Adscene Video Editor

Agentic AI video editor. One free-form `autonomous_edit` tool decomposes natural-language intents into a planned sequence of canonical edits; a deterministic allowlist (`scene_update`, `caption_compose`, `clean_audio`, `export_video`, etc.) bypasses planning for structured calls. Supports captions, viral kits, chroma key, motion tracking, B-roll, voiceover, music, multi-platform exports, and more.

OpenClaw AI Video Editor

A full agentic video editing surface. Send a prompt, get a planned, verified, executed edit — or call a known canonical action directly with structured params. Both modes flow through the same brain + safety gates as the in-product editor.

Endpoint

POST {ADSCENE_API_URL}/api/v1/misc/openclaw/v1/execute

Auth: Authorization: Bearer {ADSCENE_API_KEY}

Accepts either single-shot JSON (default) or SSE (Accept: text/event-stream or ?stream=true).

Request body:

{
  "tool": "autonomous_edit" | "<allowlisted-tool>",
  "params": { ... },
  "project_id": "optional-project-id",
  "scene": { /* optional client scene; server-side committed scene wins if newer */ }
}

Two Modes

1. autonomous_edit — free-form prompt (primary)

Pass a natural-language description in params.prompt. The brain classifies the intent, decomposes into atomic steps, plans a DAG of canonical actions, executes through the safety gates, and verifies the result. Use this when the caller doesn't already know which low-level action(s) are needed.

curl -sS -X POST "$ADSCENE_API_URL/api/v1/misc/openclaw/v1/execute" \
  -H "Authorization: Bearer $ADSCENE_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "tool": "autonomous_edit",
    "params": {
      "prompt": "Make this a TikTok-ready viral clip: vertical reframe, add bold captions, remove silences, and apply motion tracking to the speaker."
    },
    "project_id": "my-project"
  }'

Behind a single autonomous_edit call the agent can compose any of:

Read / inspect

• Inspect the timeline structure, layer properties, and current scene state
• Run computer-vision analysis on frames (object/face/scene detection)
• Query the transcript with keyword, semantic, or timestamp-window search
• Pull video intelligence (narrative peaks, speaker diarization, sentiment, pacing)
• Search the asset gallery (videos, images, audio) by type, duration, name
• Poll async job status; introspect property schemas

Structural editing

• Insert / update / replace / delete layers (video, audio, text, image, shape, group, adjustment)
• Trim, split, retime layers (slow-mo 0.5×, fast-forward 2×, freeze-frame)
• Reposition on the timeline, sequence layers, snap to transcript
• Heal timeline gaps, normalize audio, reconcile durations (pre-export safety pass)
• Multi-step undo / redo

Visual editing

• Color grading (brightness, contrast, saturation, hue, lift/gamma/gain, RGB curves)
• Procedural VFX shaders: smoke, dust, fire, explosion, lightning, snow, glitch, scanlines, grain, glassmorphism, bokeh, lava, telomere/corrosion, portal
• Chroma key (green / blue screen) with similarity, smoothness, spill controls; luma / alpha / depth masks
• Geometric clip shapes (circle, dome, star, hexagon, …)
• Crop (absolute or edge-based), 3D rotation + perspective
• Glow, shadow, inner shadow, gradient fills, text gradients
• Vertical reframe (9:16) and vertical-reframe montage
• Split screen (top/bottom, left/right, picture-in-picture)
• Branding overlays (logo / watermark) from gallery or AI-generated
• Motion / face tracking with dynamic zoom-follow

Captions & text

• Auto-generate captions from transcript
• Style captions with built-in templates or an AI director that picks/generates a custom template at runtime
• Curved text paths (circle, wave, custom SVG)
• Per-word entrance / exit animations (typewriter, slide, fade, scale, rotate, bounce, flip, swing, elastic, blur, glitch, wave, plus matching exits)
• Lottie animation playback control

Audio

• Clean audio: remove silences, breaths, filler words; word-level mute or cut
• Auto-ducking on speech detection (sidechain music vs voice)
• Mix / normalize / denoise / EQ (bass boost, vocal clarity, warm, bright)
• Sync external master audio to video (offset, mute camera audio)
• Beat / kick-drum-synced cuts (provide beat_times or bpm)
• Add SFX, generate music (mood / genre / BPM), generate voiceover (TTS or cloned voice)
• Render waveform visualizers (bars, wave, circular)

Async generation

• AI video / B-roll — duration + aspect ratio
• AI images — single or batch at timestamps
• AI music — prompt + duration + mood + genre + BPM
• AI voiceover — TTS or cloned voice library
• Auto-thumbnail extraction
• Face blur (all faces or background-only)
• Image edit (generative instruction-based)

High-level kits (each is a single canonical action that orchestrates many underlying edits)

• APPLY_VIRAL_KIT — vertical reframe + captions + silence removal + motion tracking + emphasis
• APPLY_CINEMATIC_DIRECTOR — energy analysis + dynamic zooms + cinematic color grade + mood-based camera moves
• APPLY_EMPHASIS_SYSTEM — keyword detection + scaling / glow / pulse coordinated with captions
• OPTIMIZE_PACING — filler-word + silence + low-energy segment removal for retention

Export

• EXPORT_VIDEO — render to MP4 (resolution / codec / quality tier)
• GENERATE_VIRAL_CLIPS — auto-segment short-form clips packaged as ZIP
• GENERATE_MULTI_PLATFORM — TikTok + Reels + Shorts + YouTube + Instagram aspect ratios in one pass

2. Deterministic tool allowlist — structured params, no planning

When the caller already knows the action and has structured params, use one of these tool names. The request is dispatched directly to the canonical action (skips intent decomposition / planning entirely):

Tool	Canonical action	Use when
read_scene	READ_SCENE	Inspect the current timeline
read_media	QUERY_ASSETS	List gallery assets
read_visual	READ_VISUAL	Run CV analysis on frames
query_transcript	QUERY_TRANSCRIPT	Search transcript by text or timestamp
scene_update	UPDATE_LAYER	Mutate a known layer's properties
scene_insert	CREATE_LAYER	Add a video / audio / text / image / shape layer
scene_timing	SCENE_TIMING	Trim, retime, reposition a layer
scene_mask	APPLY_MASK	Apply a chroma / luma / alpha / depth mask
chroma_key	APPLY_MASK	Convenience for green-screen / blue-screen keying
split_screen	SPLIT_SCREEN	Grid layout: top-bottom, left-right, PIP
caption_compose	GENERATE_CAPTIONS	Generate captions from transcript
media_treat	COLOR_GRADE	Apply color correction
scene_track	TRACK_MOTION	Face / object tracking with zoom-follow
clean_audio	CLEAN_AUDIO	Silence / breath / filler removal
audio_mix, audio_mixing	AUDIO_MIXING	Ducking, normalize, denoise, EQ
voiceover_add	GENERATE_VOICEOVER	Generate voiceover (text + voice_id)
music_generate	GENERATE_MUSIC	Generate background music
export_video	EXPORT_VIDEO	Render to MP4

Any unknown tool name returns 400 UNKNOWN_TOOL with a hint pointing at autonomous_edit.

Example — direct layer update, no planning round-trip:

curl -sS -X POST "$ADSCENE_API_URL/api/v1/misc/openclaw/v1/execute" \
  -H "Authorization: Bearer $ADSCENE_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "tool": "scene_update",
    "params": { "layer_id": "layer_123", "opacity": 0.5, "rotation": 15 }
  }'

Auto-export follow-up

After any mutating tool call (i.e., not read_*/query_* and not export_video itself), if the scene was actually changed and the agent did not already queue an EXPORT_VIDEO, the route fires one automatically as a second run. The final response carries videoUrl (when ready) or jobId (for polling). Read-only and conversational autonomous_edit calls do NOT trigger auto-export.

Optional input parameters (parity with the in-product editor)

Pass any of these inside params (or at the top level of the body) to drive advanced features:

• prompt — required for autonomous_edit; ignored for deterministic tools (structured params win)
• workingMemory — durable working-memory snapshot. Re-send to resume after awaiting_approval
• requirePlanApproval — true makes the agent stop after planning and emit awaiting_approval; resume with the same workingMemory + an approval prompt ("yes", "approve", "do it", …)
• attachedImages — array of base64 screenshots / reference images
• flaggedIssues — array of strings describing specific problems the user wants fixed
• captionTemplatePreset, captionTemplateMode — style preset routing for caption generation
• core_only (also accepted via ?core_only=true) — return a minimal scene shape (rendering-only, no debug fields)
• assets — additional asset descriptors to make available to the agent

Response shapes

JSON (default)

{
  "type": "success" | "partial_success",
  "tool": "<tool-name>",
  "success": true,
  "status": "completed" | "failed" | "awaiting_approval",
  "scene": { /* updated scene */ },
  "reply": "Human-readable summary of what changed",
  "videoUrl": "https://.../output.mp4",
  "jobId": "task_...",
  "viral_clips": [ /* clip metadata if generated */ ],
  "zip_url": "https://.../clips.zip",
  "activeTasks": [ /* queued background jobs */ ],
  "pendingAsyncJobs": [ /* in-flight job status */ ],
  "workflowStepsDetailed": [ /* every executed step */ ],
  "workflowSummary": { "title": "...", "summary": "..." },
  "verificationPassed": true,
  "verificationIssues": [],
  "committedToProjectScene": true,
  "processingTime": 12.3,
  "message": "Same as reply",
  "workingMemory": { /* return this in the next call to resume approval-paused runs */ }
}

Failure response (HTTP 4xx/5xx):

{ "success": false, "error": "...", "code": "UNKNOWN_TOOL" | "MISSING_TOOL" | "EXECUTION_ERROR" }

SSE (Accept: text/event-stream or ?stream=true)

The same event stream the in-product editor uses. Notable event types:

• heartbeat — every 15s, keeps the connection alive
• status — phase transitions (request_received, runtime_start, …)
• mode_select — { mode: "qa" | "action" }
• thinking, tool_call, tool_result — per-step reasoning visibility
• background_job_completed — async job finished (B-roll, viral clips, …)
• workflow_completed — main brain loop done, verification may continue
• success / partial_success — final terminal payload (same shape as JSON above)
• error — terminal failure

Async job lifecycle

Generation actions (generate_*, EXPORT_VIDEO) return immediately with a jobId in activeTasks / pendingAsyncJobs. Poll status with:

GET {ADSCENE_API_URL}/api/v1/misc/openclaw/v1/jobs/{jobId}
Authorization: Bearer {ADSCENE_API_KEY}

Response:

{
  "success": true,
  "jobId": "task_xxx",
  "status": "queued" | "processing" | "completed" | "failed",
  "progress": 0.74,
  "message": "Rendering frame 142 of 192",
  "result": { /* artifact URLs / clip metadata on completion */ },
  "error": null,
  "createdAt": "...",
  "updatedAt": "..."
}

To pull async-generated content into the timeline once jobs settle, the agent uses APPLY_PENDING internally — autonomous_edit callers don't need to manage this, but direct callers can issue an autonomous_edit prompt like "apply any pending generated content" to harvest.

Plan approval flow

If you pass requirePlanApproval: true, the agent stops after planning and the response carries status: "awaiting_approval" + a populated workingMemory. To proceed, call again with:

{
  "tool": "autonomous_edit",
  "params": {
    "prompt": "yes",
    "workingMemory": { /* the workingMemory from the previous response */ }
  }
}

Accepted approval phrases: yes, y, approve, approved, go, proceed, go ahead, do it, confirm.

Safety, verification, and limits

Every run flows through three deterministic gates (ActionPermissionGate, ArchitectureControlPlane, EditorSafetyPolicy). Destructive actions (CLEAR, mass deletes) require explicit confirmation params. Verification runs after execution and may trigger up to 2 repair loops; failures surface in verificationPassed: false + verificationIssues[]. Concurrent identical requests for the same (user, project, prompt, scene fingerprint) are deduplicated server-side.

Rate-limited per API key. Processing times vary: read-only ~1–3s, structural edits ~3–10s, async generation 30s–5min per artifact, viral-clip / multi-platform exports several minutes.

Supported formats

• Video in: MP4, MOV, WebM (HTTP/HTTPS URLs, YouTube URLs, gallery IDs)
• Image in: JPG, PNG, WebP
• Audio in: MP3, WAV, M4A, AAC (or extracted from video)
• Output: MP4 (export), ZIP (viral clips / multi-platform bundles)
• Max video length: depends on plan; soft limit ~30 min for synchronous edits, async generation handles longer
• Recommended resolution: 1080p or 4K; canvas is configurable per project

Example: end-to-end viral-clip generation

# 1) Kick off the viral-clip pipeline (auto-export follow-up queues rendering)
curl -sS -X POST "$ADSCENE_API_URL/api/v1/misc/openclaw/v1/execute" \
  -H "Authorization: Bearer $ADSCENE_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "tool": "autonomous_edit",
    "params": {
      "prompt": "Generate 5 viral clips, 15-30 seconds each, focused on the most engaging moments. Add bold captions, vertical reframe, remove silences."
    },
    "project_id": "my-project"
  }' | tee /tmp/result.json | jq -r '.jobId // .activeTasks[0].intent.job_id'

# 2) Poll job status until done
JOB_ID=$(jq -r '.jobId // .activeTasks[0].intent.job_id' /tmp/result.json)
while true; do
  STATUS=$(curl -sS "$ADSCENE_API_URL/api/v1/misc/openclaw/v1/jobs/$JOB_ID" \
    -H "Authorization: Bearer $ADSCENE_API_KEY" | jq -r '.status')
  echo "Status: $STATUS"
  [ "$STATUS" = "completed" ] || [ "$STATUS" = "failed" ] && break
  sleep 5
done

# 3) Fetch the final artifact URL(s)
curl -sS "$ADSCENE_API_URL/api/v1/misc/openclaw/v1/jobs/$JOB_ID" \
  -H "Authorization: Bearer $ADSCENE_API_KEY" | jq '.result'