Roboflow is a computer-vision platform and a family of open-source CV tools. The commercial side handles dataset management, annotation, training, hosted inference, and deployment; the open-source repos give an agent or developer the building blocks to self-host and drive directly. The reason it earns a leaf here is the one most CV vendors miss: a genuine agent-drivable surface — an official hosted MCP server, an OpenAPI inference server that runs on a Raspberry Pi, composable Workflows, and a published Skills repo — not just a SaaS UI. This leaf maps what's genuinely open source, what's metered, and where it fits a POPIA-sensitive edge deployment.
Roboflow is maintained by a US-based, GitHub-verified organisation that controls the roboflow.com domain. Two things ship under the name, and keeping them apart is the whole game:
supervision, inference, rf-detr, trackers, maestro) that a developer or agent can install, self-host, and run offline.You can use the open-source core without ever touching the cloud, and you can use the cloud without ever reading the source. Most real deployments mix the two, which is exactly why the licence boundary (§05) matters more than the marketing line "Roboflow is open source."
The selection criterion for this node is agent-drivability. Roboflow exposes its capabilities through an official hosted MCP server and an OpenAPI inference server — a sub-tree agent can build, train, and deploy a vision pipeline from a prompt, or call a self-hosted inference server as a standard REST tool. That is a sharper reason to track it than "it annotates images well," which a dozen tools do.
Adoption numbers from the GitHub org, fetched 2026-06-14. Star counts are point-in-time and drift — treat them as order-of-magnitude, not a live leaderboard.
supervision stars — the most depended-on piecenotebooks stars — tutorials, ResNet→RF-DETRrf-detr stars — the detection modeltrackers stars — multi-object trackingmaestro stars — multimodal fine-tuninginference stars — the serverRF-DETR is published by Roboflow as an ICLR 2026 real-time object-detection / segmentation architecture, stated SOTA on COCO and designed for fine-tuning. The "SOTA on COCO" line is the vendor's framing — reported here as Roboflow's claim, not an independently re-run benchmark. The architecture being Apache-2.0 and fine-tunable is the part that matters for a self-host story.
Five repos do the heavy lifting. Each carries its own licence (see §05) — the self-hostable core is the permissive subset.
Reusable CV glue: detection post-processing, annotation, tracking glue, zone / line counting. The most depended-on piece in the ecosystem.
Turns any machine or edge device into an inference server. REST API with OpenAPI (/docs) + Redoc, a Workflows engine, RTSP / webcam pipelines. Runs on Linux, Windows, Mac, Jetson, Raspberry Pi.
The detection / segmentation architecture. Real-time, fine-tunable, ICLR 2026.
Modular re-implementations of leading multi-object tracking algorithms. Combine with any detection model.
Fine-tuning toolkit for multimodal models — PaliGemma 2, Florence-2, Qwen2.5-VL.
Official SDK for datasets, models, and deployments. The cloud-connected control surface in code.
Workflows are the composable unit on top. Each block gives models a common interface for chaining, swapping, and adding business logic — the thing an agent assembles into a CV micro-service or a self-contained visual agent running on a video stream (detect → track → count → notify).
# self-host the inference server (validated quickstart) pip install inference-cli && inference server start --dev # Jupyter quickstart then runs at http://localhost:9001/notebook/start
Roboflow runs an official hosted MCP server at https://mcp.roboflow.com/mcp, exposing 30 tools across projects, datasets, training, Workflows, and Universe. The server, the Skills, and the plugin are free and open source under Apache-2.0, in the repo roboflow/computer-vision-skills.
A single hosted URL means any MCP client — Claude Code, Codex, Cursor — picks up new capabilities on connect, with no SDK pin and no version drift. Through it an agent can create projects, upload and auto-label images, pull Universe datasets, generate dataset versions, trigger training (e.g. RF-DETR), and build or deploy Workflows.
# validated install (Claude Code) claude mcp add -s user roboflow \ --transport http https://mcp.roboflow.com/mcp \ --header "x-api-key: YOUR_ROBOFLOW_API_KEY" \ --header "Accept: application/json, text/event-stream"
Several third-party Roboflow MCP servers also exist (eusef, nickedridge-wq, Composio, viaSocket). For production use prefer the official mcp.roboflow.com server — it is first-party, Apache-2.0, and capability-tracked at one URL. Supporting signal that the codebase is authored for agents: roboflow/inference ships an AGENTS.md and a .claude/skills/add-inference-model directory.
True OSI open source vs. metered / source-available divides cleanly. Do not flatten it into a single claim — the difference is the difference between "an agent can run this offline" and "production needs a contract."
| Component | Licence | OSI open source? |
|---|---|---|
| supervision | MIT | Yes |
| rf-detr, trackers, maestro, roboflow-python | Apache-2.0 | Yes |
| inference (core) | Apache-2.0 (LICENSE.core) | Yes |
| inference (enterprise dir) | Enterprise, source-available | No — production needs a contract |
| computer-vision-skills / MCP server | Apache-2.0 | Yes |
| Hosted cloud (registries, training, deploy) | Proprietary, metered | No |
| Some pre-trained / foundation models | Architecture-dependent; paid commercial licence | Varies |
The self-hostable core — inference core + supervision + rf-detr + trackers — is genuinely permissive and agent-drivable offline. Cloud-connected functionality (model / Workflow registries, dataset management, monitoring, managed infra) requires a Roboflow account + API key and is metered. Assume nothing is free until you've checked which side of this line a feature sits on.
maestro.• You need agent-drivable CV with a real MCP surface, not just a labelling UI.
• You want a permissive self-hostable core (MIT / Apache-2.0) for edge / on-prem.
• You need fast dataset → train → deploy loops and can accept metered cloud for the managed parts.
• RF-DETR or multimodal fine-tuning is on the roadmap.
• The workload is privacy-critical and you cannot self-host — hosted cloud is US, metered, and phones home with an API key.
• You only need a single static model with no pipeline — supervision + a raw model may be lighter.
• You need production use of inference/enterprise — that requires an active Enterprise contract, not the open-source path.
• You assumed "open source" means everything is free — registries, training, monitoring, and dedicated deploy are proprietary and metered.
Data residency is the whole POPIA story. Self-hosting inference keeps image / video frames and the inference itself on-prem or in a chosen SA cloud region. That matters where CV data includes biometric or identifiable persons — faces, plates. The hosted Roboflow cloud is US-centred with no stated SA region, so sensitive workloads should route through self-host.
Biometric CV is special-personal-information processing under POPIA — high-sensitivity. Prefer on-device inference; keep training data and predictions inside the SA boundary.
Cloud-region fit for a self-host target: GCP africa-south1, Azure southafricanorth, AWS af-south-1 — inference runs in any of them as a container. For where Workers-fronted services actually serve from in-region, see tech/cloudflare/edge-africa-me.
The metering caveat is the trap. API-key features phone home to Roboflow cloud (US). For data that must not leave the boundary, run open-access mode (no key) or fully self-hosted — accepting the loss of registries and monitoring. Document this boundary so an agent never silently routes a sensitive frame to US cloud.
computer-vision-skills — the same folder-plus-SKILL.md packaging standard, applied to vision.