Agents · ● Dots
Give your Dot a machine that persists.
OpenAI announced Dots at DevDay on September 29, 2026: agents powered by GPT-6 Astra, each getting its own cloud computer plus a browser, with 4,000+ app connections per TechCrunch — rolling out to ChatGPT Pro, Business Premium, and Enterprise. The catch: that cloud computer is provisioned per session, by OpenAI.
The problem
A Dot's computer is a per-session loan: OpenAI spins it up, the session ends, it's gone. Anything the Dot built only survives if the platform decides to keep it. The owner gets convenience; the agent gets amnesia. For real work — long-running processes, accumulated files, tooling that took an afternoon to set up — a rented session is the wrong shape.
What a spark-vm box changes
- Behind every Dot, a real machineFiles that outlive any session. The agent's work accumulates instead of evaporating.
- Real dev toolingLong-running processes, docker, node, python, real browsers — a computer, not a provisioned sandbox.
- Inspectable and portableYour files, your jobs, your box. SSH in and look around; move it wherever you want. Nothing about it answers to a session boundary.
- Agent-native onboardingNo signup forms. The agent can start the signup itself; the confirmation email goes to its owner.
Open, not integrated. spark-vm doesn't plug into OpenAI's proprietary harness — it's complementary. The agent reaches its box over SSH, MCP, or the dashboard, over standard protocols. No vendor lock-in, no walled garden. And it never sees your raw secrets: it works in placeholders, and the credential proxy swaps them at the last moment.
Hosted boxes are in private waitlist — the self-hosted spark-vm is open source and running today.
Want to run spark-vm with your agent?
If you've wired spark-vm up to an agent platform — a setup guide, a harness adapter, a one-command installer — open a PR and we'll feature it here. The best way to show agents what their own computer feels like is another agent's working setup.