platform-cli skill

Platform CLI (ul cloud)

ul needs Python 3.11+ and comes with pip install ultralytics or pip install ultralytics-platform. ul cloud <resource> <operation> calls the Platform API through the generated SDK; ul login, ul train, ul predict, the ul cloud train|predict|export|download shortcuts (see yolo), and the other local commands delegate to the ultralytics package. This skill covers Platform operations, not model selection or hyperparameter tuning.

pip install ultralytics                   # or `pip install ultralytics-platform`
export ULTRALYTICS_API_KEY="YOUR_API_KEY" # or `ul login API_KEY`
ul cloud --help                           # every resource and operation
ul cloud training start --help            # one operation's arguments, types, and choices

Canonical command shapes

ul cloud account summary        # plan, credits, counts; `username` is your workspace
ul cloud datasets list owner=WS # owner=TEAM for a team
ul cloud projects create owner=WS project=helmets name="Helmet Detection" visibility=private
ul cloud models create body='{"owner":"WS","project":"helmets","model":"exp1","name":"Experiment 1"}'
ul cloud training start model_id=MODEL_ID gpu_type=l4 \
  train_args='{"model":"yolo26n.pt","data":"ul://WS/datasets/helmets","epochs":50}'
ul cloud models training owner=WS project=helmets model=exp1 # live status, epoch progress, metrics
ul cloud models files owner=WS project=helmets model=exp1    # short-lived weights download URL

Argument rules:

• Use the command shapes shown here without a preliminary --help call. Check operation help when a needed argument is missing or after an argument error.
• Arguments are SDK Python names as key=value (gpu_type, project_id), never --key value. A boolean key given without =value means true. Quote JSON for the shell.
• An operation takes one body= JSON object exactly when its help lists body (dict[str, Any]). These are the operations whose API request is a union of shapes: models create, deployments update, images update, lifecycle delete-trash, datasets ingest, datasets create-batch, upload signed-url, storage-integrations create/discover, and models predict/deployments predict. Every other operation takes flat fields; nested objects such as train_args, metadata, and args are still JSON values.
• Object and array values accept @file.json or @- for stdin. Multipart binaries such as the predict file field accept @path only.
• For body keys and value types, read only the operation's request schema, such as paths["/api/models"].post.requestBody, from the production contract at https://platform.ultralytics.com/openapi.json, instead of guessing or loading the whole file.

Behavior rules:

• Output is the complete API response on stdout, printed as JSON, text, or bytes according to its content type. omit=key1,key2 drops those keys from JSON at any depth when the question does not need them, such as omit=trainArgs on projects retrieve when run configurations are not compared, or omit=labels,scatterSample on models retrieve ... analysis=1 (add examples to drop the per-image rows and keep the summary, cohort metrics, and comparisons). Download operations return expiring signed URLs, not bytes: datasets export, datasets create-export, models files, and a completed exports retrieve. Preserve them verbatim and fetch them separately.
• Failures go to stderr: exit 1 for API/connection errors, 2 for argument/file errors, 130 when interrupted. Interrupting does not cancel a submitted job; use its cancel operation (models delete-training, exports delete, datasets delete-batch).
• There is no --json, --fields, --dry-run, or automatic pagination. datasets list (limit, include_samples, include_image_urls) and projects list (limit) take no offset or search; pass include_samples=false to keep dataset listings small. models list requires project=, returns the 20 most recently updated models by default and at most 100 with limit=, and takes no offset. Operations that page expose page, offset, cursor, or page_token. Follow returned continuation fields until exhausted. A limit-only listing may still be incomplete.
• Pass owner= on every command that takes it: the signed-in username for the caller's own resources, owner=TEAM for team workspaces, and the resource's own owner for public or shared resources. An omitted path owner defaults to the logged-in username but costs an extra account lookup on every call. Account, billing, trash, storage-integration, and Roboflow commands act on the credential's own account, and account summary does not list teams for API keys.
• Display names, URL slugs, database IDs, and URIs are distinct. Training data is ul://OWNER/datasets/DATASET; starting weights are a checkpoint name or ul://OWNER/PROJECT/MODEL. Carry returned IDs and slugs into the next command; retrieve missing identifiers instead of inferring them from names or URLs; dataset= takes a dataset slug, never a datasetId (a model's dataset is model.dataset.owner and model.dataset.dataset from models retrieve). Take a named dataset's or project's slug from its listing before retrieving it.

Working method

1. Resolve the requested outcome and target. Use exact supplied identifiers; otherwise list and pick one unambiguous match. For several matches, inspect distinguishing metadata and ask when the target or consequence stays ambiguous. When a named resource is not found, say so and offer the closest matches; never substitute another one. A bounded listing does not prove absence; broaden discovery or report what was searched. Look up datasets the user names in their own workspace; find new public datasets with explore search q=... type=datasets, one short keyword at a time (aerial, then UAV), narrowed with task, min_images, or max_images. Explore dataset search ranks exact and whole-word name, tag, class, and description matches first, then datasets whose images match the term by content; project search matches the term at the start of a word. Before recommending a public dataset for training, verify clone eligibility, labels, and splits. If none match or search fails, say so.
2. Read current state when it affects the change (visibility, status, existing children).
3. Execute the smallest requested change, then verify from the response. Retrieve again when the response omits needed state, the write is uncertain, or the job is asynchronous.
4. Report what actually changed, current status, warnings, and a verified resource link when one exists. Creating an entry, accepting a job, and completing it are separate outcomes.

Execute clearly requested actions without repeated confirmation. For spending, sharing, or irreversible changes, resolve ambiguity about target or consequence first. A request for advice is not authorization to act; "clean things up" is not permission for account-wide deletion. Do not invent commands, flags, or status edits to simulate operations the CLI does not expose; point the user to the Platform UI for those.

Workflows

Commands below omit the ul cloud prefix and owner=; add owner= as described above.

Goal	Commands
Inventory	projects list, datasets list, models list project=P. explore search q=... type=datasets searches public content, which is not private inventory; public models appear under type=projects. Retrieve a dataset to check task, classes, splits, and readiness.
Create	projects create project=slug name="Display", datasets create dataset=slug name="Display" task=detect. Set visibility= deliberately.
Copy	datasets clone dataset=D owner_body=DEST_OWNER, projects clone project=P owner_body=DEST_OWNER, models clone project=P model=M owner_body=DEST_OWNER project_body=DEST_PROJECT. Path fields name the source; *_body fields name the destination. A model's destination project must exist.
Rename, edit, move	<resource> update with the changed fields only; a new name also changes the URL slug, so use the returned slug afterwards. Move a model with models update project=P model=M project_id=DEST_ID as the only changed field.
Compare runs	projects retrieve project=P omit=trainArgs lists every model in the project with its status and metrics (datasets models dataset=D for runs on a dataset); large outputs can still be clipped, so never claim a complete ranking from a clipped or capped listing. Rank by the requested metric before naming a leader. Use models retrieve project=P model=M for evidence the summary lacks: configuration, dataset/version, or plots. Tabulate status, dataset/version, configuration, and requested metrics with links. Missing metrics are unknown, not zero. There is no compare command.
Analyze model	models retrieve project=P model=M analysis=1, then models find-similar-training-images project=P model=M when needed. See Model analysis and similarity for coverage and hash selection.
Weights	models files project=P model=M returns a temporary checkpoint URL. No training or export is needed.
Convert	exports create project=P model=M format=onnx, then exports retrieve project=P model=M export_id=ID for progress and the download URL. format=engine needs gpu_type. Exporting does not deploy.
Dataset versions	datasets create-export dataset=D saves an immutable numbered snapshot with a signed NDJSON URL; datasets export dataset=D (v=N for a saved version) returns a download URL; datasets restore dataset=D version=N rolls the dataset back; datasets update-export dataset=D version=N description="TEXT" updates a saved version's description.
Deploy	deployments create project=P model=M deployment=slug name="Display" region=us-central1. deployments update deployment=D body='{"action":"replace","project":"P","model":"M"}' swaps the model on the same URL; body='{"action":"stop"}' and {"action":"start"} pause and resume it. deployments health deployment=D warms it.
Usage, services	account storage, billing usage-summary, and billing transactions inspect the credential's own account usage; deployments list, retrieve, logs, and metrics inspect endpoints. See Usage and billing and Compute for filters and pagination.
Inference	models predict project=P model=M body='{"file":"@image.jpg"}' or deployments predict deployment=D body='{"file":"@image.jpg"}'. images predict image_id=ID model_id=ul://... predicts on a dataset image without saving labels.
Agents	agents save name="Display" graph='{...}' version=0 creates a canvas Agent; read agents save --help for blocks first. Use real datasets, models, deployments, and detectable classes; fix returned errors. Edit: read the current graph and version with agents list id=ID, then agents save id=ID name=N version=V graph='{...}' with the full edited graph.
Import data	datasets create dataset=D name="Display", then upload signed-url body= (assetType, assetId, filename, contentType, totalBytes) for one .zip, .tar, .tar.gz, .tgz, or .ndjson file, PUT the bytes with the returned headers, datasets ingest dataset=D body='{"sessionId":"S"}' (completes the upload; or {"sourceUrl":...}), then retrieve until ready. Queued is not imported.
Trash	projects delete project=P, datasets delete dataset=D, and models delete project=P model=M move them to 30-day trash. lifecycle trash lists it, lifecycle restore id=ID type=model undeletes, lifecycle delete-trash body='{"id":"ID","type":"model"}' purges one item, and body='{"all":true}' permanently empties all workspace trash.

Platform exports a dataset only as NDJSON that links its images; there is no YOLO-format or ZIP export. For local training, pass the file, or ul://OWNER/datasets/DATASET with ULTRALYTICS_API_KEY set, straight to model.train(data=...), which downloads the images and builds the YOLO folders; for the folders alone, run asyncio.run(convert_ndjson_to_yolo("dataset.ndjson", output_path="datasets")) after import asyncio and from ultralytics.data.converter import convert_ndjson_to_yolo.

Cloud training

1. Resolve the workspace, project, a ready dataset, starting weights, and requested settings. Check training gpu-availability before choosing gpu_type; cloud training spends credits. data must use the dataset's actual owner, which may differ from the model's.
2. models create body=... with owner, project, the model slug, and name for a new experiment. Reuse a pending entry from a failed start after checking its state.
3. training start model_id=ID gpu_type=... train_args=... with a fresh train_args holding model, data, and numeric epochs plus requested settings. Omit device. Do not copy stored worker arguments wholesale; saved data paths may be worker-local.
4. models training or models retrieve for progress. Return the run link and actual status without waiting for completion unless asked. models delete-training cancels a run and keeps its entry; never call training start on a model whose run is still active.

Recover

• 401: check the credential source without exposing secrets. ULTRALYTICS_API_KEY overrides the saved login key, so logging in again does not replace a stale environment value. If the execution environment injects credentials, use its flow. Verify with a read afterwards.
• 403: read the error for workspace access, permissions, or operation constraints; do not assume an expired login.
• Fix clear validation errors within the requested action. After a timeout or uncertain write, inspect state before retrying. Stop on unresolved inputs, permissions, or repeated failure; report partial results and remaining work instead of duplicating writes.

Resource links

Resource links use https://platform.ultralytics.com: projects /{owner}/{project}, datasets /{owner}/datasets/{dataset}, models and training /{owner}/{project}/{model}, and deployments /{owner}/deploy/{deployment} with the returned deployment slug. After status is ready, report the exact serviceUrl returned by deployments retrieve for inference; never construct a service URL.

Operation gotchas

Operation help already says what each command does. These are the side effects and constraints it omits; --help and the error text still win when they disagree.

Clone

• Project, dataset, and model clones need a different destination workspace. Cloning is not how to create a same-workspace experiment; use models create.
• Model sources need usable weights and must not be training. Project clones carry completed models with weights but not datasets, deployments, exports, or unfinished runs.

Visibility

• Creation may default to public and clones may keep the source's visibility, so set visibility= explicitly. Public projects expose their models. Public visibility is not collaborator edit access.

Update

• Arrays and metadata objects replace prior values. Send the complete array or object, including unchanged entries.
• An update that includes starred ignores every other field; send it alone.
• Starting training again on an existing model clears its prior run metrics and results. Create a new model to keep them. Fine-tuning from existing weights does not resume the optimizer or epoch count.
• capture_dataset_version=true saves a new snapshot; it does not select an existing one.

Delete, trash, restore

• lifecycle delete-trash is irreversible. Run it only when the user explicitly requests permanent deletion of the specified items; body='{"all":true}' requires an explicit request to empty all workspace trash. Otherwise, leave purging to the user in the UI.
• projects update archived=true only organizes; it frees nothing.
• Trashing a project also trashes its models and cancels their training. Both projects delete and models delete permanently delete attached deployments during the trash operation; restoring the project or model does not restore deployments. If deployments are attached (lifecycle trash id=ID type=project or type=model previews them), explain this consequence and obtain explicit authorization before trashing, unless the user's request already covers their permanent deletion. A 502 response means cleanup is incomplete; inspect deployments before reporting them deleted.
• Restore a parent before its children; independently trashed children need their own restore. Dataset version restore is separate: datasets restore dataset=D version=N replaces the current images, splits, classes, and annotations with the selected snapshot. It cannot be undone unless the current state was first versioned. When preservation matters, first run datasets create-export dataset=D. Stop active annotation before restoring, then wait for the dataset to return to ready.
• images delete and images delete-bulk are permanent and bypass trash.
• datasets create-export reuses an identical existing snapshot instead of creating a duplicate.