Vellum Ambient: a Mac assistant that sees your screen but will not click it

Ambient launched on September 24 for macOS users and does the one thing most screen-aware assistants skip: it annotates your display instead of driving it. That single design choice is more interesting than the model behind it.

If you have used an AI assistant that claims to “see your screen,” you have probably already learned to be careful. The pitch sounds passive — point at the problem — while the implementation is an automation that clicks through your applications on your behalf. Vellum's Ambient, released on September 24, 2026, takes the opposite approach, and that is the entire story worth telling.

What Ambient actually is

Ambient is a lightweight voice agent that floats over your desktop. It can see what you see, it can hear you, and it answers. The headline capability is what Vellum calls on-screen guidance: you ask how to do something, and instead of taking over, it points.

Ask it to help blur a face in a video in CapCut, or to match a payment against an invoice in QuickBooks, and it draws a ring or an arrow around the relevant control with a caption telling you what to do. You still perform the click.

This is a meaningfully different contract from the rest of the category. The promise is not “your work is done,” it is “you will not get lost in a UI you do not know.” For anyone who has watched an agent misclick their way through a settings page, that is a much more believable product claim than autonomous execution on day one.

How you talk to it

The interaction model is deliberately short:

• Double-press Fn to start a session.
• Hold Fn to dictate into any text field you already have focused.

There is no hotkey configuration maze, and nothing to learn beyond two gestures. For a tool whose entire pitch is being ambient, that is the right call — a global assistant that needs a setup wizard is not ambient.

The screen annotation is a public skill

One detail suggests this is more than a marketing layer: Vellum ships the screen-annotation capability as a public skill in its open vellum-assistant repository, under an MIT licence. You can read how the guidance mode actually works rather than taking the vendor's word for it.

The repository is also where a piece of engineering pragmatism shows. Vellum's code explicitly handles the case where another app has already claimed the Fn key — Wispr Flow is named in the code as a known conflict — and detects it so the same dictation does not fire twice. That is the sort of detail you only find in a real codebase, and it tells you something useful about the product's maturity.

What it costs

Ambient is not a separate subscription. It runs on Vellum's existing plans, and Vellum states plainly that it passes model provider costs through at cost with no markup on token usage.

A Custom configuration starts at $50/mo: a $10/mo platform fee plus your choice of machine tier, storage and usage bundle. Credits are bought in steps and expire after twelve months; included usage in a bundle expires at the end of each billing cycle and does not roll over.

Requirements, and the honest catch

Ambient is a macOS product. It needs macOS 15 Sequoia or later, roughly 500 MB of free space, and it runs on both Apple Silicon and Intel Macs. The screen-aware features are tied to the Mac — a Windows app exists but is described as a dev build, and there is no Linux desktop app.

So the real-world audience for this launch is Mac users who want help navigating software, not developers who want a cross-platform agent. If you are on Windows or Linux, the honest answer is that this does not help you today.

How it compares to what else is shipping

Ambient did not arrive alone. The same week saw Google's Gemini Task mode reach trusted testers, coverage of Rabbit's desktop agent, and Anthropic's Cowork browser for Claude Desktop arriving earlier in August. All of these sit in the same “agent that operates your computer” category, and they are converging on the same benchmark: OSWorld 2.0, which measures how well models complete long real-world workflows on a real desktop.

There, the picture is sobering. Claude's own Opus 5.5 launch materials report a score of 81.8% on OSWorld 2.0 — strong, and explicitly labelled as partial-credit. Other published figures for leading models on long-horizon desktop workflows are considerably lower. The gap between a benchmark number and “it works on my machine for ten minutes” is exactly where products like Ambient are competing.

What differentiates Ambient is not benchmark performance. It is that it declines the hardest part of the problem. Where competitors pursue autonomy, Ambient pursues legibility. On a category where agents are famously confident and frequently wrong, being visibly, structurally incapable of clicking the wrong button is a defensible product position rather than a limitation to apologise for.

Who should actually try it

Try it if you are on a recent Mac, you work in software you do not know well, and you would rather be shown than be automated. The guidance mode is genuinely useful for the “which button is this” class of problem, and the Fn ergonomics are better than most.

Skip it if you are on Windows or Linux, you want a hands-off agent that completes a task end to end, or you are sensitive about always-on screen capture. That last point deserves stating plainly: a tool that can see your screen will be capturing whatever is on it, including anything confidential. Read the permissions it requests on install rather than waving them through.

Bottom line: Ambient is the most interesting release this week precisely because it is modest. It ships the unglamorous capability — screen awareness without screen control — and it makes that capability public. If you have been burned by an agent clicking through your windows, a version that refuses to is worth more than one that promises to do it all.

Prices and benchmark figures are standard published rates as of September 26, 2026 and can change without notice. Some outbound links may be affiliate links — see our disclosure. Found an error? Tell us and we'll correct it in public.