Kling AI Video Generation
Translate a user brief into a coherent Kling motion plan and one approved remote generation request. Use only live kling-ai MCP tools and schemas at https://klingai.com/mcp.
Contract
- Use host-managed OAuth. Never request or expose API keys, tokens, cookies, authorization headers, or signed URLs.
- A user request to generate authorizes one submission after materially missing inputs are resolved. Do not add a credit-cost warning or a separate confirmation step.
- Submit once per approved intent. Never automatically retry a failed or ambiguous generation.
- Discover live tools and schemas at runtime. Do not hard-code model names, input roles, duration values, or multi-shot fields from examples.
- Upload attached media through the remote upload tool when required and preserve the exact returned reference.
Workflow
- Classify the request using the mode table below.
- Read scene patterns for the matching format.
- Read motion and shot planning for camera choreography, image-to-video constraints, multi-shot continuity, or timed narration.
- Ask only for missing creative facts that materially change the result: duration, destination ratio, required references, narration/copy, or shot structure.
- Build a motion-first prompt describing subject action, camera action, environmental motion, timing, continuity, and protected elements.
- Call the selected live generation tool once when the request has enough information. Preserve
generationId and any taskTraceId.
- If the submission is not terminal, poll its status at provider-allowed intervals until success or failure. On user cancellation or current-turn timeout, return the current state and task number.
- Provide the primary video or result link returned by Kling. Show
generationId as 任务编号 and keep taskTraceId internal unless troubleshooting requires it.
Generation modes
| User intent |
Mode |
Required interpretation |
| 文生视频 / text-to-video |
Generate |
No source image controls the opening frame. Define the opening composition from text. |
| 图生视频 / image-to-video |
Image-to-video |
One or more images control the first frame, last frame, identity/product reference, or visual reference. Assign each role explicitly. |
| 多镜头 / storyboard |
Single approved video plan |
Split timing and continuity deliberately; do not submit one task per shot unless the user explicitly approves separate tasks. |
| 查进度 / status |
Read-only |
Do not call a generation tool; query the existing task. |
For image-to-video, distinguish these roles before submission:
- first frame: lock opening composition and animate forward from it;
- last frame: define the intended destination only when the live schema supports it;
- identity/product reference: preserve subject facts without assuming the input is the first frame;
- style reference: transfer only named visual traits, not identity or composition.
Do not silently fall back from 图生视频 to 文生视频 when upload, reference count, or schema validation fails. Report the limitation and let the user revise the request.
Before calling the tool, check the selected mode, reference roles, duration,
resolution, shot structure, and protected elements internally. Do not show a
pre-submission process message unless you need the user to clarify a missing
creative requirement.
Defaults
- Use
720p, 5 seconds, and 16:9 only when the user gave no alternative and the live schema supports them.
- For image-to-video, derive composition from the source and avoid passing a ratio unless required.
- Prefer one continuous shot for a single moment. Use multi-shot only for explicit narrative progression, multiple locations/times, or a requested sequence.
- Keep the first generation focused. Do not add narration, on-screen copy, extra characters, or product claims that the user did not request.
Quality gate
Before submission, check that the subject action can fit the duration, camera instructions do not conflict, first/last frame intent is clear, reference identity/product geometry is protected, and multi-shot durations form a coherent whole. For ads and explainers, ensure every shot has a single communication job.
Failure behavior
- Authorization failure: direct the user to WorkBuddy's native MCP connection flow.
- Invalid argument/model: refresh the live schema and revise only the unsupported field.
- Insufficient credits: tell the user to recharge and stop. Do not retry automatically.
- Lost response: treat task creation as unknown and query existing tasks before any new submission.
- Provider failure: report the message and preserve IDs; never resubmit automatically.
1---2name: kling-ai-generate-video3description: Plan and generate videos with the OAuth-protected remote Kling MCP server. Use for text-to-video, image-to-video, cinematic clips, product showcases, UGC-style ads, unboxing, explainers, presenter scenes, multi-shot stories, camera-motion studies, social ads, and animating an existing image. Use when the user asks to make, animate, storyboard, or generate a video, clip, commercial, reel, short, motion scene, or image-to-video result.4---5
6# Kling AI Video Generation
7
8Translate a user brief into a coherent Kling motion plan and one approved remote generation request. Use only live `kling-ai` MCP tools and schemas at `https://klingai.com/mcp`.
9
10## Contract
11
12- Use host-managed OAuth. Never request or expose API keys, tokens, cookies, authorization headers, or signed URLs.
13- A user request to generate authorizes one submission after materially missing inputs are resolved. Do not add a credit-cost warning or a separate confirmation step.
14- Submit once per approved intent. Never automatically retry a failed or ambiguous generation.
15- Discover live tools and schemas at runtime. Do not hard-code model names, input roles, duration values, or multi-shot fields from examples.
16- Upload attached media through the remote upload tool when required and preserve the exact returned reference.
17
18## Workflow
19
201. Classify the request using the mode table below.
212. Read [scene patterns](references/scene-patterns.md) for the matching format.
223. Read [motion and shot planning](references/motion-and-shots.md) for camera choreography, image-to-video constraints, multi-shot continuity, or timed narration.
234. Ask only for missing creative facts that materially change the result: duration, destination ratio, required references, narration/copy, or shot structure.
245. Build a motion-first prompt describing subject action, camera action, environmental motion, timing, continuity, and protected elements.
256. Call the selected live generation tool once when the request has enough information. Preserve `generationId` and any `taskTraceId`.
267. If the submission is not terminal, poll its status at provider-allowed intervals until success or failure. On user cancellation or current-turn timeout, return the current state and task number.
278. Provide the primary video or result link returned by Kling. Show `generationId` as **任务编号** and keep `taskTraceId` internal unless troubleshooting requires it.
28
29## Generation modes
30
31| User intent | Mode | Required interpretation |
32| --- | --- | --- |
33| 文生视频 / text-to-video | Generate | No source image controls the opening frame. Define the opening composition from text. |
34| 图生视频 / image-to-video | Image-to-video | One or more images control the first frame, last frame, identity/product reference, or visual reference. Assign each role explicitly. |
35| 多镜头 / storyboard | Single approved video plan | Split timing and continuity deliberately; do not submit one task per shot unless the user explicitly approves separate tasks. |
36| 查进度 / status | Read-only | Do not call a generation tool; query the existing task. |
37
38For image-to-video, distinguish these roles before submission:
39
40- **first frame:** lock opening composition and animate forward from it;
41- **last frame:** define the intended destination only when the live schema supports it;
42- **identity/product reference:** preserve subject facts without assuming the input is the first frame;
43- **style reference:** transfer only named visual traits, not identity or composition.
44
45Do not silently fall back from 图生视频 to 文生视频 when upload, reference count, or schema validation fails. Report the limitation and let the user revise the request.
46
47Before calling the tool, check the selected mode, reference roles, duration,
48resolution, shot structure, and protected elements internally. Do not show a
49pre-submission process message unless you need the user to clarify a missing
50creative requirement.
51
52## Defaults
53
54- Use `720p`, `5` seconds, and `16:9` only when the user gave no alternative and the live schema supports them.
55- For image-to-video, derive composition from the source and avoid passing a ratio unless required.
56- Prefer one continuous shot for a single moment. Use multi-shot only for explicit narrative progression, multiple locations/times, or a requested sequence.
57- Keep the first generation focused. Do not add narration, on-screen copy, extra characters, or product claims that the user did not request.
58
59## Quality gate
60
61Before submission, check that the subject action can fit the duration, camera instructions do not conflict, first/last frame intent is clear, reference identity/product geometry is protected, and multi-shot durations form a coherent whole. For ads and explainers, ensure every shot has a single communication job.
62
63## Failure behavior
64
65- Authorization failure: direct the user to WorkBuddy's native MCP connection flow.
66- Invalid argument/model: refresh the live schema and revise only the unsupported field.
67- Insufficient credits: tell the user to recharge and stop. Do not retry automatically.
68- Lost response: treat task creation as unknown and query existing tasks before any new submission.
69- Provider failure: report the message and preserve IDs; never resubmit automatically.
infometa