Claude Desktop vs Antigravity 2.0 vs Codex: Three Bets on the Agentic Desktop

By May 2026, all three frontier AI labs have shipped a flagship desktop product for the same idea: you supervise AI agents doing real work instead of typing every line yourself. Anthropic has the Claude desktop app, Google has Antigravity 2.0, and OpenAI has Codex.

From a distance they look like the same product three times. They are not. Each company made a genuinely different structural bet, and the differences decide which one fits your work. This is an honest comparison — no winner declared, because there isn't one.

A note on scope: this compares the three companies' *top-level* products. Claude's "Cowork" is a mode inside the Claude desktop app, not a separate competitor — so it is folded into Claude Desktop here, where it belongs.

The 30-Second Version

Strategy 1 — Anthropic: Breadth (One App, Three Modes)

The Claude desktop app is one downloadable application with three tabs: Chat, Cowork, and Code. One sign-in, one bill, a switcher in the sidebar.

• ●Chat — the conversational assistant, the same Claude you know from the web, for thinking and writing.
• ●Cowork — agentic knowledge work for non-developers. Point Claude at folders, give it a goal, it plans and executes. Runs locally in a sandbox on your machine.
• ●Code — the full developer surface (this is Claude Code), rebuilt in April 2026 around parallel agent sessions, git worktree isolation, and a choice of Local, Remote, or SSH execution.

Under the hood, Cowork and Code are the same engine — both are Claude Code — and they share CLAUDE.md, MCP servers, hooks, skills, and settings. One app that serves a non-technical office worker, a knowledge worker, and a developer.

The catch: it is one app at the plumbing layer but three products at the experience layer. Switch tabs and your context resets. The word "Project" means three different things across the three tabs. Instructions written in one tab are invisible in the others. The most-cited critique is bluntly titled "Claude Desktop is 3 Apps Pretending to Be One," and a GitHub request to merge the tabs was closed "not planned." The app was accreted one tab at a time, not designed as a whole, and it shows.

The bet: cover every user and every task type in a single download. Most complete scope of the three. Least coherent experience of the three.

Strategy 2 — Google: Focus (Delete the IDE)

Antigravity 2.0, launched May 19, 2026, made the boldest move: it has no code editor. Antigravity 1.0 was a VS Code fork. Version 2.0 threw the editor away and became a standalone, agent-first desktop app where you watch agents work and review what they produce.

Around that app, Google ships a CLI and a Python SDK — but there is exactly one experience, one purpose. No mode-selection tax, no "which tab do I need." Google's stated reasoning: bundling an IDE with an agent surface is "confusing," and agents are expanding past coding into general knowledge work anyway.

Antigravity's agents run in Google's cloud — ephemeral Linux sandboxes. Your laptop is just a control surface; close the lid and the work continues. The default model is Gemini 3.5 Flash, served fast.

The catch: it is cloud-only, so your work executes on Google's servers, and the day-one launch was rough — Hacker News filled with crash reports, an installer bug that bricks the old IDE, even a date typo in Google's own announcement.

The bet: one focused product, do it cleanly, accept that you serve fewer scenarios. Most coherent of the three. Narrowest. Brand new and still buggy.

Strategy 3 — OpenAI: Ubiquity (Seven Surfaces, One Engine)

Codex is not an app. It is a family of surfaces over one shared agent engine: a CLI, a cloud web dashboard, an IDE extension, a desktop app, a mobile remote, a Chrome extension, a GitHub code-review bot, and an SDK. Same agent, same login, same skills follow you everywhere — you just pick the surface that fits where you are.

Codex is the only one of the three that is genuinely both local and cloud, your choice per task. The CLI and desktop app run on your machine; Codex Cloud and the GitHub bot run in OpenAI's containers and open pull requests; the mobile app remote-controls a machine running elsewhere. If your code can't leave your network, Codex has an answer. If you want fire-and-forget cloud delegation, Codex has that too.

It is also, by a wide margin, the most mature — roughly a year old, 4 million weekly users, shipping updates weekly.

The catch: Codex has no GUI knowledge-work mode for non-developers — nothing like Cowork. It is a developer platform with many doors. Reviewers also flag that the desktop app has no built-in editor (you jump out to VS Code) and you are locked to OpenAI's models.

The bet: meet the developer on every surface they already use. Most flexible. Most mature. Coding-focused.

The Axis That Actually Decides It: Where Your Code Runs

Strip away the marketing and one technical fact separates these three more than anything else — where the agent executes:

• ●Antigravity 2.0 — cloud only. Work runs on Google's servers. Upside: your laptop can sleep. Downside: your code and data leave your machine to run a task.
• ●Claude Desktop — mostly local. Cowork runs in a sandbox on your machine; the Code tab lets you pick Local, Remote, or SSH. Upside: files stay on your device by default. Downside: for local work, your machine must be awake.
• ●Codex — both, per task. You decide each time. The most flexible answer, full stop.

If you work in a regulated environment — legal, finance, healthcare, anything with a data-residency rule — this axis is the whole decision. Antigravity's cloud-only model is a hard stop for some teams. Claude Desktop and Codex both let you keep execution local.

A Maturity Reality Check

Honest comparisons have to weigh age, not just features:

• ●Codex — about a year in market, 4M weekly users, weekly releases. Rough edges are known and getting filed down. This is a settled, working product.
• ●Claude Desktop — the Code tab redesign (April 2026) drew real praise; the three-mode app concept is genuinely contested. Mixed, but stable.
• ●Antigravity 2.0 — launched the same day this comparison was written. Day-one crash reports, an installer bug, auth failures. The vision is strong; the build is not yet ready for production-critical work.

If you need something dependable today, that ordering matters as much as any feature table.

Who Each One Is For

Choose Claude Desktop if:

• ●You want one app that covers thinking, knowledge work, and software development
• ●You are a mixed team — some coders, some not — and want a single tool and a single bill
• ●You want local execution by default and you are already an Anthropic shop
• ●You can live with the mode-switching friction

Choose Antigravity 2.0 if:

• ●You want the cleanest, most focused single-purpose agent app
• ●Cloud execution is a feature for you, not a dealbreaker — you want your laptop free
• ●You are in the Google ecosystem and want Gemini-native agents
• ●You can wait out the buggy launch, or you are just kicking the tires

Choose Codex if:

• ●You want the most mature, most battle-tested option
• ●You want to choose local or cloud per task, not be locked to one
• ●You live across many surfaces — terminal, IDE, browser, phone, CI — and want the agent in all of them
• ●You are a developer first (there is no non-coder knowledge-work mode)

The Honest Verdict

There is no winner, because the three companies are not actually building the same product:

• ●Anthropic bet on breadth — every user, every task, one app. The most ambitious scope, paid for with a fragmented experience.
• ●Google bet on focus — one agent app, no IDE, cloud-first. The cleanest concept, paid for with a narrow footprint and a rough launch.
• ●OpenAI bet on ubiquity — one engine, every surface, local or cloud. The most flexible and most mature, with no answer for non-developers.

The convergence is real — three giants independently decided the desktop should be a place to supervise agents. But the divergence is where your decision lives. Pick by the axis that constrains you: if it is data residency, the local-vs-cloud split decides it. If it is team composition, Claude Desktop's breadth or Codex's developer focus decides it. If it is "I need it to work today," maturity decides it.

The agentic desktop era has clearly arrived. It just arrived three different ways at once.

Sources

• ●Claude Code desktop documentation
• ●Redesigning Claude Code on desktop for parallel agents
• ●Introducing Google Antigravity 2.0
• ●OpenAI Codex product hub
• ●Introducing the Codex app
• ●"Claude Desktop Is Three Apps Pretending to Be One" (age-of-product.com)
• ●TechCrunch: Google launches Antigravity 2.0