and a single active execution. Pause stops the next step; Cancel ends the run; reloading stops it and preserves completed records. Nothing runs in the background after the page closes.

Policy-check allowances are configuration values, separate from funded Resource Credits. Usage estimates apply quoted provider cost, execution overhead, service margin, and credit value; the calculator releases unused reservations and identifies extra authorization requirements.

Portable drafts, no secret fields

Exports are schema-versioned JSON containing application metadata, model selections, tool permissions, and resource limits. Imports validate their shape and create a new local draft. Activity history is not included in draft exports.

Use HTTPS for optional public metadata URLs. Never enter API secrets, recovery phrases, or private keys. Local storage is not authenticated ownership and is not an encrypted vault.

Prepare application requests

Select a capability, enter a request, and review its resource budget in the playground’s Request planner. The output is a request plan. For an inference result, use Browser AI to classify sentiment, compare meaning, retrieve passages, or extract answers on your device.