Roboflow Computer Vision Skills
https://github.com/roboflow/computer-vision-skills- Category
- AI Agents
- Rank
- No. 1509Tools index
Previous survey · No. 1485 ·
- Pricing
- Open Source
- Type
- AGENT
- Builder
- roboflow
- GitHub
- 35 stars
- Date
About
Agent-ready skills for Roboflow and computer vision workflows.
What it does
This plugin gives coding agents detailed playbooks for building and operating Roboflow computer vision projects. Its guidance covers datasets, cloud storage, training, evaluation, deployment, inference, workflows, public models, pricing, and APIs. Live project actions come through Roboflow’s authenticated MCP service, while the skills explain which actions to take and why.
Why it's ranked here
The separation between operational tools and written guidance is practical. Agents get live access to projects and models without losing deployment comparisons, security advice, response-size warnings, or workflow design rules. Broad coverage makes it useful across a project lifecycle, although much of its value depends on Roboflow services and an API key.
What's good
The guidance makes consequential choices explicit. It compares serverless, dedicated, self-hosted, batch, and live-video deployment paths. It warns that segmentation polygons and rendered images can overwhelm agent context. Cloud credentials are collected outside chat, validation precedes credit-consuming mirror runs, and long batch jobs can use webhooks instead of repeated polling.
Tradeoffs
This is primarily product guidance and integration configuration, not a standalone computer vision framework. Live actions require Roboflow services and authentication. The standalone skill installation omits the bundled MCP connection. Codex installation also requires adding a marketplace source, restarting, and completing installation through the plugin browser. The macOS, Linux, and Windows installer scripts are placeholders with no real installation logic.
How to use it well
Use it with an agent that helps manage Roboflow datasets, design saved workflows, select deployment modes, run inference, or automate batch processing. Install only the relevant skill when you need reference guidance, or use the full plugin when authenticated actions matter. Keep separate API keys per project when workspaces or billing accounts differ. Look elsewhere for model implementation or a vendor-neutral vision stack.
Technical notes+
The canonical content is Markdown under skills/, with manifests in .codex-plugin/plugin.json and .claude-plugin/plugin.json pointing to that shared directory. skills/inference/SKILL.md separates deployment strategy from protocol reference, while skills/api-reference/SKILL.md documents hosts, authentication, SDKs, and limits. skills/inference/bin/poll_batch_job.py is an executable Python CLI that uses inference_cli helpers, polls batch metadata, reports state changes, and returns distinct success, error, timeout, and interruption exit codes. agent-install/agent.sh and agent-install/agent.ps1 contain TODO markers and only print placeholder messages.
Observed
- License
- Apache-2.0
- Content format
- Markdown skills with one Python batch-job helper and plugin configuration files
- Plugin interfaces
- Claude Code and Codex plugin manifests share the same skills directory
- Standalone install
- Individual or complete skills can be installed through the npx skills CLI
- Live interface
- HTTP MCP service for authenticated Roboflow project, dataset, model, workflow, and inference operations
- Agent support
- Plugin or standalone-skill workflows are documented for Claude Code, Codex, Cursor, OpenCode, and other SKILL.md readers
- Installer status
- The macOS, Linux, and Windows installer scripts are placeholders
Read from README.md, skills/inference/bin/poll_batch_job.py, LICENSE, .mcp.json, agent-install/agent.sh, agent-install/agent.ps1, .codex-plugin/plugin.json, .claude-plugin/plugin.json, .claude-plugin/marketplace.json, skills/universe/SKILL.md, skills/inference/SKILL.md, skills/api-reference/SKILL.md, skills/cloud-storage/SKILL.md, skills/inference/workflows.md, skills/inference/batch-jobs.md.
What it can do
Train custom object detection models
Labeled image dataset → Trained computer vision model
Annotate images for training data
Raw images → Labeled bounding boxes and classifications
Deploy computer vision models via API
Trained model → REST API endpoint for inference
Perform real-time object detection
Image or video stream → Detected objects with coordinates and confidence scores
Augment training datasets
Original image dataset → Expanded dataset with variations and transformations
Convert between annotation formats
Annotations in source format → Annotations in target format
Generate synthetic training data
Base images or parameters → Synthetic images with annotations
Tags
Tech Stack
Comments (0)
No comments yet
Editorially curated, with community endorsements as a secondary signal. Corrections welcome.