Cursor Cloud Agents now spin up from a prompt, no repo required
It scaffolds an Origin repo, serves a live browser preview, and can publish to Vercel — plus Codex extension hooks and a free finance MoE.
Copy markdown
Prompt in, repo scaffolded for you
Pick "Start from scratch" in the Cloud Agents repo picker, type a prompt, and Cursor spins up a working project with no GitHub repo attached — it creates an Origin repo in the background that you can save under a real name once the output holds up.
A live preview, then a real URL
Port-forwarding streams the agent's running environment straight into your browser, design-mode tools included, and a connected Vercel account turns the finished build into a live URL — so one session can go from prompt to shippable site.
Codex extensions can now rewrite tool output
OpenAI's Codex 0.151.0 (Aug 29) lets extensions inspect or replace MCP tool results before the model ever sees them — a clean hook for trimming noise, redacting secrets, or reshaping payloads. It also adds a configurable grace period for optional MCP servers and permission profiles that persist across terminal-UI sessions.
A free finance-tuned MoE on Vercel's gateway
InclusionAI's Ling 3.0 Flash Fin — a 124B-param MoE with ~5.1B active and a 256K context, tuned for financial research and long-horizon agentic planning while keeping coding and math — is free right now on Vercel AI Gateway via inclusionai/ling-3.0-flash-fin-free. The -free ID stops serving rather than billing you when the window closes.
Transformers 5.16.1 runs the newest drops
Hugging Face Transformers 5.16.1 adds native GLM-5.3-Flash support — the first multimodal GLM-5 — on top of 5.16.0's Qwen4-Exp, Granite Speech 5.0, and Step-3.7-Flash, so you can pull this week's open models locally without hand-patching the library.