Connection check and machine overview: GPUs (VRAM, load), queues, active jobs, and data folders.

List dataset folders in ai-toolkit's datasets folder, with full paths and file counts.

Use the path value as datasets.0.folder_path when creating a job.

Create a dataset (or add to an existing one) by uploading local files into ai-toolkit.

paths are files or folders on the machine running this MCP server; folders are searched recursively for images, videos, audio, and matching caption files (e.g. photo1.jpg + photo1.txt). Files are sent over the UI's API, so this also works when ai-toolkit runs on another machine.

Captions for each file in a dataset (empty string = uncaptioned), plus a count of missing ones.

Write captions. captions maps a file name (as shown by get_captions) to its caption text; existing captions for those files are replaced.

Queue ai-toolkit's built-in captioning job (Qwen3-Omni) for a dataset, like the UI's Caption button.

Only uncaptioned files are captioned unless recaption is set. prompt overrides the captioning instruction; by default an image prompt is used for image-only datasets and ai-toolkit's video/audio prompt otherwise. Uses GPU time and downloads the captioner model on first use. Track it with get_job.

List training jobs (newest first) with status, progress, and model.

Status and progress of one job. job is a job id or name.

Full training config of a job as YAML. Useful as a starting point for a new run.

The ai-toolkit UI's default job config (Flex.1 LoRA) as YAML, to edit and pass to create_job.

For other models, cloning a previous job with create_job(base_job=...) is more reliable, or see ai-toolkit's config/examples/*.yaml for per-model settings.

Create a training job in the ai-toolkit UI.

Config source (pick one): config as YAML/JSON text (full job file or just its config: body), or base_job (id/name) to clone an existing job's config. Neither = the default template. overrides sets values inside config.process[0] by dotted path, e.g. {"train.steps": 4000, "datasets.0.folder_path": "D:/datasets/me", "trigger_word": "ohwx", "network.linear": 16, "sample.samples": [{"prompt": "ohwx at the beach"}]}. gpu_ids is the GPU index (e.g. "0" or "0,1"). Set start to queue it immediately.

Edit a job that isn't running: replace its config (YAML/JSON) and/or apply dotted-path overrides inside config.process[0] (same format as create_job).

Queue a job and start its GPU queue. Resumes from the last checkpoint if one exists.

Stop a job gracefully (saves a checkpoint first). force kills the process immediately, only for jobs stuck and ignoring a normal stop.

Ask a running job to save a checkpoint or generate samples on its next step.

Last lines of a job's training log, with progress-bar noise collapsed.

Loss curve summary: first/last/min values, a smoothed trend, and evenly spaced points.

Checkpoint (.safetensors) files and sample media produced by a job, as local paths.

Return the most recent sample images from a job so you can judge training progress visually.