Iterate video prompts through MCP without runaway spend

Set a credit ceiling before the first video call

Treat the iteration loop as a bounded experiment, not an instruction to keep generating until an agent likes the result. Start by calling `list_models` and `get_credits`, then record the selected model, its displayed cost, the available balance, and the maximum number of paid attempts that a human has approved. OfflineCreator currently publishes Kling 2.6 Pro text-to-video at 42 credits, Kling 2.6 Pro Motion image-to-video at 42 credits, and Veo 3.1 Fast text-to-video at 96 credits. Those are current catalog prices, not quality rankings, and they can change.

Convert the ceiling into an explicit run count before writing prompts. At the current catalog values, an 84-credit test budget permits at most two 42-credit Kling attempts and does not permit a 96-credit Veo attempt. A 192-credit ceiling permits two 96-credit Veo attempts, but that arithmetic does not justify using them. Require a named question for each generation, such as whether the camera path is readable or whether the source image remains recognizable, and stop when the approved count is reached.

Use the narrowest MCP scope that can perform the step

The published `@offlinecreator/mcp` 0.1.2 package divides its tools among `models`, `read`, and `generate` scopes. Model discovery uses `models`; balance, generation history, status, waiting, and download use `read`; starting, uploading to, and cancelling a generation use `generate`. A planning or review connection therefore does not need generation authority. Use a models-only or read-capable connection while choosing a model, checking the balance, and reviewing prior jobs, then expose generation authority only to the controlled execution step.

Scope limits reduce what the connected agent can call; they do not create a spending budget, approve a prompt, or cancel a job automatically. Keep the maximum attempts and allowed model IDs in the human-approved run sheet. The current package lists nine MCP tools and none performs a credit top-up or Checkout action; the current Studio page presents claiming credits as a browser action. This workflow therefore must not imply that an MCP agent can charge a payment method to continue iterating.

Define cancellation before submitting any source media

Cancellation has a specific boundary in the published MCP surface. `cancel_generation` is described as cancelling a reserved generation before provider submission and refunding its credits. That makes it applicable to a reserved image-to-video generation that has not yet been submitted to the provider. It is not evidence that an already submitted or in-progress provider job can be stopped, and the page must not promise a refund for that later state. Preserve the generation ID and use cancellation only within the documented pre-submit boundary.

If a text-to-video request has already queued at the provider, continue checking that same generation ID with `get_generation` or `wait_generation` rather than starting a replacement. The reviewed package sources document no post-submit cancellation or voluntary-abandonment refund. This draft therefore treats pre-submit cancellation as verified and post-submit cancellation as an unresolved product gap. Stop issuing new generation calls while the existing job remains unresolved, and record the gap without inventing a cancellation or refund outcome.

Separate a bounded wait from a new generation

The 0.1.2 `generate` tool accepts `wait: true` and polls for up to 120 seconds. The separate `wait_generation` tool accepts a timeout from 5 to 300 seconds. Its client implementation checks the existing generation repeatedly and returns when the state is completed or failed; when the timeout expires, it fetches the latest status and returns `timedOut: true`. A timeout is therefore a polling outcome, not proof that generation failed and not permission to spend on a replacement.

Use a fixed control loop: submit once, save the generation ID, wait once for a bounded interval, then call `get_generation` or `wait_generation` again with the same ID. Never call `generate` merely because the client stopped waiting. The published client does not call cancellation in its timeout path. This distinction prevents duplicate paid jobs while allowing the agent session itself to remain bounded.

Submit limit

One generation IDCreate no replacement until the recorded job reaches a verified terminal state.

Wait limit

5 to 300 secondsChoose a bounded poll interval; a timeout reports status and does not cancel.

Terminal check

Completed or failedRetrieve or revise only after the existing job's state is known.

Change one prompt variable and attach a pass condition

Write a baseline prompt and label the single clause under test. Example baseline: “Five-second product reveal of a ceramic speaker on a dark walnut table. Slow clockwise orbit, steady camera height, warm rim light, no cuts, no text.” Iteration B changes only the camera clause to “slow dolly in.” Keep the model, source image, ratio, subject, lighting, and continuity language fixed. The purpose is not to claim that either wording performs better; it is to make the reason for another paid call inspectable.

Before submission, define a pass condition that can be checked on the file: the speaker remains recognizable, the requested move occurs once, the table geometry does not jump, and no unrequested text appears. Classify a failure as camera path, subject drift, geometry, continuity, or unwanted content. Revise only the clause linked to that failure. Do not switch both model and prompt in one attempt, because the next result would not identify which change mattered.

Choose the next guide from the unresolved decision

Return to the MCP creative-workflow directory when model choice, cloud processing, or the asset type is still undecided. Use the image-prompt iteration guide when the actual task is improving the source still rather than spending another video call. Use the aspect-ratio guide when composition fails only after fitting a placement. These are next-action handoffs: this page retains ownership of bounded video prompt changes, scoped MCP calls, wait behavior, and the verified cancellation boundary.

Keep spend, cancellation, and evidence claims narrow

This route owns a bounded MCP video-prompt loop: current video credit examples, least-authority tool scopes, one-generation-ID polling, pre-submit cancellation, single-variable revisions, and a human stop rule. It does not promise a lower total cost, a successful prompt, a maximum provider latency, a post-submit cancellation, a refund for voluntary abandonment, or consistent behavior across models. Current package and catalog facts expire on the page's quarterly research cadence and must be rechecked before publication.

Recent community coverage remains degraded and insufficient for qualification. The page therefore makes no claim about practitioner consensus, common savings, cancellation reliability, or popular prompt tactics. Detailed source failures and irrelevant-item counts remain in the research brief and ignored raw evidence rather than appearing as unsupported reader-facing facts. If the documented cancellation boundary, scope mapping, and current costs cannot be reverified, the overlap guard requires consolidation into the MCP creative-workflow directory rather than advancing this route.