> **Best if you want one provider for image edits and reference-conditioned short video.** Grok covers both image generation/editing and video generation under one key.
**Tools unlocked:** `grok_image`, `grok_video`
**Env var:** `XAI_API_KEY`
#### Setup
1. Create an xAI developer account
2. Generate an API key in the xAI developer console
3. Add to `.env`: `XAI_API_KEY=xai-...`
#### What it's best for
- Image editing and style transfer
- Multi-image composites into one generated frame
- Short reference-image videos where a person, garment, or product must carry into motion
#### Pricing
Current xAI docs pricing for the Grok media models:
| Model | Price |
|------|-------|
| `grok-imagine-image` | $0.02 per generated image |
> **Direct ByteDance API via V4 signing.** Calls the Volcengine visual API (visual.volcengineapi.com) with HMAC-SHA256 request signing using IAM AK/SK credentials. Supports text-to-video and image-to-video via Jimeng 3.0 Pro.
1. Go to [console.volcengine.com/iam/keymanage](https://console.volcengine.com/iam/keymanage)
2. Create a Volcengine account if you don't have one
3. Create an Access Key pair (AK + SK)
4. Ensure your account has access to Jimeng AI (即梦) video generation service
5. Add to `.env`: `VOLC_ACCESSKEY=...` and `VOLC_SECRETKEY=...`
#### What it's best for
- Direct ByteDance/Volcengine API quota usage
- Jimeng 3.0 Pro text-to-video and image-to-video
- Chinese-language prompt understanding
- Configurable frame count (121=5s, 241=10s) and aspect ratio
#### API notes
Authentication uses Volcengine IAM V4 signing (HMAC-SHA256), not a Bearer token. The signing process builds a canonical request, derives a signing key from SK → date → region → service, and signs the request.
The implementation uses the compatible generic `CVSync2Async*` route (API version `2022-08-31`) rather than the model-specific `2024-06-06` actions presented in the public API explorer. This is intentional — the generic route supports the same Jimeng 3.0 Pro model via `req_key` while remaining stable across model updates.
**Authoritative API reference:** [Jimeng TI2V V30 Pro SubmitTask](https://api.volcengine.com/api-docs/view?action=JimengTI2VV30PROSubmitTask&serviceCode=cv&version=2024-06-06)
**Schema constraints** (enforced by `input_schema` to prevent paid-call failures):
-`prompt`: max 800 characters
-`frames`: must be exactly `121` (5s) or `241` (10s) at 24fps
-`seed`: `-1` for random, or any non-negative integer
### Volcengine Ark — Direct Seedance 2.0 Video Generation
> **Official direct Seedance API.** Calls Volcengine Ark without routing through fal.ai or Replicate, while keeping those existing provider paths available as independent fallbacks.
**Tool unlocked:** `seedance_ark`
**Env var:** `ARK_API_KEY`
#### Setup
1. Open the [Volcengine Ark API key console](https://console.volcengine.com/ark/region:cn-beijing/apiKey)
2. Enable the Seedance 2.0 model family and confirm that the account has balance or a valid resource package
3. Create a long-lived API key
4. Add the key body to `.env`: `ARK_API_KEY=...`
Do not include the `Bearer ` prefix in the environment value. The tool adds the authorization scheme when it sends a request.
| Standard | `doubao-seedance-2-0-260128` | 480p, 720p, 1080p, or 4K |
| Fast | `doubao-seedance-2-0-fast-260128` | 480p or 720p |
| Mini | `doubao-seedance-2-0-mini-260615` | 480p or 720p |
The adapter supports:
- text-to-video, first-frame image-to-video, and multimodal reference-to-video
- local image and audio inputs encoded as validated Data URIs
- remote reference image, video, and audio URLs
- task create, query, cancel, and bounded polling
- synchronized audio, optional last-frame return, web search for text-only requests, and output download
- pre-submit dry-run and token-based cost estimates
Local reference videos are intentionally rejected because the public API does not document video Data URI support. Use a provider-accessible HTTPS URL or an Ark asset reference instead.
#### API and billing notes
The asynchronous API flow is:
`POST /contents/generations/tasks` → `GET /contents/generations/tasks/{id}` → download the successful result URL.
Queued tasks can be cancelled with `DELETE /contents/generations/tasks/{id}`. Task records are retained for a limited period, and successful result URLs are short-lived, so the tool downloads outputs promptly.
Ark bills Seedance 2.0 by completion tokens. Rates vary by model, resolution, and whether the request includes reference video. OpenMontage estimates cost before submission and reconciles against provider-returned usage when available. Check the Ark console for current rates before a paid run; custom endpoint IDs require an explicit custom price so unknown pricing is never treated as free.
Official references: [model list](https://www.volcengine.com/docs/82379/1330310?lang=zh), [create task](https://www.volcengine.com/docs/82379/1520757?lang=zh), [query task](https://www.volcengine.com/docs/82379/1521309?lang=zh).
> **Best for Chinese-language production.** One key unlocks Qwen-Image generation, Qwen-TTS Mandarin narration, and Qwen-ASR with word-level timestamps — the only DashScope path that provides word-level granularity for subtitle alignment.
- Word-level timestamp transcription for subtitle alignment (Qwen-ASR filetrans)
- Replacing the broken `whisperx` slot for ASR
#### API notes
DashScope's `/compatible-mode/v1/` only supports `/chat/completions` and `/embeddings`. Image gen, TTS, and ASR all use DashScope-native endpoints with nested `{model, input, parameters}` request shape — not OpenAI-compatible paths.
The ASR tool (`qwen3-asr-flash-filetrans`) uses an async submit-poll pattern. Audio must be at a publicly accessible URL (local files are not supported). Word timestamps are in milliseconds, normalized to seconds by the tool.
#### Pricing
| Model | Price |
|------|-------|
| `qwen-image-2.0-pro` | ~$0.02 per image (check console for current rates) |
> **Official Kling path.** This is separate from `kling_video` via fal.ai: it uses Kling's official `Authorization: Bearer <KLING_API_KEY>` API, provider name `kling_official`, and direct Classic/Turbo/Omni task protocols.
- Direct official Kling API provenance rather than fal.ai gateway routing
- Text-to-video, image-to-video, and deep Video Omni reference workflows via `kling_official_video`
- Text-to-image, image edit/reference, and Image Omni multi-reference or series workflows via `kling_official_image`
- Text-to-speech via `kling_tts` when you already know the official Kling `voice_id`
- Cloud avatar presenter clips via `kling_avatar`, without replacing local `talking_head`
- Cloud lip-sync via `kling_lip_sync`, with explicit face selection for multi-person videos
- Accounts that need to use official Kling model permissions, resource packs, or regional endpoints
#### Notes
-`provider="kling_official"` is intentionally different from fal.ai's `provider="kling"`.
- Official Kling is a paid remote API. OpenMontage uses conservative cost estimates and includes high-cost factors such as Omni references, series output, 4k mode, and native sound.
- Local image paths are sent as raw base64 for supported Classic/image-generation fields. Turbo image-to-video requires a URL and will not silently upload through fal.ai.
- Video Omni and Image Omni can pass official `element_id` references through `element_list`; Elements remain an internal Kling Official helper, not a standalone OpenMontage capability.
- Account Usage is available as a low-frequency diagnostic helper under `tools/_kling/account.py`; it is not a selector or pipeline tool.
-`callback_url` is passed through and recorded when supplied, but OpenMontage still polls tasks by default.
-`kling_tts` requires an explicit `voice_id`; OpenMontage does not guess a default official voice.
-`kling_avatar` and `kling_lip_sync` register under the existing `avatar` capability and coexist with local SadTalker/Wav2Lip tools. Current avatar pipelines must opt into them explicitly; registry discovery alone does not replace local tools.
- Official Kling audio effects and video effects are documented but intentionally not registered as OpenMontage tools yet, because current pipelines do not have a stable sound-effects or video-effects capability slot for them.
| Scale | $330/mo | 2,000,000 | Priority support |
**Free tier:** 10,000 characters/month (roughly 2-3 minutes of narration). API access included. Music generation and sound effects also available on free tier with limited credits.
> **High-emotion narration and reusable cloned voices.** S2-generation models support inline emotion tags (`[laugh]`, `[whispers]`) and 80+ languages. Voices created in the fish.audio playground are reused across runs via `reference_id`.
**Tools unlocked:** `fish_audio_tts`
**Env var:** `FISH_AUDIO_API_KEY`
#### Setup
1. Sign up at [fish.audio](https://fish.audio)
2. Create an API key at [fish.audio/go-api/api-keys](https://fish.audio/go-api/api-keys/)
3. Add to `.env`: `FISH_AUDIO_API_KEY=your-key-here`
4. (Optional) Build or pick a voice model in the fish.audio playground and pass its id as `reference_id` to reuse a cloned voice
#### Backend models
`model` is **required — there is no default**. Pass one of:
**`s2.1-pro-free` caveats — promotional, not a durable free tier.** Per the [fish.audio announcement](https://fish.audio/ko/blog/s2-1-pro-free-api/?articleLocale=en), free API access runs **through August 31, 2026** and is subject to Fair Use limits, carries **no SLA or latency guarantee**, requests **may be retained** by fish.audio, and **commercial use is restricted**. Don't route client work or production narration through it, and don't plan long-term costs at $0 — `fish_audio_tts.estimate_cost()` falls back to the paid `s2.1-pro` rate after the promotional window ends.
Billing is **per UTF-8 byte of input text** (not per character) — CJK text and emoji cost 3-4x an ASCII character of the same visible length. Current list pricing: `s1` / `s2-pro` / `s2.1-pro` = $15 per 1M bytes; `s2.1-pro-free` is $0 only during the promotional window (through August 31, 2026 — see caveats above). Verify current pricing in the [official pricing guide](https://docs.fish.audio/developer-guide/models-pricing/pricing-and-rate-limits) before large batches.
> **Strong Mandarin narration.** Volcengine Doubao Speech is a good choice for Chinese explainer voiceovers and long-form narration that needs subtitle timing metadata.
Do not pass a new-console API Key as `X-Api-App-Id` or `X-Api-Access-Key`. That mismatch can produce `load grant: requested grant not found`.
#### What It Is Best For
- Natural Mandarin narration for Chinese-language explainers
- Async long-form narration via `/api/v3/tts/submit` and `/api/v3/tts/query`
- Character-level timing metadata for subtitle alignment
- Calm educational pacing where the video duration can follow the approved voice rhythm
#### Pacing
Start with `speech_rate: 0` for natural Mandarin delivery. If the approved format needs a tighter runtime, compare short samples at `speech_rate: 25` or `50` before generating the full narration. Do not force Doubao to match another provider's duration unless the user explicitly wants that tradeoff.
#### Pricing
Doubao Speech 2.0 is billed by character package or usage in Volcengine. OpenMontage estimates cost from text length and prefers provider-returned usage metadata when available.
Resolution options: **720p** (default) or **1080p**.
A watermark (`logo_add`) is added by default. Set `logo_add: 0` to disable it (requires console approval from Tencent).
**Schema constraints:**
- **Prompt:** max 200 UTF-8 characters
- **Image:** max 10MB, 50–5000 px per side, aspect ratio 1:4 to 4:1
- **Formats:** jpg, png, jpeg, webp, bmp, tiff
#### Fallback Tools
If `hunyuan_cloud_video` returns an error, the agent may retry with: `jimeng_video`, `kling_official_video`, `minimax_video`
#### Pricing
Tencent TokenHub uses a credit-based pricing system (1 credit = 1.2 RMB ≈ $0.167 USD):
| Model | Resolution | Credits | Estimated USD |
|-------|-----------|---------|---------------|
| HY-Video-1.5 | any | 1.5 | ~$0.25 |
| YT-Video-2.0 | 480p | 2 | ~$0.33 |
| YT-Video-2.0 | 720p / 1080p | 5 | ~$0.83 |
> **Free tier:** Tencent occasionally offers new-user credits for TokenHub. Check the [TokenHub console](https://console.cloud.tencent.com/tokenhub) for current promotions.
> **Cloud transcription.** Azure AI Speech Fast Transcription turns local audio into text with word-level timestamps, speaker diarization, and multi-language identification — no GPU required. Optional: the local faster-whisper `transcriber` remains the default offline STT path. When `AZURE_SPEECH_KEY` is set, the agent prefers `azure_stt` for cloud transcription.
> **Cloud neural narration.** Azure neural TTS delivers high-quality multilingual voices with SSML prosody control and express-as styles — same Speech resource as `azure_stt`, so one key/region unlocks both directions. Optional: the local `piper_tts` remains the default offline TTS path. When `AZURE_SPEECH_KEY` is set, the agent may prefer `azure_tts` for cloud narration.
> **One key, five tools.** Google Cloud TTS has 700+ voices in 50+ languages — the strongest localization option. `google_imagen` supports both Imagen 4 and Gemini 2.5 Flash Image, including projects without Imagen catalog access. Google Lyria generates high-quality background music. Gemini Omni Flash supports conversational video editing, and direct Veo generation covers premium short video clips.
| **Studio** | — | $24.00 | Professional studio voices |
| **Chirp** | — | $4.00 | Conversational style |
The free tiers apply *independently* — you get 1M Standard AND 1M WaveNet AND 1M Neural2 characters per month free. That's roughly 250+ minutes of narration per month at zero cost.
| `gemini-omni-flash-preview` | ~$0.10 per second of video | Billed as 5,792 output tokens/sec of 720p video at $17.50/1M tokens |
Generates 3–10 second clips at 720p/24fps with synthesized audio, plus stateful conversational editing (`edit_video` via `previous_interaction_id`). **Paid tier only — no free tier.** A typical 8-second clip costs ~$0.80; each edit turn generates a new clip and bills again.
| `lyria-3-pro-preview` | $0.08 (flat rate, up to 184s duration) |
**Free tier for Music:** None. Paid tier only.
**New account bonus:** Google Cloud offers **$300 in free credits** for new accounts (90-day trial), applicable to TTS, Imagen, Music, Gemini Omni video, and direct Veo video.
> **Note:** DALL-E 2/3 were shut down by OpenAI on 2026-05-12, and the `gpt-image-1` family (`gpt-image-1-mini`, `gpt-image-1.5`) retires 2026-12-01 — `gpt-image-2` is OpenAI's recommended replacement ([deprecations](https://developers.openai.com/api/docs/deprecations)).
> **Highest-rated AI video quality.** #1 on Elo rankings. Professional-grade video generation with Gen-3 Alpha Turbo, Gen-4 Turbo, and Gen-4 Aleph models.
> **Multi-model video platform.** Routes to Kling 3.0, Veo 3.1, Sora 2, WAN 2.5, and proprietary Soul Cinema through a single API. Includes Soul ID for character consistency across clips.
> **React-based video rendering.** Turns still images into animated video with spring physics, animated text cards, stat cards, charts, and transitions. **This is the key fallback when no video generation providers are configured** — the agent generates images and Remotion animates them into professional-looking video.
**Tool:** `video_compose` (with `operation="render"` — auto-routes to Remotion when needed)
**Runtime:** CPU (Node.js required)
**Env var:** None
#### Setup
```bash
# Included in make setup, or install manually:
cd remotion-composer && npm install && cd ..
```
Requires **Node.js 18+** and `npx`. The `remotion-composer/` project is included in the repo.
#### What Remotion Renders
| Component | What it produces |
|-----------|-----------------|
| **TextCard** | Animated title/body text with spring physics entrance |
| **StatCard** | Animated statistics with count-up animations |
> **GSAP-native local rendering.** HyperFrames is the preferred runtime for motion-graphics-heavy HTML compositions and the `character-animation` pipeline's rigged SVG character acting.
**Tool:** `hyperframes_compose` directly, or `video_compose` with `edit_decisions.render_runtime="hyperframes"`
**Runtime:** CPU (Node.js >= 22, FFmpeg, and `npx` required)
**Env var:** None
#### Setup
```bash
node --version
ffmpeg -version
npx --yes hyperframes doctor
```
The CLI is consumed as `npx hyperframes`. Do not use `npx @hyperframes/cli`; that package name is not the OpenMontage runtime path.
#### What HyperFrames Renders
| Use case | What it produces |
|----------|------------------|
| **Kinetic typography** | HTML/CSS text animation driven by GSAP timelines |
| **Product / launch videos** | Structured HTML scenes, registry blocks, and transitions |
| **Website-to-video** | Browser-captured site compositions with HyperFrames validation |
| **Character animation** | SVG character rigs, pose/action timelines, and GSAP acting beats rendered to `renders/final.mp4` |
HyperFrames workspaces live under `projects/<project-name>/hyperframes/`. Final videos still follow the normal OpenMontage convention: `projects/<project-name>/renders/final.mp4`.
**Available voices:** ~30 English voices plus voices for German, French, Spanish, Italian, and other languages. Lower variety than cloud providers but completely free and offline.
**Quality:** Good for drafts, internal videos, and budget projects. For client-facing narration, use ElevenLabs or Google TTS.
---
### Local Video Generation (GPU Required)
> **Free AI video generation.** Requires an NVIDIA GPU with sufficient VRAM.
A: FFmpeg + Node.js (both free, local). FFmpeg handles video assembly, audio mixing, and subtitles. With Node.js, Remotion renders still images into animated video — so even without any video generation API, the agent generates images and Remotion turns them into professional-looking video with spring animations, text cards, and transitions. Add Piper TTS for free narration and Pexels/Pixabay for free stock footage.
**Q: I don't have any video generation providers. Can I still make videos?**
A: Yes. The agent generates still images (via any image provider — even free stock from Pexels/Pixabay) and Remotion composes them into animated video with spring physics transitions, text cards, stat cards, and charts. This is the default path for explainer and animation pipelines when no video gen is configured.
**Q: What's one low-friction way to get AI-generated images and video?**
A: fal.ai (`FAL_KEY`) is one pay-as-you-go option with broad single-key coverage. It unlocks FLUX images plus multiple video providers. No subscription — pay only for what you generate.
**Q: I have a GPU. What can I run locally for free?**
A: Set `VIDEO_GEN_LOCAL_ENABLED=true` and install `diffusers`. You get WAN 2.1, Hunyuan, CogVideo, and LTX video generation plus Stable Diffusion image generation — all free, all offline.
**Q: Which TTS provider should I use?**
A: For quality → ElevenLabs. For localization (50+ languages) → Google TTS. For budget → Google free tier (1M chars/month). For offline → Piper.
**Q: Do I need all these providers?**
A: No. Start with what you have. The selector pattern auto-routes to whatever's available. Missing a provider? The system falls through to the next one automatically.
