Claude Code vs Gemini CLI: seven ways developers split the work, and which ones survived June 18

On June 18, 2026 Gemini CLI stopped answering everyone who used it on a free or Google One account. Google’s transition post had announced the date a month earlier: Gemini CLI would “stop serving requests for Google AI Pro and Ultra, as well as those using it free of charge.” Those users moved to a different tool, Antigravity CLI, whose command is agy. Search for the two names together and the first results still list a free tier and a million-token window under Gemini.

This is the third piece in a series, after Claude Code and Codex and Claude Code and Cursor. The other two tools changed this year too. This one stopped serving most of its individual users. In May I wrote that if a harness is wired to one CLI specifically, the next change in that CLI’s meter rebuilds it. June 18 was a bigger change than a meter, on Google’s side, and the seven setups below are the test.

Same labels as before. Documented means a vendor’s docs, or a project’s own repository, show the mechanism. Reported means a named person published their setup or their findings and I am relaying them without having reproduced them. Everything about Gemini CLI and Antigravity CLI below comes from Google’s documentation and those sources, read on November 8.

Who can still run Gemini CLI after June 18, 2026?

Organizations with a Gemini Code Assist Standard or Enterprise license, Code Assist for GitHub through Google Cloud, and anyone on a paid API key. Everyone else, free users and Google AI Pro and Ultra subscribers included, was moved to Antigravity CLI. The repository README puts it in two lines:

Enterprise Code Assist only: Designed for Gemini Code Assist Standard /
Enterprise licenses and Code Assist for GitHub on Google Cloud (your
access remains unchanged).
No free tier or Google AI plans: Free individual access and Google AI
Pro / Ultra (Google One) moved to Antigravity CLI on June 18, 2026

The project is not dead. Version 0.67.0 shipped on November 4, and the transition post says Google will keep giving enterprise users “access to the latest Gemini models.” But the README says its harness is not tuned for the Gemini 4 series, and the docs site sends anyone who wants Gemini 4 Argon to Antigravity. Google’s newest frontier model is recommended on a different tool than the one this article is named after.

The docs disagree with each other, which matters if you are reading them to decide. The quota and pricing page still opens with “Gemini CLI offers a generous free tier” and a table of 1,000 requests a day for individuals, under a site banner announcing that Gemini CLI “was replaced by Antigravity CLI on June 18th, 2026.” The README and the Google Cloud quota page list only Standard and Enterprise. Trust the README and the banner. A page body can go stale. A banner gets added on purpose.

Why did developers send Claude Code’s long-context reads to Gemini CLI?

Because on a personal Google login it cost nothing and its window was bigger. That login now goes to Antigravity, and the window gap closed. It is also one of the earliest, and it got copied a lot. Peter Steinberger published a gist on June 29, 2025 (reported) that teaches Claude Code to hand big reads to Gemini:

When analyzing large codebases or multiple files that might exceed
context limits, use the Gemini CLI with its massive context window.
Use `gemini -p` to leverage Google Gemini's large context capacity.

The examples pass whole directories with @ paths, like gemini -p "@src/ Summarize the architecture of this codebase". The same day it was committed to amantus-ai/vibetunnel as a slash command (documented in that repository), and the heading “Using Gemini CLI for Large Codebase Analysis” reappears in copies across gists and skill collections. Comfy-Org’s prompt library (documented) has Claude estimate the size of the job and ask before handing anything over 200,000 tokens to Gemini, which it then runs in the background with -y, auto-approving every action. Ashik Nesin’s July 2025 write-up (reported) gives the reasons plainly: the free tier and the 1M context.

The free login went to Antigravity CLI on June 18. The window gap closed from the other side: Claude Code’s model configuration page now says current models, Opus 4.7 and later, Sonnet 5 and later, Haiku 5.5 and Fable, “run with the 1M window by default,” and that on the Anthropic API there is “no premium for tokens beyond 200K,” Haiku 5.5 excepted. Gemini CLI budgets 1,048,576 tokens per its source code. Both now sit at roughly a million.

The snippet still runs for a developer with a paid key or an Enterprise license. It now sends the read to a second meter. Whether that meter is cheaper depends on the Gemini model and on how much of your Claude allowance is left, so the reason to do it is a second model’s reading of the repository, with cost as a case-by-case question.

Can Claude Code call Gemini through an MCP server?

Yes, and the most-starred bridge I found shows what a forced migration does to a tool in the middle. jamubc/gemini-mcp-tool (documented in its repository, about 2,300 stars) wraps the Gemini command line as an MCP server, so Claude Code calls it as a tool. Its README pitch: “Use Gemini’s powerful analysis capabilities directly in Claude Code to save tokens and analyze large files.”

On June 18 the maintainer added an Antigravity backend and made it the default: “From 2026-06-18 this tool selects the agy backend automatically.” An enterprise user who updated the package and set nothing else was moved to agy, and has to set GEMINI_MCP_BACKEND=gemini to keep the old behavior. The repository’s migration notes flagged the headline feature, @ file references, as not confirmed to exist in agy. The bridge now inlines those files itself before the call, so the feature survived as the maintainer’s code rather than Google’s. The README calls the agy backend experimental. One more detail before you install it: the license file says “MIT License (Non-Commercial),” which is not the MIT license the badge implies.

The rules changed too. Google’s Antigravity FAQ answers the question this pattern raises:

Antigravity CLI headless mode ( agy -p ) is intended for direct
scripting and CI workflows. Wrapping agy inside third-party coding
agents or applications, or automating account rotation, is not
supported under consumer plans.

The same FAQ names Claude Code among the third-party tools that may not use an Antigravity login, and recommends “a Gemini Enterprise or Google AI Studio API key” instead. Google ships its own route for that: the Gemini API CLI, gemini-api, which the Gemini CLI README describes as giving “agents (agy, Claude Code, Codex), scripts, and CI direct commands for Gemini models.” It is stateless and runs on an API key, and its README tells you to paste an instruction block into CLAUDE.md. MCP servers that call the Gemini API directly, like RLabs-Inc/gemini-mcp, never depended on the CLI login and were not affected.

Can Gemini or Antigravity review the code Claude Code wrote?

People built it both ways, before and after the switch. The earlier version is documented in vibetunnel’s repository: its review-pr command, committed June 29, 2025, runs “parallel sub-agents powered by Claude and Gemini, with results intelligently merged.” It still calls gemini, so on a personal account it now fails.

The newer version is documented too. simplybychris/antigravity-plugin-cc, last updated November 6, adds /agy:review to Claude Code, plus commands to ask, delegate and research through agy. Its README borrows the sentence OpenAI used for the Codex plugin: it is for Claude Code users who want Antigravity “from the workflow they already have.” The author’s post says Claude Code stays where “my git, tests, edits, and agent-assisted work happen.”

Two things separate this from the Codex version of the same split. OpenAI wrote and publishes its plugin, and its review command is read-only by design. Here a third party wrote the plugin, and on a personal Google plan it wraps agy in exactly the way the FAQ above says is not supported. The FAQ says nothing that clears an Enterprise login either, and its advice for third-party agents is an API key. Read it before you install.

Should Claude Code plan while Antigravity writes the code?

Some people run it that way, and the one field report shows where it goes wrong. yuting0624/antigravity-for-claude-code (documented in the repository) was created the day after the switch and splits by role: Claude takes “requirements · architecture · the hard 20% · verification · review,” and agy takes “scaffold · implementation · test generation · search.” The README’s summary: “Claude stays the conductor; the bulk, token-heavy read ran on cheaper Gemini, and Claude verified the result.”

The same author filed the most useful bug report in this whole topic, in gemini-mcp-tool issue 105 (reported, July 9). Since agy 1.1.0 made review-first the default, a headless write task without write permission, per the report:

writes them to agy's own scratch dir (~/.gemini/antigravity-cli/scratch/)
and returns a confident success, while the target workspace stays
untouched.

The fix the report gives is --mode accept-edits, which applies file edits without opening the terminal and other tools. Antigravity’s headless page documents the general case: an action that needs approval it cannot get is “soft-denied: the run continues, exits 0,” with a notice on stderr naming the tool. Gemini CLI treats an approval prompt in headless mode as a denial, and Claude Code’s claude -p now starts in Manual mode and denies what nothing approved, reporting the denials in its result. So the conductor’s verification step is the one part of this split you cannot skip: exit 0 says the run finished, not that your workspace changed. Read the denial output and the diff.

Can Gemini CLI run Claude Code’s web searches?

It was a common split in 2025 and is the easiest one to replace now. mizchi wrote on Zenn in June 2025 (reported) about a Claude Code slash command that sends searches to Gemini’s Google Search tool through gemini --prompt 'WebSearch: <query>'. He had to write “ALWAYS use this for web search instead of builtin Web_Search tool” into it, because Claude kept reaching for its own.

The reason holds. Claude Code’s tools reference says its WebSearch runs on Anthropic’s backend and “is not configurable,” and that a different provider needs an MCP server that exposes a search tool. The mechanism does not hold on a personal account: it is the same gemini call that stopped working there. An API-key MCP server with a Google Search tool does the same job without the CLI and without the terms question.

Do multi-model hubs like PAL still connect Claude Code to Gemini?

The biggest one I found is stale. PAL MCP server, formerly Zen MCP, has about 11,800 stars, the most of any project here. Its repository documents consensus between models, multi-model code review, and clink, which starts other CLIs as subagents: “Claude Code can spawn Codex subagents, Codex can spawn Gemini CLI subagents.” Its last commit was December 15, 2025. Two open issues report the state: fresh installs crash because an unpinned dependency resolves to MCP SDK 2.0 (issue 470), and the Gemini clink client is “effectively dead for individual accounts” (issue 465). Its API-key tools, consensus and code review, still make sense in principle if you pin the dependency yourself.

Here the adoption numbers run backwards. The repositories with thousands of stars, PAL and vibetunnel, still call gemini on the paths described above and have not touched them since June 18. The ones that moved to agy are small and new, some created after the switch. A star count records when a project was popular, not whether it still runs.

How do Claude Code and Gemini CLI share one set of instructions?

With one AGENTS.md and a one-line pointer file for each tool, and this is the split that came through June 18 untouched. Committed files in public repositories show it (documented). Red Hat’s Quay added a CLAUDE.md that reads @AGENTS.md and a GEMINI.md that says “Follow instructions from @./AGENTS.md” in December 2025. Percona’s PMM did the same in August 2026, after the switch, and so did AlmaLinux’s website in July. AlmaLinux’s GEMINI.md explains itself:

This repository's conventions and guardrails live in `AGENTS.md`, the
tool-neutral source of truth that every AI agent follows. The import
below pulls them into Gemini CLI.

It survived because Antigravity CLI kept the file convention. Its migration guide says both tools “use identical workspace context rules” and read GEMINI.md and AGENTS.md. Gemini CLI can also skip the pointer: its context.fileName setting takes a list, and the docs example includes AGENTS.md. Claude Code does not discover GEMINI.md on its own, and it reads AGENTS.md by default only when no CLAUDE.md or CLAUDE.local.md exists, so the import line on the Claude side is what makes the arrangement work.

The rest of the configuration travels through importers, which copy and translate it. Claude Code’s commands reference lists /import gemini, which brings over instruction files, MCP servers, commands, subagents and skills, and /init offers it when it finds Gemini CLI configuration. Going the other way, agy plugin import claude turns installed Claude Code plugins into Antigravity plugins, MCP servers included, and Gemini CLI’s source has gemini hooks migrate --from-claude, which renames events (PreToolUse becomes BeforeTool) and converts timeouts from seconds to milliseconds. That command appears in the v0.67.0 changelog but not in the docs.

What does not travel is a shared file. Skills almost meet: Gemini CLI and Antigravity read .agents/skills/, Claude Code reads .claude/skills/, and one symlink joins them, though Antigravity made everyone move workspace skills out of .gemini/skills/ first. MCP servers have no common file at all. Claude Code uses .mcp.json, and Antigravity moved its servers to .agents/mcp_config.json, where the URL key is now serverUrl.

Which Claude Code and Gemini CLI split should you start with?

This is my read of the sources, not a measurement. Start with the shared AGENTS.md and the two pointer files. It costs nothing and needs neither vendor’s login.

Then decide which Gemini account you are talking about before you copy any snippet. On an Enterprise license or a paid API key the gemini -p patterns still work, as a second reading rather than a reliably cheaper one. On a personal Google plan, route Gemini into Claude Code through an API key: gemini-api, or an MCP server that calls the API. Calling agy from inside Claude Code on that plan is a setup Google’s FAQ says it does not support.

Who should skip the second tool: anyone whose reason was the free login or the bigger window. The login moved to another tool, and the window is the same size on both sides now.

In May I said I want a CLI swap to be a routing decision and not a rebuild. June 18 ran that test on other people’s setups, and the dividing line turned out to be the account behind each call. Files carried on, binaries on a paid key carried on, and everything that leaned on a personal Google login stopped that morning. By star count, those are still the setups people find first. Google, meanwhile, ships Claude Opus 5.5 inside Antigravity on paid plans, per its models page. The vendor that built the replacement did not pick one model either.