AI

Send a conversation to a chat model: OpenAI-shaped messages, or a string as one user message. model: "agent:<name>" talks to one of your agents instead: it brings its own tools and keeps the conversation (conversation: reply.conversation continues it). The answer streams, so a long turn never times out. A request with no conversation whose messages end in an assistant reply the agent gave recently continues that reply's conversation, so a chat UI that resends the whole history keeps one thread (invite-only beta).

const reply = await AI.chat("Summarize: the meeting moved to Friday at 3pm.", { maxTokens: 200 });
console.log(reply.text, reply.costUsd);

Turn text into vectors: one string, or up to 2,048. dimensions shortens them, for models that allow it.

Make one image, kept as one of your account's files (its url keeps working); size is "WIDTHxHEIGHT" in an aspect ratio the model makes (its own shape unless you say).

Start one video. It takes minutes, so it runs as an Operation whose wait() resolves to the GeneratedVideo, kept as one of your account's files. Its price is quoted before it starts (usdPerSecond in AI.models({ kind: "video" })) and charged when it's done.

const job = await AI.video("Waves rolling under a lighthouse at dusk", { seconds: 8, aspectRatio: "9:16" });
const clip = await job.wait({ timeoutS: 600 });
console.log(clip.url, clip.seconds, clip.costUsd);

Write down what is said in a recording: an https URL to it, or its bytes, up to 2 hours (25 MB as bytes, 200 MB behind a URL), in MP3, WAV, M4A, MP4, OGG, WebM, AAC, FLAC or AIFF. language (ISO 639-1) skips detection. The default model transcribes a recording over 10 minutes in pieces cut at pauses, each charged, and joins them into one transcript; a failed piece fails the call and says which.

speakers (invite-only beta) tells the voices apart, on noclick/transcribe-1, which takes the whole recording in one job so each voice keeps its label from start to end (a model that cuts recordings into pieces refuses it). true labels each segment's speaker speaker_1, speaker_2, ... in the order they first speak; a list of names (["rep", "customer"]) says how many voices there are, and NoClick works out which is which from what they say, giving a voice it can't tell the next unused name in speaking order. It's priced as usdPerMinuteWithSpeakers in AI.models.

{ background: true } (invite-only beta) returns an Operation at once, whose wait() resolves to the transcript: for a recording that takes longer than one request waits (90 seconds), as one with speakers over about an hour can. A call that doesn't run in the background and passes 90 seconds rejects with RequestTimeout naming its operation, which still finishes and holds the transcript.

The headers a call to AI.baseUrl needs: your key as the bearer and, inside asUser (or with user), the NoClick-User that attributes the call to that user and holds it to the user's spend cap. Give them to an OpenAI client as defaultHeaders; without the user header the account pays.

const gpt = new OpenAI({ baseURL: AI.baseUrl, apiKey: "-", defaultHeaders: await AI.headers({ user: "user_42" }) });

The models you can call, the default of each kind first, with their prices in US dollars; agent lists the agents you (or the user you act for) can talk to as agent:<name>.