Gemini 3 Pro beats Claude Sonnet on most reasoning benchmarks. Gemini Live is the most capable multimodal voice API on the market. NotebookLM is arguably the best RAG system any frontier lab has shipped. Veo 3 is state of the art for AI video. Nano Banana 2 leads on image editing.

And yet, in April 2026, the single biggest complaint you’ll find about Gemini on Hacker News, the official Google AI Developer forum, and Google’s own support community is this: the product around the model is broken.

This post explains why Google’s long-awaited native Gemini Mac app — confirmed by Bloomberg in March 2026 and likely to ship around Google I/O 2026 — is almost certainly going to disappoint the exact power users waiting for it. And it lays out what a truly native Gemini desktop experience should do instead.

If you want to skip the analysis and just try the beta, go ahead. The rest of this is the argument for why we’re building it.

Yes, Google is really building a native Gemini Mac app

Let’s start with the facts. According to Bloomberg reporting surfaced via MacRumors on March 19, 2026:

Google has shared an early version of a native Gemini Mac app with beta testers.
The beta only ships “critical features,” with more to come before release.
Google has not confirmed a release date, but the product appears to be on track for an announcement around Google I/O 2026 (May–June).
The app includes a feature called Desktop Intelligence, which lets Gemini “see what you see” (screen context) and “pull content directly from apps” on your machine.

That last bullet sounds promising. It is not quite what it appears to be.

The one sentence that tells you everything

Buried in the Bloomberg piece is a sentence most readers have skimmed past:

“The app apparently looks similar to the Gemini apps designed for iPhone and iPad.”

If you have used any Google mobile app on macOS over the last decade, your reaction to this should be: that’s bad news. Google has a long track record of shipping Mac apps that are effectively iPad layouts in a window frame — with broken keyboard shortcuts, non-native scroll behavior, ignored system preferences, and a menu bar that exists purely for show.

This is not bias. It is pattern recognition.

John Gruber’s “Electron Turd” review of Claude’s Mac app made the same point in a different direction:

“The Mac apps for Claude are simply websites residing in an app’s shell.”

Anthropic shipped Claude as an Electron app. The result, per multiple long-form critiques, is a 200+ MB binary that lags, doesn’t support macOS window-drag gestures, and breaks in small but constant ways. Mac users paid a subscription and got a webview.

ChatGPT’s Mac app, by contrast, is genuinely native. Small binary. Proper macOS gestures. Global hotkey that actually works. It’s the difference between software that was made for a Mac and software that was compiled for a Mac.

Google’s reported Gemini Mac app is going to land somewhere on this spectrum. Given (a) Google’s web-first DNA and (b) Bloomberg’s iPad-similarity tell, the likely outcome is: a product that is more iPad than Mac, shipping somewhere between Claude-quality polish and Google-quality polish. Neither of those is what power users are paying for.

The real problem isn’t polish — it’s structural incentives

Let’s be precise about why Google keeps under-delivering on desktop. This is not a story about engineering competence. It’s a story about misaligned incentives.

Reason 1: Google’s business model punishes local-first

Every Google consumer product is measured on metrics that require you to stay inside Google’s web surfaces: daily active users inside gmail.com, docs.google.com, notebooklm.google.com, drive.google.com. Local file integration — reading your ~/Documents folder directly without uploading — actively damages those metrics, because it routes the user workflow around Google’s web silos.

This is why Google Drive exists as a mandatory intermediate layer between you and your files. It is also why, when the Bloomberg report describes Gemini’s Desktop Intelligence feature, it carefully says Gemini will “pull content directly from these apps” — but says nothing about direct filesystem access, watch folders, or local file writes. The feature is designed to look at your screen, not work with your disk.

Reason 2: “Desktop Intelligence” is a late, cloud-bound copy of Claude Cowork

When you read the Bloomberg description of Desktop Intelligence carefully, it is functionally identical to Anthropic’s Claude Cowork, which shipped months earlier.

Google’s version is late. It will also be cloud-constrained, because that’s how Google’s computer-use infrastructure is architected. Project Mariner — Google’s browser agent built on the same Computer Use technology — runs in virtual machines in the cloud, not on your local machine, and is explicitly limited to browser tasks. From the Project Mariner docs:

“For now, it’s limited to browser-related tasks — it can’t access or control your computer.”

This is 2026’s state of the art from Google: an agent that works inside Chrome and runs in someone else’s data center. And it’s locked behind the $249.99/month Google AI Ultra plan, US-only.

Anything Gemini ships on Mac in the Desktop Intelligence branding will almost certainly inherit the same architectural constraints: cloud-based, browser-centric, Drive-mediated. Your Mac becomes a camera that points at whatever Gemini happens to be watching.

Reason 3: Gemini’s best features are fragmented across six different products

This is the most damaging structural issue, and it’s the one people don’t talk about. Gemini is not one product. It is at least six:

Gemini web app (gemini.google.com) — chat interface
Google AI Studio (aistudio.google.com) — developer playground with parameter controls
NotebookLM (notebooklm.google.com) — RAG-AI with Audio Overview, Video Overview, Mind Maps
Google Opal (opal.withgoogle.com) — no-code agent workflow builder
Google Antigravity — Google’s IDE-embedded Gemini CLI experience
Project Mariner — browser agent (Gemini Ultra only)

Each lives on its own domain. None of them share state. If you create a Gemini chat in the web app, you cannot reference it from NotebookLM. If you build an Opal workflow, you cannot call it from inside a Gemini chat. If you have documents in NotebookLM, Deep Research in the main app cannot see them.

This fragmentation is a direct consequence of Google’s org structure. Different teams own different products, measured on different goals, launched at different cadences. There is no unified product vision because there is no unified product team.

A native Mac app that just mirrors the Gemini web chat interface — which is what the Bloomberg beta sounds like — will not fix this. It will give you a dedicated window for one sixth of Gemini’s actual capability, while NotebookLM, AI Studio, Opal, and Antigravity remain stranded in separate browser tabs.

Reason 4: The Gemini web app is already demonstrably broken

You don’t need to speculate about Google’s execution capacity on this. The Gemini web app that the Mac app will inherit its UX from is currently generating a steady stream of complaints from paying customers.

A representative sample from April 2026:

From a Gemini Ultra subscriber posting on Google’s own AI Developer Forum: “the experience has been incredibly frustrating and, frankly, unusable for professional workflows”. The same user reports Gemini Ultra taking “30 minutes to make 10 lines of changes.”
From the Hacker News front page: “Google needs to fix their Gemini web app at a basic level. It’s slow, gets stuck…”
From Google’s own Gemini support community: Multiple threads titled “Gemini is slooooow” and “why gemini pro is too slow? it’s frustrating.”
Lag with moderately large inputs: Users report the Gemini 2.5 Pro web interface lagging on 40–50k token files, which is well below the advertised 1M-token context window.

Google is reportedly working on a “UX 2.0” overhaul for the Gemini app. That project’s existence is itself a confession that the current experience is not good enough.

The Mac app is going to be built on top of this. Unless Google rewrites the entire Gemini web frontend before shipping the Mac shell — which is not what internal sources suggest — the Mac app will inherit the lag, the clunky navigation, and the organizational fragmentation.

What power users actually need from a Gemini desktop app

Let’s ignore what Google is likely to ship and ask a different question: what should a native Gemini desktop experience actually do in 2026?

Grounded in the real complaints above, here’s the specification:

Local file system as a first-class primitive

Your ~/Documents, ~/Pictures, ~/Movies, ~/Music, and project folders should be directly readable and writable by the AI. No “upload to Drive” round-trip. No separate web UI. No temporary links. When Gemini generates an image with Nano Banana 2, it lands in ~/Pictures/Gemini/ and macOS Spotlight indexes it within a minute. When Gemini generates a Veo 3 video, it goes to ~/Movies/.

Local-first RAG is now trivially doable with sqlite-vec or similar embedded vector stores. The technology has existed for two years. Google just won’t ship it, because it breaks Drive’s growth metrics.

A real macOS binary, not an iPad port

Native SwiftUI or Tauri, not Electron. Real menu bar. Real keyboard shortcuts that respect system preferences. Real Quick Look integration. Real global hotkey. Real system tray. Real window chrome. Real dark mode that reads from the OS instead of a JS hack.

Claude’s Electron app is 200+ MB and lags. Tauri-based apps are typically 5–15 MB and feel native because they use the OS webview instead of bundling Chromium. This is the right tradeoff for AI desktop apps in 2026.

All six Gemini surfaces in one window

If a user has data in NotebookLM, the main chat interface should be able to reference it. If a user builds an agent skill, both the chat and the workflow runner should be able to invoke it. If AI Studio developer parameters are useful for a specific task, switching to “studio mode” should not require opening a new tab.

This unification is impossible for Google itself because of internal team boundaries. It is trivial for a third-party app built with a single product vision.

Gemini Live is one of the most impressive real-time AI capabilities shipped by any lab in 2026. It runs over a WebSocket, supports streaming audio input and output, and accepts continuous video frames from your screen or webcam. It supports barge-in (you can interrupt the model mid-sentence) and affective dialog (the model adjusts style to match yours).

On a desktop app, this should be a single keyboard shortcut. Hit Cmd+Shift+L, point Gemini at a specific window, say “watch this and tell me when you see a problem.” The model watches your screen until you stop it. This does not exist in any product today. It is trivial to build on top of the Live API.

Computer Use without the $249.99/month paywall

Project Mariner’s Computer Use capability is powerful but locked to Google AI Ultra at $249.99/month, US only, browser-only, running in cloud VMs. For most users, this is not usable.

The Gemini 3 Computer Use model is available in the public API. A desktop app can call it directly, run the actions locally against the user’s actual machine, and charge a fraction of the AI Ultra price — all while offering more flexibility because it is not cloud-constrained.

Multi-model routing

Gemini is excellent, but it is not always the right tool. For coding tasks, Claude Sonnet is frequently preferred. For certain reasoning tasks, GPT-5 has the edge. A desktop app should let users route tasks to the right model without reconfiguring anything — through OpenRouter’s OAuth PKCE flow, which gives you all three with a single sign-on.

Google will never ship multi-model routing. A third-party product can, and this is where the “vastly more useful than the official app at the same price” pitch actually lives.

MCP server built-in

Model Context Protocol is the emerging standard for how AI agents access external tools and data. A Gemini desktop app should expose an MCP server so that other agents — Claude Code, Cursor, Zed’s Agent Panel, Antigravity itself — can query your Gemini notebooks and workflows from their own environments. This turns your local Gemini Desktop into a backend for your entire AI tool stack.

This is the app we’re building

GeminiDesktop is the product Google’s internal incentives prevent them from shipping. It is not a web wrapper. It is a native Tauri application for macOS, Windows, and Linux that:

Reads your local folders directly (no Drive upload)
Treats any folder as a notebook, with NotebookLM-style Audio and Video Overviews saved back to the same folder
Exposes Gemini Live screen sharing as a global keyboard shortcut
Calls Gemini 3 Computer Use for local desktop automation — not browser-only, not cloud-sandboxed
Routes between Gemini, Claude, and GPT through OpenRouter OAuth (one sign-in, all models)
Ships an MCP server so Claude Code and Cursor can query your Gemini notebooks
Runs in ~15 MB, not 200 MB
Costs $20/month, not $249.99/month

You can launch the beta here. It is early and it has rough edges. It is also, in our opinion, the Gemini experience the product actually deserves.

What happens when Google ships the official app

Google’s Mac app will land around Google I/O 2026. It will be welcome. It will also, almost certainly:

Be an iPad layout in a Mac window
Require Drive for anything that touches files
Omit NotebookLM integration
Omit Computer Use (or put it behind the $249.99 Ultra tier)
Be slower than the experience you’re reading this on
Be subject to the same “merged into Google Photos in six months” pattern that users are already joking about

We are going to write about it when it ships. We are not going to stop building our app because Google builds theirs. The differentiators above are structural — they come from Google’s incentives, not Google’s execution — and they will remain valid no matter what Mountain View announces on stage.

Frequently asked questions

When will the official Google Gemini Mac app launch?

Google has not announced a public release date. Based on Bloomberg’s March 2026 report and Google’s historical pattern of using Google I/O for major consumer AI announcements, the most likely window is May–September 2026.

Is GeminiDesktop affiliated with Google?

No. GeminiDesktop is an independent product built by developers who use Gemini every day. Gemini and NotebookLM are trademarks of Google LLC. We are not endorsed by, sponsored by, or affiliated with Google.

What’s the difference between the Gemini web app and a native desktop app?

The Gemini web app is a browser-based interface that runs inside Google’s infrastructure and stores everything in Google’s cloud. A native desktop app runs as a compiled binary on your machine, can read and write local files directly, integrates with your OS (keyboard shortcuts, system tray, global hotkeys), and can operate offline for many use cases.

How does Gemini Computer Use compare to Claude Cowork?

Both let the AI observe your screen and take actions on your behalf. Claude Cowork runs natively on macOS and Windows and can control arbitrary applications. Google’s Project Mariner — the productized version of Gemini 3 Computer Use — is currently limited to browser tasks, runs in cloud VMs, requires the $249.99/month AI Ultra plan, and is US-only. The underlying Gemini 3 Computer Use model is, however, available via the public Gemini API, which is what GeminiDesktop uses.

Does Gemini work well with local files?

Not natively. Gemini’s file understanding requires uploading files to Google’s servers. NotebookLM’s source limits are 500,000 words or 200MB per file, with notebook-level caps of 50 sources on the free tier. GeminiDesktop uses local embeddings to let you query folders of thousands of files without uploading anything.

Why is the Gemini web app so slow?

Google acknowledges performance issues. Users have reported lag with moderately sized inputs, slow response times at peak hours, and regressions in stability. Google is working on a “UX 2.0” overhaul, but the Mac app is expected to launch on top of the existing web stack, so the underlying performance profile will be similar.

Stop waiting. Start using.

If you’re a Mac user who has been holding out for Google to ship something you can actually use, you have two choices in 2026:

Wait another three to nine months for a Mac app that is probably going to be an iPad layout with Claude Cowork-style screen features bolted on, locked into Drive, missing NotebookLM integration, and running on top of a web stack your Ultra subscription already complains about.
Try GeminiDesktop now.

We built option 2 because option 1 is not good enough.

And if you want the deeper story on why NotebookLM — arguably Google’s single best AI product — also has no native app, read our companion post. It’s part of the same structural story.