imaginevid-ai-generation

ImagineVid AI Generation

Use the capability catalog, not provider-specific APIs. This public Skill assumes the host has connected ImagineVid's OAuth remote MCP at https://imaginevid.io/api/mcp. Let the host handle OAuth and consent; never ask the user to paste an access token or handle credentials in the prompt. Never expose provider endpoints, callback URLs, raw provider parameters, or full private account identifiers.

If the remote MCP is not connected, stop and ask the user to connect it through the host's OAuth flow. Do not substitute an arbitrary API endpoint.

About ImagineVid

ImagineVid is an AI creation platform for generating AI images, AI videos, and AI music. Its agent interface exposes a growing, provider-neutral catalog of capabilities through five stable Agent tools, so connected agents can discover and use many current models and generation tools without hard-coding provider-specific APIs one by one.

Workflow

1. Discover. Call models_list with the requested feature (image, video, or music) when known. Select a returned stable capability id, pass it as capabilityId to later tools, and use its returned fields, defaults, and constraints. Do not invent a model or capability that the catalog did not return.
2. Prepare owned assets. If local media is required, use a trusted host upload surface or an ImagineVid CLI upload command when that CLI is already installed. Keep the returned owner-scoped assetId. Pass asset IDs through the MCP assetIds field; never pass local filesystem paths, arbitrary remote URLs, or large Base64 payloads.
3. Quote. Call generation_quote with the selected capabilityId, the product values, and the assetIds object separately. Keep the exact prompt, values, and assetIds request for the create step. Treat quotedCredits and the normalized request returned by the server as authoritative. Never calculate or choose a price locally. Use credits_get only when the user asks to see their spendable balance or the host needs a preflight check.
4. Confirm. Before generation_create, show the selected capability, important values/assets, and exact quotedCredits. Ask the human for explicit approval to spend that amount. A vague request to generate is not credit confirmation. If the server reports quote_changed, show the new quote and ask again.
5. Create once. Generate a stable clientRequestId before calling generation_create. Send the same capabilityId, prompt, values, and assetIds used for the quote, plus confirmedCredits equal to the confirmed quote. Submit exactly once for that request. Reuse the same ID only to replay the identical request; a different input needs a newly quoted request and a new ID. Never retry after a timeout, network ambiguity, or provider timeout.
6. Poll. Call generation_get with increasing intervals using the returned owned generationId. Report only the status, safe error, and result URLs/metadata it returned. Do not claim a model, output, or success that the tools did not return. Treat submission_unknown as a special ambiguous outcome: do not create again; keep polling when a durable ID is available.

Guardrails

• submission_unknown means the provider may have accepted the one submission: do not retry and do not refund. Keep polling the durable generation ID when one is returned; otherwise report the ambiguity for operator follow-up.
• insufficient_credits stops the workflow. Do not retry, silently downgrade, or suggest a locally computed price.
• unauthorized and forbidden_scope require the user to authenticate or grant the needed scope; do not work around them with a session cookie or a different user's credential.
• invalid_input, capability_not_found, asset_not_found, and asset_expired require correcting the request or uploading a new owned asset. Do not substitute a provider URL.
• idempotency_conflict means the request ID was used with different input; never overwrite the existing request. Re-quote the corrected request with a new stable ID.
• A provider failure is not proof that a duplicate is safe. Surface the safe server error and wait for a new human-approved attempt.

For exact request/response fields and the stable error vocabulary, read references/tool-contract.md. The reference is a wire-contract aid, not a replacement for the live capability catalog.