---

name: roboflow-inference description: "Route Roboflow Inference users through CLI operations, SDK/WebRTC, workflow processing, and model runtime selection." metadata: disco-role: operating disable-model-invocation: true license: NOASSERTION

Roboflow Inference

Use this skill when the task is about Roboflow Inference's public runtime surfaces: the inference CLI and server, inference_sdk, workflow image/video processing, or inference-models runtime selection.

This skill is self-contained. Use the bundled references and scripts in this tree instead of reopening the source repository once the route is clear.

Start here

1. Identify the user-facing surface the request names.
2. Open the matching sub-skill.
3. If the problem is installation or a predictable failure mode, open the cross-cutting references first.

Route map

• sub-skills/cli-operations/SKILL.md
  • Server lifecycle, infer, benchmarks, cloud deploy, Roboflow Cloud staging, batch processing, and enterprise compilation.
  • Also covers the experimental benchmark inference-models-speed path, but backend/package-choice questions still belong in model-runtime.
• sub-skills/sdk-webrtc/SKILL.md
  • InferenceHTTPClient, workflow and pipeline HTTP APIs, and WebRTC streaming with webcam, video file, RTSP, MJPEG, local stream, or manual sources.
• sub-skills/workflow-processing/SKILL.md
  • inference workflows process-image, process-images-directory, and process-video, including parameter files, output layout, and overwrite controls.
• sub-skills/model-runtime/SKILL.md
  • AutoModel, backend selection, local packages, cache/offline behavior, environment inspection, and runtime load failures.

Package family map

• inference / inference-cli → CLI, server, cloud, rf-cloud, and workflow command families.
• inference-sdk → HTTP client and WebRTC streaming.
• inference-models → model loading, backend negotiation, and package/runtime inspection.

If the user asks about a surface that spans more than one family, start from the family that owns the user-facing command or API, then hand off to the neighbor sub-skill for shared troubleshooting or backend selection.

Common request cues

Use the owning sub-skill by the exact surface the user mentions:

• inference server start|status|stop → cli-operations
• inference infer → cli-operations
• inference benchmark api-speed or python-package-speed → cli-operations
• inference cloud ..., rf-cloud ..., enterprise inference-compiler ... → cli-operations
• inference workflows process-* → workflow-processing
• InferenceHTTPClient, run_workflow, start_inference_pipeline_with_workflow, client.webrtc.stream(...) → sdk-webrtc
• AutoModel.from_pretrained(...), describe_compute_environment(), backend extras, local packages, cache/offline mode → model-runtime

Read these next

• Overview for the high-level surface map.
• Installation when a package is missing or the environment needs to be prepared.
• Troubleshooting when a command, stream, or model load fails.
• Repository provenance when you need the source snapshot for staleness checks.
• Router metadata when a machine-readable import or routing record is needed.

Helpful probe scripts

• sub-skills/cli-operations/scripts/inspect_cli_help.py
• sub-skills/sdk-webrtc/scripts/check_webrtc_surface.py
• sub-skills/workflow-processing/scripts/inspect_workflow_outputs.py
• sub-skills/model-runtime/scripts/describe_compute_environment.py

Use the probe that matches the surface before reopening source files.

When to switch sub-skills

• If the user is actually asking for workflow image/video processing, switch from CLI operations to workflow processing.
• If the user is choosing a backend, debugging cache/offline behavior, or loading a local package, switch to model runtime.
• If the user is trying to stream frames or configure the HTTP client, switch to SDK/WebRTC.
• If the user is trying to start the server, run cloud, rf-cloud, or enterprise commands, stay in CLI operations.