start_async_invoke returns immediately and the remaining tools answer what happened afterwards: list_async_invokes for everything tracked with its status, get_async_invoke for one in detail. That is the right shape for video, where a blocking call would simply time out. get_prompting_guide is the unusual inclusion and probably the one most likely to change your output, because it puts the model's own prompting conventions in front of the caller instead of leaving them to be discovered through failed generations. The seed parameter is what makes a run repeatable, which matters when a prompt is being iterated rather than sampled.
An Amazon Nova Reel client with four tools: start an asynchronous generation, list tracked invocations, read one in detail, and fetch the prompting guide.
- Generation as an asynchronous job: start_async_invoke takes the prompt with a task_type, dimension, duration_seconds, fps and seed, and returns rather than blocking on the render.
- Job tracking: list_async_invokes returns every tracked invocation with its current status, and get_async_invoke returns one in detail by identifier.
- The guidance in-band: get_prompting_guide returns Nova Reel prompting best practice with examples, so a prompt can be written against the model's own guidance.
AWS credentials and an S3 bucket for the output — the server takes an access key id, a secret access key and a bucket name at startup.
One command plus a key — docker pull ghcr.io/mirecekd/novareel-mcp:latest-stdio, then supply credentials