
Product & Updates

Every Agent Task Starts With a Wait
An agent decides its next action in milliseconds. Then it waits: for a browser to boot, a sandbox to provision, a desktop to come online.
A few seconds per task is invisible at ten runs a day. At ten thousand, the environment sets the pace and the bill, not the model.
Those environments also have to be contained. Real credentials, generated code, and scraped data should never share a machine with someone else's run. Fast and isolated usually pull against each other. Solari was built to deliver both.
One API. Browsers, Sandboxes, Desktops.
Solari is the full environment layer for AI agents behind a single API key:
Cloud Browsers. A real Chrome, ready in 8ms. Stealth, residential proxies, captcha solving, persistent logins, and session replay built in. Drive it with Playwright, Puppeteer, Selenium, or raw CDP.
AI Sandboxes. Isolated Linux compute for generated code, tool calls, builds, and data jobs. Ready in 90ms.
Computer Desktops. Full Linux machines an agent can see and control over VNC, for software that has no API. Resumes from a snapshot in 0.78ms.
Every environment runs on a hardware-isolated microVM. Sandboxes and desktops start from memory snapshots, so they're live in milliseconds instead of booting. SDKs ship for TypeScript, Python, Go, Rust, and C++, plus a raw HTTP API.
Solari is engineered by Pinetree Research, the team building computer-use agents for real-world workflows. It's the infrastructure our own agents run on, opened up to everyone.
Fastest, by 4x
On Steel's own open-source benchmark — create a browser, connect, navigate, release, measured end to end — Solari finishes in 199ms. Kernel takes 778ms. Steel takes 867ms. Browserbase takes 2,888ms.
Time to ready: 32ms, against 317ms for the next-fastest provider. Across 5,000 measured runs, zero failures.
Sandboxes tell the same story. On the open-source nibzard benchmark — create the machine, run the same heavy workload, tear it down — Solari finishes in 8.2s, ahead of E2B (10.8s), Modal (11.9s), Daytona (13.6s), and CodeSandbox (24.8s).
Learn One Client. You Know All Three.
Swap the import for @solarisdk/sandbox or @solarisdk/desktop. The key, authentication, and session model stay the same. Start with one primitive and add the others as your agents grow.
What Comes Next
As we enter beta, browsers, sandboxes, and desktops are live and free to start. Beta also means we're still shipping and pushing hard: cold starts get shorter, isolation gets tighter, and the benchmarks get rerun as the numbers move.
Three things are coming next. A public cost benchmark at 100, 1,000, and 10,000 concurrent agents. A security page covering isolation, data handling, and compliance. Shareable recipes anyone can open and run in one click.
The goal is simple. Companies and developers should be able to scale their agents faster and cheaper, on environments that stay secure and reliable at any volume. If you're running agents in production today, we want to hear what breaks.
Learn more at docs.getsolari.com and getsolari.com.