Run a capability, poll a task or fetch a result

Runs an inspected capability with your API key; a start spends balance at the inspected price. Copy next from Inspect and fill the placeholders.

• operation: start (default): input follows the inspected input_schema; unknown fields inside input are rejected. Synchronous kinds — social and web data, text models, JEV — return the result in the response. Asynchronous kinds — image, video, workflows and data:web.research (30 s to 3 min) — return {data: task, next}; repeat next, an operation: status call with task_id, until succeeded or failed. A finished research task carries its result in data.output; a failed one is not charged.
• Text models: {"reference":"model:<id>","input":{"input":"<prompt>"}} returns {object:"text.result", model, status, output_text, usage, request_id}. input.input is a string or a messages array; instructions, max_output_tokens and temperature are optional.
• JEV (model:jev-1.13, model:jev-1.13-free): {"input":{"state":…,"questions":…}} returns {id, model, answers, usage}.
• Result views: view: preview lifts the result’s main list to items (the first max_items, default 10, each element trimmed inside) with items_path and items_total, shortens long strings and marks the reply truncated with a result_ref; fields keeps only the listed paths, items[].<key> for keys of each list element. REST defaults to full; the BeatAPI MCP server defaults to preview. A status poll takes the same view.
• operation: result with request_id returns a finished result again, free, for one hour, with any view, fields or max_items.

Send a unique idempotency_key per task as a top-level field (or the Idempotency-Key header) and reuse it only to retry the same start. An unknown reference returns 404 with suggestions.