platform-cli

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   # omit owner= for your own workspace; owner=TEAM for a team
ul cloud projects create 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 project=helmets model=exp1 # live status, epoch progress, metrics
ul cloud models files 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. 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=. Operations that page expose page, offset, cursor, or page_token. Follow returned continuation fields until exhausted. A limit-only listing may still be incomplete.
• An omitted path owner defaults to the logged-in username after one account lookup. Pass owner=TEAM for team workspaces. 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. 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=. Explore dataset search lists token-autocomplete name matches in sort order, then datasets whose images match the term by content; project search matches literal substrings. 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 the default owner.

Goal	Commands
Inventory	projects list, datasets list, models list project=P. explore search q=... type=datasets searches public content, which is not private inventory. 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 for model summaries, models list project=P, or datasets models dataset=D for runs on a dataset; models retrieve project=P model=M only for missing metrics. 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.

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.

Labels and auto-annotation

• images update body='{"labels":[...]}' replaces every label on the image, and images predict alone saves nothing. Retrieve first for a partial edit, and if the retrieve reports labelsTruncated, do not update that image's labels, because the replacement would delete the ones you did not see.
• datasets create-batch dataset=D body='{"modelId":"ul://OWNER/PROJECT/MODEL"}' auto-labels unlabeled images, persists the labels, and saves a version. On detect datasets with 1–200 classes, modelId can instead be a hosted model prompted with the class names: sam3.1 (SAM 3.1 Text), sam3, moondream, qwen, florence2, owlv2, yoloe26x, or groundingdino. Add "includeAnnotated":true to the body to also process labeled images (their existing labels are kept). datasets delete-batch cancels or dismisses a run without undoing labels already saved.
• datasets create-batch dataset=D body='{"operation":"blur"}' blurs faces in the dataset's images (or a single image when the body includes "imageId") and saves no version; run datasets create-export first when the originals matter.
• Annotation prediction needs compatible tasks and classes, or an explicit mapping: class_mapping= for images predict, classMapping inside body= for datasets create-batch. It rejects connected, depth, and more-than-three-channel datasets.

Classes, splits, task

• class_names replaces names by index; keep the order for a rename. Merging or deleting classes remaps or removes labels and shifts IDs.
• datasets redistribute-splits reshuffles the whole dataset. images update-bulk moves selected IDs, and conflict policy replace may remove target images.
• datasets adopt-images adds image references as unlabeled train images, skipping existing images; only release=false (copy), release=true (move), or class_mapping carries labels, metadata, and splits from a dataset you can edit.
• Changing task converts no annotations. Switching to or from depth needs an empty dataset.

Dataset inspection and analysis

• For duplicates or split leakage, group all image records by hash and compare IDs and splits. Storage deduplication does not prevent duplicate records. Use small pages or a downloaded export; if records are missing or clipped, report partial coverage, not zero duplicates.
• datasets retrieve dataset=D returns task, classes, splits, and counts. datasets class-stats dataset=D returns distributions and heatmaps; a set sampleSize means the stats came from a capped subset, and each histogram bin's size is its width. In objectsPerImageHistogram, bin is the lower bound: bin=0, size=6 counts images with 0–5 objects, not just zero objects. For the unlabeled-image count, use datasets images dataset=D has_label=false limit=1 and read total.
• List a dataset's images with datasets images dataset=D (there is no images list); it filters by split, has_label, has_error, class_ids, and search. For counts alone, use limit=1 and read total (included by default), not the page length. has_error means a recorded processing error. Pass include_thumbnails=false to keep inventory pages small (include_image_urls and include_labels default off); pass nextCursor as cursor while hasMore. The cursor works only with newest or oldest sorting (which includes the default); other sorts page with offset. Equal hash values mean the same stored bytes, not visual similarity.
• search=TERM matches substrings of filenames (extension optional), class names, and metadata keys, scalar values, and array entries, but not values inside nested objects. There is no field selector; retrieve matches to verify which field hit. A 32-hex term matches the content hash exactly and a 24-hex term matches the image ID, so search=HASH finds every stored copy of an image without exporting.
• Listings omit custom metadata; images retrieve image_id=ID returns it with labels. Both listing and retrieval can truncate labels: honor labelsTruncated and compare returned lengths with labelCount or retrieval's properties.annotationCount. datasets selected-images dataset=D image_ids='["ID1","ID2"]' fetches several known IDs in one call, not one images retrieve each, accepting the same include_* options as datasets images; images urls image_ids='["ID1","ID2"]' refreshes signed URLs for up to 100 IDs from one dataset. datasets export dataset=D provides NDJSON metadata and annotations for bulk aggregation; On Premise datasets cannot export.
• images find-similar-images image_id=ID returns up to 24 near neighbors from public, your own, and team datasets, excluding images the source dataset already holds and near-duplicate copies of the query. It takes only an existing image ID and needs authentication. An image without an embedding is embedded on demand, which can take minutes; a 503 means that preparation failed, so retry.
• datasets embeddings dataset=D reports analysis freshness and progress. datasets clustering dataset=D returns 2D UMAP points, not raw embeddings; pass nextOffset as offset while hasMore, including with the default page size. not_analyzed means no layout exists; use returned IDs with datasets selected-images for details. datasets create-embeddings starts analysis; datasets delete-embeddings cancels it and deletes no images.

Model analysis and similarity

• models retrieve project=P model=M analysis=1 requires authentication and returns per-image validation analysis instead of model details; null means analysis is unavailable. A completed detection run with recorded per-image results is required; a missing dataset version or manifest returns scores with traitsAvailable=false and coverage.mode="unavailable", not null analysis. Before generalizing, compare population with retained and read coverage.mode: only full covers every validation image. traitsAvailable gates comparisons (image traits versus F1); cohorts.worst and .best give whole-cohort count and metrics but only the first 100 rows as examples, whose labels come from the run's saved manifest, not the current dataset. scatterSample.rows are [f1, width, height, pixels, aspectRatio, instanceCount] points. Per-image tp/fp/fn/f1 use IoU 0.50 at a confidence threshold the run did not record, so they cannot rebuild metrics at another threshold. Check the model's reported precision/recall before diagnosing poor performance; these counts alone do not establish deployment precision or explain a particular false positive.
• models find-similar-training-images project=P model=M searches public, your own, and team datasets from the run's worst validation images, excluding images its training dataset already holds. Use credentials with access to the model's workspace. The command returns candidates without adding them to the dataset. hashes=H1,H2 must all come from analysis.cohorts.worst.examples[].hash; any other hash rejects the whole request. A run without a captured cohort returns an empty list; missing run/dataset identifiers or unavailable embeddings return errors instead.

Connected storage

• Connected datasets (cloud storage or On Premise) are indexed in place, not copied. They cannot take appended uploads, versions, clones, batch annotation, or new embedding analysis; do not start analysis to obtain a missing clustering layout.
• Disconnecting an integration leaves provider objects intact but breaks dataset access until the same account reconnects. Roboflow credentials are separate from Platform credentials.

Usage and billing

• account storage returns quotas as usage.<resource>.current/limit/percent, where limit=-1 is unlimited; details=true adds the ten largest breakdown.topItems, not a full inventory. billing usage-summary returns plan usage and credits.
• billing transactions from_=START to=END uses from_, not from, and needs full ISO datetimes such as 2026-09-01T00:00:00Z; bare dates are rejected. Timestamps are in createdAt.

Compute

• Editing a model's status starts or stops nothing.
• deployments list model=P/M filters by a project's model, not a model ID or bare name. Use the returned deployment slug for deployments retrieve, logs, and metrics. Logs page with nextPageToken as page_token; metrics range= selects the window and sparkline=true returns the compact dashboard summary instead of the full response.
• Deployment delete is permanent; there is no trash for deployments.
• Export delete cancels a running conversion or removes the finished file; the source model is untouched.

The installed CLI is the authority: ul version shows the versions, ul cloud <resource> <operation> --help lists valid arguments and choices, and error text beats any command shape or caveat in this file. For endpoint semantics, the Platform API reference and the live /openapi.json win over memory.