AI Video Tools

AI Video Tools

Rule

Every AI feature in Clips goes through the agent chat unless it is the narrow media-pipeline exception below. The UI and server should not add broad shadow agents or inline chat workflows. Exception: transcription. Transcription takes audio, not prompts — the request-transcript action calls the transcription API directly. Provider priority for transcription: native first (browser Web Speech API, desktop local Whisper/macOS SFSpeech when available, and desktop Web Speech fallback on non-mac, saved via save-browser-transcript, no key required) → cloud fallback when native text is missing: Builder.io managed Gemini (via BUILDER_PRIVATE_KEY or a connected Builder account, no extra key needed) → Groq whisper-large-v3-turbo via GROQ_API_KEY (fast, ~$0.04/hr). Clips never routes recording/meeting audio to OpenAI for transcription.

Builder.io is the primary setup path: it brings managed AI credits, object storage, uploads, and transcription together. Bring-your-own-key setup belongs in the agent sidebar's API Keys & Connections panel; Clips settings can signpost that existing panel, but should not add a second key-management surface.

Why

The agent is already the user's primary interface — it has full project context, can chain tool calls, and can ask follow-up questions. Shadow LLM calls inside UI components create a second AI that doesn't know what the agent knows and can't coordinate with it. See the framework delegate-to-agent skill for the full argument.

Features and how they delegate

Include full video

Screen recordings often have thin or misleading audio ("real quick", "um this thing"). Titles, descriptions, chapters, and workflow docs (PR / SOP / ticket / email) are more accurate when the agent can see the UI, product names, and on-screen text.

• Preference key: includeFullVideoInAi on user setting clips-user-prefs.
• UI: checkbox at the top of the recording-page AI tools dropdown (copy notes Gemini-only).
• Actions: get-clips-ai-prefs, update-clips-ai-prefs.
• Gemini only: sending / understanding the full recording video requires a Gemini model (Builder Gemini or GEMINI_API_KEY). Claude and OpenAI cannot ingest the MP4/WebM. When the preference is on, the AI-request bridge prefers Builder gemini-3-5-flash for that turn.
• When on, queued clips-ai-request-* messages include instructions to attach or upload the recording to Gemini when possible, otherwise use get-recording-player-data / create-recording-agent-link and follow recommended frames / the frame API across the timeline — not transcript only.
• Default title generation also honors this: regenerate-title skips the transcript-only Gemini cleanup path and always queues the agent. Automatic post-transcription work uses one generate-metadata request so title and description are not competing application-state entries.

Do not invent a parallel "video LLM" path in the UI. Keep the preference on the shared user-prefs object and keep delegation through the agent chat.

The delegation pattern

From an action, kick work over to the agent chat in background mode so the user doesn't see a new message bubble mid-playback:

import { defineAction, sendToAgentChat } from "@agent-native/core";
import { z } from "zod";
import { getRecordingOrThrow } from "../server/lib/recordings.js";

export default defineAction({
  description:
    "Generate AI metadata for a recording. Delegates to the agent chat in the background so it can use its full toolchain.",
  schema: z.object({
    id: z.string(),
    kind: z
      .string()
      .default("title,summary")
      .describe("Comma-separated: title, summary, tags"),
  }),
  run: async ({ id, kind }) => {
    const rec = await getRecordingOrThrow(id);
    const kinds = kind.split(",").map((s) => s.trim());

    await sendToAgentChat({
      background: true,
      message: `Generate ${kinds.join(" + ")} for recording "${rec.title}" (${id}). Read the transcript via \`get-transcript --id=${id}\` and write results via \`update-recording --id=${id} --title=... --description=...\`.`,
      context: {
        recordingId: id,
        title: rec.title,
        durationMs: rec.durationMs,
        kinds,
      },
      submit: true,
    });

    return { queued: true, kinds };
  },
});

Key rules:

• background: true — the request runs in a hidden agent thread. The user's main chat is untouched.
• context — structured data the agent gets but the user doesn't see. Keep it small — ids, titles, durations. Don't dump the whole transcript; the agent can fetch it via get-transcript.
• submit: true — auto-submit. These are routine, user-approved operations.
• Never await the agent's response from an action. Fire and forget. The agent will write results back via other actions (update-recording, apply-edit), and refresh-signal will push them to the UI.

For UI-triggered AI — no wand, no sparkles, no robot icons (all three are overplayed clichés for AI). Prefer plain text with a caret (IconChevronDown) on a dropdown, or a neutral verb icon like IconBolt only if an icon is truly needed. Call the same action via useActionMutation:

const generate = useActionMutation("generate-ai-metadata");
<Button => generate.mutate({ id: rec.id, kind: "title,summary" })}>
  Suggest
  <IconChevronDown className="ml-2 h-4 w-4" />
</Button>

Media-pipeline exception

Transcription takes an audio file and returns text + segments. That's not a prompt/response LLM interaction, so it doesn't belong in the agent chat. actions/request-transcript.ts calls the transcription API directly.

Provider priority:

1. Native (highest priority). The browser's Web Speech API, desktop local Whisper/macOS SFSpeech when available, and desktop Web Speech fallback on non-mac run during recording and save results via save-browser-transcript. This gives an instant transcript with no API key when the local/browser recognizer is available. When request-transcript runs afterward, it preserves the ready native transcript and only falls back to a cloud provider if native text is missing.
2. Cloud fallback — Builder.io managed (Gemini). When a Builder account is connected (BUILDER_PRIVATE_KEY or per-user OAuth, no separate API key needed), request-transcript calls transcribeWithBuilder(), which routes audio to Builder.io's managed Gemini transcription. Returns text plus timestamped segments. Retry a failed transcript with force: true; pass regenerate: true to replace an existing ready transcript from the stored recording media. Regeneration keeps the prior ready transcript available if the new provider attempt fails.
3. Cloud fallback — Groq Whisper. GROQ_API_KEY → https://api.groq.com/openai/v1/audio/transcriptions, model whisper-large-v3-turbo. Fast (~$0.04/hour of audio) Whisper-compatible speech-to-text used when Builder is unavailable.

The transcript included by view-screen is only a bounded preview. When its previewTruncated field is true, the preview may end mid-sentence and does not show where transcription ended. Call get-recording-player-data before judging completeness or quoting the full transcript.

Desktop native transcripts merge the microphone and system-audio streams while removing overlapping duplicate speech and low-speech Whisper hallucinations. Keep system audio enabled when the meeting audio comes from another app.

Native transcript first is a hard rule, not a preference: cloud transcription is fallback-only for Clips recordings. Recording transcripts keep the native text as captured; request-transcript does not run automatic cleanup. Never hide a usable native transcript behind failed metadata work, and keep heuristic titles replaceable until the agent refinement lands. The standalone cleanup-transcript action remains available for explicit dictation, meeting, and agent workflows.

Clips never routes recording/meeting audio to OpenAI for transcription. (Groq's endpoint is OpenAI-compatible in request shape only — the audio goes to Groq, not OpenAI.)

If no native transcript exists and no cloud fallback is available (no Builder connection and no Groq key), the action writes status="failed" so the UI can show a friendly prompt.

When a transcript becomes ready, request-transcript must await the metadata handoff before its durable post-finalize worker returns. Do not leave title/summary work as an unawaited promise in that worker: serverless runtimes may freeze it immediately. A local heuristic title uses titleSource: "context" so it remains a temporary UI fallback until the generate-metadata agent request replaces it with a transcript-backed title.

If Builder transcription fails because credits are exhausted, explain that Builder.io credits/upgrade or a Groq key are the supported speech-to-text fallbacks. Generic OpenAI or Anthropic chat keys power chat, but they do not transcribe Clips recordings.

Secret registration

Builder transcription needs no app-specific key (the connected Builder.io account or BUILDER_PRIVATE_KEY carries the grant). The only transcription API key declared in server/register-secrets.ts is the optional Groq key, so it appears in the agent sidebar settings UI:

registerRequiredSecret({
  key: "GROQ_API_KEY",
  label: "Groq API Key (recommended)",
  description:
    "Fast speech-to-text fallback via Groq. Builder Gemini Flash-Lite is preferred when connected; Groq is used only when Builder/native transcription is unavailable.",
  docsUrl: "https://console.groq.com/keys",
  scope: "user",
  kind: "api-key",
  required: false,
});

It is not marked required: true — videos still upload and play without a Groq key when video storage is connected, since native transcription (and Builder when connected) already cover the common case. The API Keys & Connections panel surfaces it so the user can add the Groq fallback if they want it.

AI setup, providers, and credits

AI setup must be visible and paid-account-backed. Lead with Builder.io Connect for managed credits, object storage, uploads, and transcription. BYOK belongs in the agent sidebar's API Keys & Connections panel; template settings may signpost that panel but must not create a second credential vault.

Which provider powers what:

Any optional third-party speech provider is limited to desktop voice dictation and is never used for recording transcripts.

Use get-builder-credit-status when the user asks whether Builder.io credit limits are pausing backup transcription, transcript cleanup, summaries, or AI title generation. Treat an exhausted status as an FYI and upgrade path, not an app error: say so clearly and point the user at Builder.io credits/upgrade. Native browser/macOS capture remains the first transcript source regardless.

Bounded voice context

Dictation cleanup, clip title/cleanup, and meeting summaries should pass bounded voiceContext to the shared cleanup/transcription path whenever active app context, learned vocabulary, user notes, or AGENTS.md preferences are available. Pass a bounded summary — never a whole transcript, document, or catalog.

Live transcription during recording

The useLiveTranscription hook (from @agent-native/core/client/transcription) runs the browser's Web Speech API alongside any recording. It accumulates final transcript text in real time with no API key required. When the user stops, the client calls save-browser-transcript to persist the result. request-transcript preserves that native result and only falls back to cloud transcription when native text is missing.

A future pass could add server-side streaming transcription (Deepgram Nova-3 / AssemblyAI) via WebSocket for even higher quality real-time output — but the browser path already gives useful text from second zero.

Don't

• Don't route Clips recording/meeting audio to OpenAI for transcription. The cloud fallback is Builder.io managed Gemini → Groq Whisper only; don't add an OpenAI transcription path or import OpenAI from "openai".
• Don't import Anthropic from "@anthropic-ai/sdk" — the agent is already Claude.
• Don't build a "Clips AI" dialog that duplicates the agent chat. Use the agent chat.
• Don't render a robot, sparkle, or wand icon for AI affordances — all three are overplayed. Prefer plain text (or a neutral verb icon like IconBolt) for AI buttons.
• Don't dump entire transcripts into sendToAgentChat context. Pass the id; let the agent fetch.
• Don't await the agent's response from an action. Fire and forget; results arrive via other actions.

Related skills

• delegate-to-agent — the framework-wide rule this skill is grounded in.
• video-editing — filler-word removal writes proposed cuts into editor-draft for user review.
• recording — transcription kicks off automatically when upload completes.
• onboarding — how Builder-first setup and BYOK links are surfaced.