Skip to content
agent-media docs
IntroductionStart here: what Agent Media does and which path to follow.↵QuickstartConnect, check your account, upload references and generate.↵Core conceptsAccounts, connectors, tools, models, credits and jobs.↵Connect your agentClaude, Claude Code, ChatGPT, Codex, Cursor and MCP setup.↵Agent workflowThe sequence from request to finished media.↵Tool referenceTools, arguments and schemas from the official connector.↵Upload imagesMultiple reference images, drag and drop, 24-hour expiry.↵Choose a modelVideo modes, image models, audio, supported inputs and limits.↵Write better promptsRealism, framing and reference guidance.↵RecipesProduct images, talking heads and consistent characters.↵Account and billingPaid credits, account checks, payments and draft recovery.↵Generation pricingQuotes and model costs.↵REST APIAuthentication, endpoints and OpenAPI.↵Jobs and retriesRecover timeouts without duplicate generations.↵
Public documentation · No sign-in required
Documentation / How agents should work

Agents

How agents should work

A reliable sequence from a user request to finished media.

On this page

Discover before choosing

Read the actual tool catalog. Check account readiness with get_account, then use list_models to choose the kind, model and supported mode. Use the current public skill for detailed prompting and tool guidance.

Resolve inputs

For user images, use the existing upload panel and retrieve the returned URLs. Reference images belong in refs or first_frame according to the intended model/mode; describing an image in the prompt does not attach it.

Quote the intended request

Call quote with the same model and relevant inputs you intend to submit. Compare the quote with the account balance and stay within the user-approved budget. An upload-only request is not authorization to spend.

Submit once, then track

Submit the generation and preserve the returned job ID. Poll that same ID until completion. Keep any request identity when recovering a lost response; changing it can create a new paid run.

Timeout and retry recovery →

Return the result

Give the user the finished media using the client’s supported output display. If the run fails, report the saved job ID and the actual error. Do not claim a refund or successful output unless the service confirms it.