Cursor Cloud Agents can now run their tools on your own machine

Tool execution moves to your own hardware while planning stays in the cloud; pools autoscale, secrets stay put — plus Claude's content checker goes live.

Nowline Sep 3 2026 3:00 AM banner

Top AI stories from the last hour

Top AI stories from the last hour

Copy markdown

  • Your agent's terminal now runs on your box

    Cursor's new My Machines executes a Cloud Agent's actual tool calls — terminal commands, file edits, browser actions — on your own laptop, devbox, or VM, while the agent loop and inference stay in Cursor's cloud. You get instant access to your real repo, dependencies, build cache, and private network instead of reproducing them in a sandbox.

  • Four commands, no firewall changes

    Setup is install the CLI, run `agent login`, then `agent worker start`, and your machine appears in the environment dropdown at cursor.com/agents. The worker opens only outbound HTTPS connections, so no inbound ports are exposed and your secrets, build caches, and test outputs stay on the machine.

  • The catch: your code still reaches Cursor

    Self-hosting keeps execution local, but tool outputs and transcripts still flow back to Cursor's cloud for inference and may be processed and stored — so sensitive code isn't air-gapped. Read it as a latency and local-access win first, and a data-residency win only in part.

  • Team pools autoscale across eight providers

    For orgs, Self-Hosted Pools scale workers up and down on queue demand with no pre-sizing, and hibernate idle workers to snapshots that restore under the same ID within a reconnect window. Runtimes plug into AWS Lambda, Modal, E2B, Cloudflare, Daytona, Namespace, Coder, and Vercel.

  • What you could ship this weekend

    Point a Cloud Agent at a repo that only builds on your box — native deps, a licensed toolchain, a service behind your VPN — and let it run tests against the real thing instead of a mock. Linux and Mac workers can also drive a local Chrome for browser automation when Chromium is installed.

  • Elsewhere: Claude's content checker goes live

    Anthropic's verification tool at claude.com/check-content is now public, letting anyone test whether a file carries a Claude-issued Content Credential. It arrives as Claude's imperceptible text watermarks and signed C2PA metadata on PNG, JPG, and SVG files roll out across the API and apps worldwide, with no opt-out.

  • Elsewhere: Adobe puts 70+ tools in Slack

    Adobe shipped an MCP-based Slack app exposing 70+ tools across Firefly, Photoshop, Acrobat, Express, and Premiere, so you can generate and edit assets straight from a thread. Access is limited to Slack Business+ and Enterprise+ for now.