Kling Official Direct API

Use this skill for OpenMontage tools with provider="kling_official". This is not the fal.ai Kling gateway. Official Kling uses KLING_API_KEY, optional KLING_API_BASE_URL, and Authorization: Bearer <KLING_API_KEY>.

Provider Split

• kling_video uses fal.ai, FAL_KEY, fal.ai queue URLs, and provider="kling".
• kling_official_video uses Kling official API, KLING_API_KEY, official task protocols, and provider="kling_official".
• kling_official_image uses the same official auth and task protocol for image generation.
• kling_tts uses the official audio TTS endpoint and stays in the existing tts capability.
• kling_avatar and kling_lip_sync use official avatar/lip-sync endpoints and stay in the existing avatar capability. They do not replace local talking_head or lip_sync.

Never silently switch between these paths. If the selected provider is unavailable, surface the blocker and ask before substituting.

Auth And Endpoint

Default base URL:

https://api-singapore.klingai.com

Users may override it with KLING_API_BASE_URL, for example for a regional endpoint. All requests send JSON and:

Authorization: Bearer <KLING_API_KEY>

Task Protocols

Classic APIs:

• Create ID path: data.task_id
• Statuses: submitted, processing, succeed, failed
• Result paths: data.task_result.videos[], data.task_result.images[], data.task_result.audios[]

Turbo APIs:

• Create ID path: data.id
• Poll path: GET /tasks?task_ids=<id>
• Statuses: submitted, processing, succeeded, failed
• Result path: data[0].outputs[]

Keep the parsers separate. Do not write a fuzzy parser that guesses between task_id and id or between succeed and succeeded.

Omni References

Video Omni and Image Omni stay inside the existing provider tools through api_family="omni". Do not create selector-level Omni operations.

Video Omni accepts official reference structures:

• image_list[] with image_url and optional type such as first_frame or end_frame.
• video_list[] with video_url, official refer_type values such as feature or base, and optional keep_original_sound.
• element_list[] with official element_id values.
• Structured multi_prompt[]; do not split natural language into shots automatically.

Local image references may be normalized through tools/_kling/media.py. Local video paths must not be silently uploaded through fal.ai; ask for or require a reachable URL.

Image Omni accepts image_list[] with official image values. Prompt placeholders such as <<<image_1>>> must map stably to the provided image order. If the prompt already contains placeholders, validate that the referenced images exist and do not insert duplicates.

Capability Boundaries

TTS, avatar, and lip sync are provider additions to existing OpenMontage capabilities. Audio effects and video effects are official Kling endpoints, but they are not registered as default OpenMontage tools until a pipeline has a stable capability slot for them.

• Do not add sound_effects or video_effects capabilities from inside a provider implementation.
• Do not let video effects enter the ordinary video_generation selector path.
• Do not disguise short sound effects as long background music unless a pipeline explicitly consumes that shape and the tool's best_for / not_good_for says so.

Video Parameters

Use operation for OpenMontage semantics:

• text_to_video
• image_to_video
• reference_to_video

Use api_family for official protocol choice:

• classic
• turbo
• omni

Important constraints:

• Official video provider input schema must not expose top-level image_url; use reference_image_url or reference_image_path.
• Classic image-to-video accepts reference_image_url or a local path converted to raw base64 in official field image.
• Turbo image-to-video requires a URL first frame. Do not upload local files through fal.ai as a fallback.
• Send aspect_ratio only where the current schema supports it: Classic text-to-video, Turbo text-to-video, and Video Omni.
• Default paid path should avoid 4k, native sound, or batch behavior unless explicitly selected.

Image Parameters

Use api_family="generation" for /v1/images/generations and api_family="omni" for /v1/images/omni-image.

Generation/edit path:

• prompt is required and should stay under the official 2500 character limit.
• image_url passes through as official image.
• image_path is converted to raw base64 and sent as official image.
• image_reference can be subject or face.

Omni path:

• Put references in image_list[] using official image values.
• Use prompt placeholders such as <<<image_1>>> only when the prompt needs to bind a specific reference image.

TTS Parameters

kling_tts uses:

• text
• voice_id
• voice_language, currently zh or en
• voice_speed

Require an explicit voice_id unless an official account-specific default has been verified. Do not hard-code a made-up voice. Download every returned audio item, set data.output_path to the first local file, and include voice_id, voice_language, voice_speed, task_id, and non-zero cost_usd.

Avatar Parameters

kling_avatar uses /v1/videos/avatar/image2video and accepts:

• avatar image via URL or local path converted to raw base64
• audio_id or sound_file
• optional prompt
• mode, such as std or pro

Keep it separate from local talking_head. Pipelines that want Kling avatar output must list and choose it explicitly.

Lip Sync Parameters

kling_lip_sync has two steps:

1. POST /v1/videos/identify-face with video_id or video_url; read faces from data.face_data[]
2. POST /v1/videos/advanced-lip-sync with session_id and one face_choose[] item containing face_id, audio_id or sound_file, and the sound start/end/insert times

Local video paths must not be silently uploaded through fal.ai or any other provider. If multiple faces are returned and the user did not pass face_id or face_choose, stop and return the face list for confirmation unless auto_select_face=True was explicitly set. If auto-selecting, record the selection reason and selected face in the result/artifact.

Audio Effects And Video Effects

Official Kling audio effects (/v1/audio/text-to-audio, /v1/audio/video-to-audio) and video effects (/v1/videos/effects) are intentionally not default OpenMontage selector tools. Record the non-mapping reason in docs/tests instead of registering tools that current pipelines might misuse.

Elements Helper

Elements are an internal Kling Official helper, not a new OpenMontage tool capability. Use tools/_kling/elements.py to normalize element_list[].element_id, optionally query read-only element endpoints, and record element metadata when queried. Do not create or delete elements from the default provider path.

Account Usage Helper

Account Usage is diagnostic only. Use tools/_kling/account.py for low-frequency /account/costs checks, with local cache and throttle protection. Do not call it before every generation and do not put it in selectors or production pipeline stages.

For 1101 or 1102, surface that the account or resource pack is exhausted and include an account-usage diagnostic hint.

Callback Notes

Providers may accept callback_url, but polling remains the default execution mode.

• Classic and Omni paths pass callback_url at the top level.
• Turbo paths pass it as options.callback_url.
• Successful results should record callback_requested=true, polling_used=true, the callback_url, and task_id.
• Validate callback URLs before sending; only absolute http or https URLs should pass.

Error Handling

Surface official code, message, and request_id whenever available.

Do not retry:

• Auth failures: 1000-1004
• Balance/resource-pack exhaustion: 1101, 1102
• Permission/model access: 1103
• Parameter errors: 1200, 1201
• Safety policy: 1301

Limited retry is acceptable for:

• 1302 request too fast
• 1303 concurrency/resource-pack slot limit
• 5000, 5001, 5002 server/maintenance/backlog errors

For 1303, explain that the account hit a concurrency or resource-pack slot limit.

Cost Governance

Official Kling is a paid remote API. Provider tools must return non-zero conservative estimates from estimate_cost() and include cost_usd on successful paid results. Treat estimates as low-confidence until account usage reconciliation is implemented. High-cost Omni inputs such as multiple references, element IDs, result_type="series", mode="4k", and sound="on" must increase or flag the cost estimate.

Prompt Notes

For video, start from the universal OpenMontage video prompt skeleton: subject, subject motion, scene, spatial framing, and camera. Kling tends to respond well to clear temporal action order, camera movement verbs, and concise negative prompts. For reference workflows, state what should stay consistent from the reference and what should change.