Roboflow

What it is for

Roboflow is a vision factory: label a dataset, train a detector, then run it on their GPUs or on your Jetson. Use it when the job is credits for images and train jobs, plus a $0 self-host inference binary, not when you need a $20 chat seat.

What makes it unique: the cover is the official roboflow/inference GitHub file list — the OSS deploy runtime — not a pricing card.

Key features

• Datasets + labeling — upload images, draw boxes/polygons, and run AI Label Assist (pricing; docs). One corpus. Review Mode and outsourced labelers are Enterprise SKU of this.
• Hosted training — train YOLO / RF-DETR / SAM-class models on their GPUs (pricing “Model Training”). A fit. Concurrent jobs and NAS are SKU of this train.
• Cloud inference — Serverless API and streaming billed in credits (RF-DETR Medium ≈ 5,330 images / credit) (pricing Deploy table). A hosted call. Model names are SKU.
• Self-host Inference — the Apache-2.0 roboflow/inference package on Jetson / Pi / your GPU. $0 inference on every plan when you bring the machine (pricing FAQ). A deploy. Ultralytics commercial weights on your hardware are an Enterprise add-on.
• Workflows — a graph of models + logic + integrations you version (pricing Workflows). An orchestrator, not a single model. Vision Events are logs of this.

Roboflow Agent (idea → deploy) and the MCP/Skills installer are assist / install. Credits are the meter. Universe publishing is optional.

Who it's for

Best fit: CV teams who will stay on Free ($0 / 10 credits/month) while they self-host Inference, then buy Core at $39/month when cloud train + API credits run out.

Pricing

Verified 2026-10-22 (Asia/Shanghai) from roboflow.com/pricing. Cloud usage is credits. Self-host inference is $0.

• Free Tier — $0/month. No card. 10 credits/month (enough to train ~30 models or run ~80,000 inferences on their example). Unlimited admin seats. 20 projects. Private by default. Weights via Inference / MCP.
• Core — $39/month, billed monthly. Credit pack beyond Free (their card: ~60 train jobs or ~160,000 inferences). Same product surface as Free plus the pack.
• Extra credits — prepaid packs from $4 (expire in 1 year). On-demand credits $6 (card required). Unused included credits do not roll; they reset each period.
• Enterprise — custom. SSO, RBAC, SLA, US/EU residency, manufacturing add-ons, labeling services from $0.10/box, $0.20/polygon, $0.05/classification.
• Research — academic email, public Universe projects.

Older seat-priced Core plans convert to credit packs; the public page says your price stays until they email you.

Limitations / costs

• Included credits reset. Prepaid is the only credit that lasts a year.
• Self-host commercial YOLO weights need an Enterprise add-on. Hosted API already includes that license.
• A vision factory, not a design mockup tool. For UI mockups see Visily. For hosted image models see Fal.

Alternatives

• Fal — hosted vision/image APIs when you will not label a dataset.
• Replicate — run a public vision model when you do not need Roboflow’s dataset UI.
• Visily — product mockups when the “vision” job is a screen, not a detector.

Official tutorial

Official Roboflow channel video: Introducing Roboflow Agent: Go from Idea to Working Vision App (oEmbed author: Roboflow). First-party idea → vision app. Agent is assist on the same factory, not a sixth KF.

Written path: roboflow.com/pricing plus the Inference README.

Website

https://roboflow.com

Repository

https://github.com/roboflow/inference

Source

Catalog row curated for AI Tools Nav. Cover is the official roboflow/inference GitHub file list. Dollars checked on roboflow.com/pricing (Free $0 / 10 credits; Core $39/mo; prepaid from $4; on-demand $6). Tutorial oEmbed author: Roboflow.

Last verified

2026-10-22 (Asia/Shanghai) editorial check.