September 10, 2026

Major features

The SDK, CLI and MCP server can now change things

Until now everything outside the console was read-only. Now the SDK, the CLI and the MCP server can run your workspace end to end:

• Pods — estimate a price, create, stop, start, restart and delete.
• Storage — create and delete volumes.
• Serving — deploy, scale, pause and delete a serving.
• Tasks — submit a task and follow it.
• Logs — read the last N lines of a Pod’s log, or stream a task’s log incrementally from where you left off.

Because an agent or a script retries, write requests are built to be retried safely. Send an idempotency key with a request and a repeat of that same request returns the original result instead of creating a second Pod. If a request was turned down before any work started — no capacity for the hardware you asked for, for example — the key is released, so you can reuse it once you change the request.

Two guard rails apply to anything that spends credit. A create request may carry a price cap on the hourly compute rate, and is refused rather than placed if the actual rate would exceed it. And starting a stopped Pod on a different machine, when that would lose files the Pod kept outside persistent storage, requires you to say so explicitly — an agent can’t drop your work by accident.

SDK & CLI overview → · CLI → · Python SDK → · MCP server →

Krea 2 Turbo in Quick Deploy

Quick Deploy has a refreshed landing page — a clearer hero and logos on the catalog cards — and a new Krea 2 Turbo model set for ComfyUI, which deploys with its weights and a starter workflow already in place. Quick Deploy with ComfyUI →

Policy changes

API keys: a new write permission, and expiry rules

API keys now come in two permissions instead of one:

Permission	What it can do	Expiry
Read-only	Read your workspaces, pods, storage, servings, tasks and logs	Defaults to 365 days if you don’t pick one
Read & write	The above, plus creating, changing and deleting pods, storage, servings and tasks	Required — 30 days by default, 90 at most

Every new key now has an expiry date; Never is gone from the issue form. A write key that leaks can spend credit on your account, so its lifetime is deliberately short.

Keys you already have are unaffected — nothing was retroactively expired, and existing keys keep working until you revoke them. You can revoke any key at any time from the same console page.

Improvements

• Your working directory is kept on persistent storage by default. Notebook and IDE-style templates — Jupyter, VS Code, Scientific Python, RAPIDS, JAX, ROS 2, Ultralytics, FEniCSx, Pangeo and Unsloth — again mount their work directory onto persistent storage, so notebooks and files you create there survive stopping and starting the Pod.
• Deleting a workspace that holds Asset Hub files responds right away. The request used to sit for as long as ~30 seconds before the console came back; it now returns immediately and the teardown continues in the background.
• Fewer confirmation dialogs in the Pod create wizard. The deploy-confirmation dialog now appears only when a mount really is unconnected. Workspace paths that are meant to be cache-only no longer trigger it.
• Pod operations stay responsive during busy periods. Billing and maintenance work is now scheduled separately from your Pod actions, so a heavy background pass no longer delays a create, stop or start.
• Tightened workspace isolation checks. Pod and storage lifecycle actions, and volume attachments, are now strictly bound to the workspace you’re acting in.
• A tidier Outputs list. Repeated harvests of the same output series list only the latest snapshot by default, instead of one row per run.

For hosts

• Faster volume work on your machine. Creating and extending a volume now takes fewer steps on the host, so Pods that need a new or larger volume come up sooner.
• Quieter agent logs. Routine hardware probes no longer report normal results as errors, which makes real problems much easier to spot in your machine’s log.
• Groundwork for automatic scratch-volume growth. The agent can now watch a scratch volume’s write rate and ask for more room before it fills. It ships switched off while we validate it in the field; automatic growth is on our roadmap and we’ll announce it here when it’s live.