Claude Code vs Codex CLI vs Antigravity: which coding agent to use

Three coding agents now cover most of what developers search for. Claude Code is Anthropic's, Codex is OpenAI's, and Antigravity is Google's. All three read your repository, edit files and run commands; all three start at $20 a month. They differ in where they run, how they ask before acting, and how far you can steer them away from the desk. This comparison is checked against each vendor's own pages as of 29 October 2026.

What are Claude Code, Codex CLI and Antigravity?

Claude Code is Anthropic's agent. It started in the terminal and now runs the same engine in a CLI, a desktop app, VS Code and JetBrains extensions, cloud sessions on the web, and the Claude mobile app. The CLI is the most complete surface; scripting and the Agent SDK are CLI-only.

Codex is OpenAI's. The Codex CLI is open source under the Apache-2.0 licence, and the same agent runs in an IDE extension for VS Code, Cursor and Windsurf, a desktop app (codex app), a cloud agent at chatgpt.com/codex, and on iOS.

Antigravity is Google's. It launched in preview in November 2025 as an agent-first IDE with an agent manager beside the editor. Antigravity 2.0, announced at Google I/O on 19 May 2026, added a standalone desktop app for running several agents in parallel, a new CLI written in Go, and an SDK. Google is steering people there from Gemini CLI: consumer access to Gemini CLI and the Gemini Code Assist IDE extensions ended on 18 June 2026.

How do they compare? (checked 29 October 2026)

Prices are list prices in US dollars from each vendor's own pages on the date checked. Usage limits are not comparable across vendors: Anthropic counts a five-hour session window and a weekly limit shared with the Claude apps, OpenAI publishes estimated messages per five hours by model, and Google describes quotas and sells credits on top. Treat the entry price as a ticket to try, not a measure of how much work you get.

PeakWatch does this from your own data. Your coding agent, on your wrist. Join the free beta →

How does each one ask before it acts?

This is where the three differ most, and it matters more than the price. An agent that edits files and runs commands on your machine needs a gate, and the gate is the safety feature, not friction.

Claude Code has six permission modes: default (manual: asks before most edits, shell commands and network actions), acceptEdits, plan, auto (a classifier model reviews actions in place of you), dontAsk and bypassPermissions. Interactive terminal and VS Code sessions on recent versions start in auto. Allow, ask and deny rules sit on top of the mode, and hooks can add a deterministic check. The permissions guide covers the rule syntax.

Codex splits the question in two. A sandbox mode decides what the agent can touch (read-only, workspace-write or danger-full-access), and an approval policy decides when it asks (on-request, never, or granular per category). The Auto preset, workspace-write with on-request, is the default in a version-controlled folder; a folder outside git starts read-only. Network access is off by default. An auto-review option sends eligible approval requests to a reviewer agent.

Antigravity sets review through a security preset per project, which decides whether you approve terminal commands and file access. The presets are named Default, Request Review and Turbo, and an Auto mode in preview approves edits and safe commands while still prompting before irreversible ones; commands that would leave the sandbox go to a separate reviewing agent first. A separate setting decides whether you review the agent's plans and other artifacts or let it proceed.

Can you run them away from your desk?

Long agent sessions stop to ask, and an ask you cannot answer is a stalled session. The three take different routes.

• Claude Code has two. Cloud sessions on the web keep running after you disconnect. Remote Control connects the Claude app on iOS or Android, or claude.ai/code in any browser, to a session running on your own machine, so the code and the files stay local. It is available on Pro, Max, Team and Enterprise, not with an API key. How to set it up is its own guide.
• Codex runs cloud tasks you can start and follow from the iOS app and the web, which suits work that needs little steering.
• Antigravity 2.0 lets the desktop app schedule background tasks and run several agents at once. We did not find a phone client or a remote control for local sessions in Google's announcement or codelab.

Which coding agent should you use?

1. You live in the terminal and want the most mature permission system: Claude Code. It also has the deepest extension story, from skills to plugins, covered in the best Claude Code plugins and skills.
2. You want an open-source CLI and a sandbox you can reason about: Codex. The two-axis sandbox and approval model is the easiest of the three to explain to a security team.
3. You want an IDE built around agents and you already pay for Google AI: Antigravity, with the caveat that its quotas have changed during 2026 and are described loosely.
4. You already pay $20 a month for one of the chat apps: start with the agent that subscription includes. All three are good enough that the habit you build matters more than the model.

Whichever you pick, keep the approvals on and move them somewhere you can answer them. An ask that waits at the desk costs far more than the second it takes to answer.

How PeakWatch fits

PeakWatch is built for Claude Code. It pairs a Wear OS watch, including Galaxy Watch, with a relay on your desktop that drives sessions through the Claude Agent SDK: approve or deny tool requests with the arguments in view, dictate prompts and read replies from the wrist. It is in private beta, and other agents are on the roadmap rather than supported today.

Hen Eytan · Founder, UPeak

Builds the UPeak apps and writes up the research behind them.