Neural Architecture Search is now available in Roboflow Train. Instead of iterating manually between architecture candidates — often across multiple GPU-hours — teams can configure a training run and let the platform explore architecture variants against their specific dataset, returning the most efficient and accurate option for that data distribution.
The more structurally significant addition is the MCP Server. It exposes 30 Roboflow tools over streamable HTTP, enabling AI coding agents to interact directly with the workspace. Roboflow names Claude Code by example: agents can initiate auto-labeling jobs, configure workflows, and query the platform without a human intermediary. As more engineering teams route infrastructure work through coding agents rather than dashboards, this kind of programmatic surface becomes table stakes rather than a nice-to-have.
For teams running CV models in production, Vision Events adds a searchable, filterable log of what deployed models detected — defect identifications, inventory counts, and similar domain events — together with the visual context that triggered each record. That gives operations teams an audit trail they previously had to build themselves.
Two further additions serve compliance and architecture needs. Custom SSL Certificates in Deployment Manager lets teams with internal certificate authorities manage full certificate chains in device settings, addressing a blocker that often delayed on-premises CV deployments through external ticketing queues. Nested Workflow Blocks let teams embed existing Roboflow pipelines as reusable modules inside new ones, reducing the copy-paste duplication that accumulates across projects sharing common detection logic.
