Cloud browsers were built for agents. But as agents become consumer products, humans need to use those browsers too. Today we're launching Portal, a cloud browser built for both: agents can hand off live sessions to humans, who can interact with them as naturally as a local browser.
Your agent is booking a flight. It needs you to confirm something. Instead of sending you a screenshot, asking you to open a website, or abandoning the task, it hands you the actual session. You interact with it naturally, with permissions that prevent you from accessing unrelated sensitive information. Then the agent takes over again.
We think websites will be replaced by shared, governed session states.
We benchmarked Portal against Browserbase and Kernel on something cloud-browser benchmarks rarely measure: what it feels like for a human to take over an agent's browser. Portal delivered Browserbase-class input latency with Kernel-GPU-class smoothness, without requiring a GPU tier — plus features no competitor has all of: typing on mobile, local cursors, and audio.
We started off building this shared sandbox for go-to-market teams running guided, agent-guided onboarding sessions and demo conversations. Those demos needed the sandbox to feel native: real click latency and smoothness, to represent products accurately. The cloud browsers on the market at the time are incredible for agents, but aren't smooth enough when evaluating software. So we built Portal from scratch with a custom live-streaming architecture and a security system.
Past evals of cloud browsers measure agent click latencies. We measure human clicks and the human experience in sandboxes. The difference is felt at every click: roughly 250-300ms input-to-frame reads as remote desktop, while under about 90ms reads as local.
Portal was the only live view in our test that's both fast and smooth out of the box.
Results break down into quantitative measurements and qualitative checks. Findings: Portal is comparable to Browserbase on input latency, and Browserbase is better on click and type. Kernel had no US West available for our account, and even in US East ran about 2x higher click latency than Portal or Browserbase.
| Metric | Portal us-west-1 · us-east-1 | Browserbase us-west-2 · us-east-1 | Kernel GPU us-east only | |
|---|---|---|---|---|
| Smoothness · testing page | Presented fps | 39–42 (39.8) | 3.0 (2.7) | n/a (42–46) |
| Median frame gap, ms | 16.7 (16.7) | 310–340 (369.6) | n/a (16.7) | |
| Smoothness · nytimes.com | Scroll p50, ms | 68.3 (159.4) | 374.4 (438.6) | n/a (276.2) |
| Presented fps | 41.9 (42.0) | 10.5 (7.3) | n/a (42.2) | |
| Input→frame, ms p50 · testing page | Interactions (click · scroll · type) | 91–100 · 68–95 · 87 (141.6 · 164.6 · 153.5) | 67–83 · 78–104 · 80 (135.6 · 137.6 · 131.7) | n/a (229–248 · 258–304 · 314) |
On the qualitative side: a local cursor makes the session feel like your own cursor on desktop, and on mobile the cursor disappears (Portal, Browserbase) while Kernel shows a remote hand icon. GPU comes built-in with Portal instead of a $200/month tier. Typing works out of the box without Kernel's overlay bars, alongside native scrolling, horizontal fullscreen, and audio.
| Metric | Portal | Browserbase | Kernel · us-east, GPU | |
|---|---|---|---|---|
| Features & terms | Cursor feel | local overlay | local overlay | remote-rendered |
| Just-in-time human security (.ptl) | .ptl, natural language | n/a | n/a | |
| Stream transport | WebRTC video | canvas push feed · ~0.2–8.3 updates/s | WebRTC video | |
| GPU tier needed for smooth video | no | n/a | yes, $200+/mo plan · non-GPU tier (single run, eu-west): 1736 jank gaps >33ms | |
| Mobile qual · iPhone via Safari | Cursor on tap | local, hidden | local, hidden | remote + monopoly-hand icon |
| Native scrolling | yes | no (side scrollbar only) | yes | |
| Typing works | yes | no · docs: "mobile keyboards aren't officially supported" | yes | |
| Typing UX | clean | n/a | shortcut bar + clipboard modal, cut off | |
| Supports fullscreen | yes | no | no | |
| Audio | yes | no | yes | |
Median frame gap reflects rAF batching: browsers emit frames in bursts, so the mean and distribution matter more than the median. Browserbase's live view is a devtools-inspector canvas feed; it's the human-viewer surface they expose out of the box, so that's what we measured.
Blue = best value in that row, any vendor (ties share the highlight). "not run" = no measurement yet. Viewer = SF; We tested from San Francisco. Portal serves from us-west-1 (San Jose); Browserbase serves from us-west-2 (Oregon), so BB was the farther endpoint, yet still edged Portal on click latency. Credit where it's due. Kernel has no us-west column because us-west was not available for our account. Portal's local cursor overlay is a design choice that could be said to mask pointer latency; the input-to-frame metric is unaffected. Run by Portal.
performance.now() at input capture in viewer tab; t1 = first decoded requestVideoFrameCallback frame whose ROI shows the expected change. Excludes display lag.Cloud-browser infra is commoditizing: anti-bot handling, session lifecycle, and policies to persist authentication and agent identity are all must-haves. We think there are two gaps.
The first is smooth streaming. We hope this benchmark starts a pattern in the industry of making sandboxes better for human beings: for people evaluating software, learning it, and using it. There is no open-source methodology for measuring the human experience in a sandbox, so as one team we wrote our own to establish one — and we welcome the industry to run it, retest it, and work with us on a shared methodology. For future collaboration we welcome sharing our harness for folks to run broader runs in further situations, and hope this can serve as a starting point.
The second is a governance layer, where sessions can be securely shared between humans and agents to step in and step out during multi-agent interactions. .ptl is our attempt at the layer MCP doesn't cover: it compiles natural-language intent into deterministic policy with certain actions off limits.
You can imagine the same for captcha solving, or sharing software: Portal lets you try software natively in the Twitter feed, through work objects in Slack, and anywhere a link can embed.
Bottom-right HUD = the same in-page instrument on every vendor (frames seen = presented frames over identical wall clock). These are the sessions we recorded.