For agents: This skill is authored in roboflow/computer-vision-skills. If your client already has it loaded as a skill (for example from a Roboflow plugin), use that copy. The skill://roboflow/<name>/SKILL.md MCP resources serve the same skills for clients that don't have them loaded.

Model journey: choose, try, deploy

Understand the outcome, choose a model, try it on the user's media, put it on the project's live endpoint when the task needs one, then improve or train. These rules are the Roboflow Agent's; follow them as written.

1. Choose the model

Choose the simplest prediction type that delivers the user's requested output, then apply its Model default. Make a concrete recommendation and proceed; ask only when a missing detail materially affects the choice.

• Check whether the workspace already has a relevant custom Model (models_list). Treat a Model the user has not named as a candidate, not an automatic default: its name or class list alone does not show that it fits the user's data. Evaluate a promising existing Model on representative input before choosing it. Reuse it directly when the user requested it, it is already established in the conversation, or representative results have validated it.
• Roboflow Instant previews are identified in workspace Model results. Use one only when the user explicitly asks for that Instant Model. Otherwise, prefer the appropriate foundation Model or a deliberately trained Model.
• When the user provides a model ID or explicitly requests a specific Model, accept it exactly as given. Do not invent model IDs.
• Explicit supported choices and existing saved models take precedence over the defaults below.

• Sizes: Nano is the cheapest and fastest starting point. Move to larger sizes when evaluation shows that the additional accuracy is worth the added cost and latency. RF-DETR and RF-DETR Segmentation come in -nano, -small, -medium, -large, -xlarge and -2xlarge. YOLO26 goes up to xlarge: yolo26{n,s,m,l,x}-640, and yolo26{n,s,m,l,x}-pose-640 for keypoints.
• YOLO26 detection: only when CPU execution, genuinely low-power hardware, or strict cost constraints make it the better fit. It still runs through Roboflow Inference.
• Boxes or outlines: Sol, Astra and Gemini provide bounding-box object detection, not segmentation masks. Keep SAM 3 and RF-DETR Segmentation for segmentation. Start counting and localization with the detection default; use segmentation when the output requires contours or pixel areas, or tests show boxes can't meet the requirement. SAM 3's weaker language understanding often needs several text-prompt attempts.
• Explicit choices: honor Sol ({type: "sol", classes}), Astra ({type: "astra", classes}), SAM 3 ({type: "sam3", classes}) and Gemini 3.8 Flash ({type: "gemini", classes}) for the endpoint instead of substituting the default.
• CLIP ({type: "clip", classes}): only when the user explicitly requests CLIP or for an existing CLIP setup.
• Universe models: universe_search with result_type: "model" lists each result's modelVersions. Use the exact model_id, including a listed older version, as {type: "model_id", modelId}.
• Auto Label is separate: its model selection has its own defaults. Don't apply these endpoint defaults to a labeling-only task, and name an endpoint by its endpoint model, never by the Auto Label one.
• Video: Gemini, Sol, and Astra video processing runs frame by frame and can use credits quickly; confirm with the user before running them on video. For ongoing video, recommend sampling frames, labeling them, and training RF-DETR.
• Briefly name the Model and explain in plain language what kind of Model it is and why it fits the workload.

2. Workspace detection default and OpenAI availability

project_deployment_get (and project_deployment_launch) return workspaceDetectionDefault and openAiDetectionAvailability.

• workspaceDetectionDefault.objectDetectionEndpoint is the endpoint model for new non-COCO object detection with no explicit model choice: GPT-6.1 Sol on free and trial workspaces, GPT-6 Astra on paid workspaces or unknown usage. workspaceDetectionDefault.objectDetection (Gemini 4 Argon, gemini-4-argon-boxes) is the default for Workflow detection steps and catalog labeling.
• If workspaceDetectionDefault is absent, use the paid default: GPT-6 Astra. Never infer the tier from user messages.
• When openAiDetectionAvailability is available or absent, follow the workspace detection default.
• A managed-key-denied or region-unavailable restriction does not establish another provider's availability: retain a non-OpenAI default only when its own access and deployment policies permit it. When no permitted VLM default is available, use the native SAM 3 fallback without another selection question. Never switch providers, credentials or global routes to evade plan, workspace-key or residency restrictions.
• If the user explicitly requests Sol or Astra:
  • managed-key-denied: explain that its managed-key access is unavailable in this workspace and offer SAM 3 or resolving that access.
  • region-unavailable: explain that OpenAI detection is unavailable in the EU deployment and offer SAM 3. Don't suggest upgrading the plan or adding a provider key, and don't configure a new Sol or Astra endpoint.
  • For either restriction, don't configure a substitute without the user's agreement or claim the requested detector ran. Leave existing configured endpoints, COCO RF-DETR, other task defaults, and Auto Label's separate selection unchanged.

3. Try a model

Ordinary model testing needs no endpoint configuration. Never set an endpoint model just to try one.

• models_infer: a trained model id or pretrained alias on one image; returns JSON predictions.
• autolabel_preview: a foundation model from autolabel_models_list on one image of a project.

4. Put it on the endpoint

Configure this optional endpoint only when the requested task needs a stable serving endpoint or active learning. Each project has one stable endpoint; changing its model keeps the inference path.

• New object-detection project outside COCO, no explicit model choice: a new project's detection default can only be read once the project exists.
  1. Call project_deployment_launch without model or Active Learning.
  2. Read the returned workspaceDetectionDefault and openAiDetectionAvailability (section 2).
  3. Call project_deployment_set_model with {type: workspaceDetectionDefault.objectDetectionEndpoint.type, classes}, the classes to detect.
  4. Call project_deployment_enable_active_learning if wanted. Active Learning needs a model on the endpoint.
• Any other new project: project_deployment_launch with model, and enable_active_learning if wanted.
• Existing project: project_deployment_set_model. It saves the configuration, so don't describe it as temporary. Never replace a configured serving model or saved prompt unless the user explicitly requests that persistent change. Training completion, Auto Label setup, and trying a preview don't authorize replacement. Never use it for temporary classes or confidence changes. Active Learning settings are preserved.

5. Improve or train

If the results miss, look at them with the user first. For SAM 3, try a few simple descriptions of the target's appearance. When quality, latency or video cost needs a custom model, sample and label representative frames and train: default to RF-DETR for object detection, including non-COCO targets, and RF-DETR Segmentation for instance segmentation, unless the user explicitly chooses another supported architecture (see roboflow-training-and-evaluation).

Put a trained model on the endpoint only when the user explicitly asks for it; finishing a training doesn't authorize it. Then use project_deployment_set_model with {type: "model_id", modelId}.