Media and image generation

I wanted any agent in my estate to be able to make images, video and narration on my own GPUs, without anyone hand-starting a ComfyUI server or copying files between machines. The hard parts are around the rendering: a GPU is shared with other work, model files are large, a render can finish while its delivery fails, and an agent that retries blindly burns GPU time on work it already did. So these 33 tools treat every step as a durable operation with its own identifier. Starting a service goes through the compute budget. Model downloads are planned before they run and checked after. A render and its delivery are tracked separately, so a failed copy is repeated without rendering again. The service runs as a controller on each media host and the hub forwards these tools to it, as described on the FAFOHive page. Narration uses the same voice service as Helix.

Hosts and services

media_list_hosts lists the media machines, their GPUs and whether they are reachable.

media_list_service_profiles lists the prepared ComfyUI service configurations on a host, with their mounts and GPU compatibility.

media_start_service starts a managed ComfyUI instance on a chosen GPU from one of those profiles, after the compute budget admits it.

media_stop_service stops one instance, either letting its jobs drain or cancelling them. Installed models and files stay.

media_restart_service restarts one instance with the same drain or cancel choice and a fresh budget admission, without touching other GPU work.

media_service_status reports an instance's health, GPU, queue and errors, which separates a start request from a service that is ready.

media_service_logs reads a page of an instance's logs, with secrets redacted and a cursor for the next page.

media_operation_status returns the stored outcome of any media operation by its identifier. It is how an agent resolves an uncertain start, transfer or narration.

Models

comfy_list_models lists model and LoRA files, showing separately whether each is present, intact and visible to the loaders.

comfy_get_model shows one model's origin, installed locations, dependencies and what the loaders report about it, which explains why a file that exists can still be unusable.

comfy_inspect_model_source examines named files in a Hugging Face repository at an exact revision before anything is downloaded.

comfy_check_model_access tests whether the configured Hugging Face access can reach those files, separating a working login from repository access and licence consent.

comfy_prepare_model_download builds a fixed download plan with the exact files, revision, destinations and bytes required, so the installation can be checked before it starts.

comfy_download_model carries out a prepared plan with server-held access, checks the files and registers them in the inventory.

comfy_cancel_model_download stops one download, keeping its partial progress and anything already installed.

comfy_verify_model checks the bytes of an installed model when it looks damaged.

comfy_refresh_models compares the inventory with what a running instance's loaders can see, without interrupting its work.

comfy_reconcile_models finds models that were copied into the model folders by hand and adds them to the inventory.

comfy_release_models unloads cached models from the GPU of an idle instance. The files stay on disk.

Workflows and jobs

comfy_list_nodes lists the node types a particular instance actually has installed.

comfy_get_node_schema shows one node type's inputs, outputs and loader choices, which is the wiring contract for a workflow.

comfy_validate_workflow checks a workflow graph for problems before it runs.

comfy_submit_workflow queues a workflow with explicit input bindings and an exact output destination, and returns a receipt that tracks rendering and delivery separately.

comfy_job_status shows a job's state and errors next to its delivery result, so a failure is placed in generation or in transfer.

comfy_job_events reads a job's progress events, with optional waiting and a cursor.

comfy_cancel_job cancels one job without clearing the rest of the queue.

comfy_job_outputs lists the files a job produced and where each one was delivered.

Media files and narration

media_stage_input copies an exact source file from a host into an instance's staging area and returns a reference a workflow can load.

media_deliver_output copies a finished file to an exact host and filename, checking size and hash. It repeats a delivery without repeating the render.

media_probe reads a media file's metadata with FFprobe without changing it.

media_transform applies FFmpeg operations: transcode, resize and crop, concatenate, normalize audio, burn in subtitles, or put narration under video.

media_list_voices lists the configured narration voices.

media_synthesize_narration turns text into a WAV file with a chosen voice and delivers it to an exact destination.

These tools are how I make an agent a reliable user of expensive hardware: every action is admitted, identified, resumable and checked, and the agent always knows which part of a job failed.