Playground
Fire a single video or a batch of up to 20 and watch clips land in real time. Each job streams its log over SSE; the grid fills as workers finish.
Generate
720p is the base render; 1080p/4K add an upscale step. Batch (count>1) is text-to-video only.
Live jobs
Rendered clips
Accounts
The Google account pool. One account is bound per worker (ban-safe); the pool rotates the least-busy healthy account and degrades any that error.
Pool
| Status | Active | Uses | 403s | Proxy | Session | |
|---|---|---|---|---|---|---|
| loading… | ||||||
Add account
cookie export or passwordImport from VeoUnlimited
source: bothAnalytics
Live metrics polled every 5s, plus a GitHub-style hourly heatmap of generation volume. History accumulates client-side from metric samples (and uses /analytics/heatmap if the server exposes it).
Queue depth
now 0 · peak 0Accounts by status
proxy: —Generation heatmap
API Reference
VeoForge is a direct-to-Google-Flow generation control plane: a concurrent worker pool (one Google account per worker, ban-safe), per-account residential proxy support, and API-level batch coalescing (N prompts in one call). Every endpoint except static UI assets requires Authorization: Bearer <token>.
System Architecture
How a prompt becomes an mp4 — the request lifecycle, session establishment, the worker/account topology, and the concurrency model.
1 · Request lifecycle
A generate call is persisted to a SQLite JobStore with an atomic claim, then picked up by a free worker. The worker drives a stealth Chrome tab authenticated to labs.google, mints a fresh reCAPTCHA v3 token in-page, submits to the aisandbox batch endpoint, polls, downloads the fifeUrl, optionally upscales, transcodes, and exposes the result.
fetch beats reconstructing headers. The generation request is issued from inside the real browser tab, so the cookies, User-Agent, client hints and the reCAPTCHA token are all the browser's genuine values. Reconstructing those headers from Python is exactly what Google's anti-abuse stack is built to catch — issuing the call in-page sidesteps the whole class of 403s.2 · Session establishment
Auth is cookie-first: inject the __Secure-next-auth.session-token into the labs.google origin, let NextAuth exchange it for a session, and lift the ya29 bearer. From then on every request rides the in-page fetch path.
3 · Worker & account topology
The WorkerPool owns N workers. Each worker holds exactly one Google account and its own browser session, guarded by a per-worker CircuitBreaker that trips on repeated 403s and backs off. The AccountPool hands out the least-busy healthy account and degrades any that error; every account binds its own residential proxy.
4 · Concurrency & throughput
Effective parallelism is min(VF_WORKERS, healthy accounts). On top of raw parallelism, the queue coalesces up to N prompts of the same tier/aspect into a single batchAsyncGenerateVideoText call, so one API round-trip yields many videos. Measured steady-state is around 16 concurrent generations.