Gemini CLI — Google's Open-Source Terminal Agent

Source: ai-research/gemini-cli-github-readme-2026-07-03.md, ai-research/gemini-cli-vs-claude-code-datacamp-2026-07-03.md, ai-research/gemini-cli-quota-pricing-docs-2026-07-03.md, ai-research/gemini-cli-service-update-github-discussion-22970-2026-07-03.md, ai-research/gemini-cli-to-antigravity-google-blog-2026-07-03.md, ai-research/gemini-cli-stopped-serving-individuals-github-discussion-28017-2026-07-03.md, ai-research/gemini-cli-antigravity-migration-digitalapplied-2026-07-03.md

Gemini CLI is Google’s open-source AI agent for the terminal — “an open-source AI agent that brings the power of Gemini directly into your terminal” — the closest direct analog to Claude Code from the Google side, and the tool this wiki has referenced in passing across many articles (Hyperframes, Open Design, agent bundles) without a dedicated page. It launched in June 2025 with an unusually generous free tier, a ReAct agent loop, built-in tools, MCP support, an extensions system, and GEMINI.md context files, all under Apache 2.0. Its story took a sharp turn in 2026: Google progressively restricted the free tier (March 25, 2026) and then, on June 18, 2026, stopped serving Gemini CLI requests for all individual, Google AI Pro, and Google AI Ultra accounts — steering those users to a new closed-source Go rewrite called Antigravity CLI (agy). This article documents both what Gemini CLI is and where it now stands, written for a Claude Code-first audience evaluating open-source terminal agents.

Key Takeaways

• What it is: an open-source (Apache 2.0), terminal-native AI coding agent from Google, npm-installed as @google/gemini-cli, invoked with the gemini command. ~106,000 stars / ~14,200 forks at fetch (2026-07-03) — one of the most-starred agent repos on GitHub.
• Models: Google’s Gemini 3 family — defaulting to Gemini 3 Flash, with Gemini 3.1 Pro for harder tasks ^[ambiguous] (point-version naming varies across sources; the wiki’s Cursor roster lists “Gemini 3.1 Pro / Gemini 3.5 Flash”). 1M-token context window, ~64K max output.
• The launch hook was the free tier: sign in with a personal Google account for 60 requests/min and 1,000 requests/day, no credit card — materially more generous than Claude Code, which has no free tier.
• Core features: built-in tools (Google Search grounding, file ops, shell, web fetch), MCP support, an Extensions system (90+ installable packages bundling MCP servers + context files + slash commands as of late 2025), GEMINI.md project-memory files, a ReAct loop over a PTY (pseudo-terminal) shell, conversation checkpointing (/restore, --checkpointing), token caching, and headless JSON/stream-JSON output modes.
• Two 2026 restrictions changed the calculus: (1) March 25, 2026 — Gemini Pro models became paid-only; free-tier users limited to Flash; abuse detection + license-based traffic prioritization added. (2) June 18, 2026 — Gemini CLI (and the Gemini Code Assist IDE extensions) stopped serving free-individual, AI Pro, and AI Ultra accounts entirely.
• What survives June 18: the open-source repo stays public; enterprise (Gemini Code Assist Standard/Enterprise) access is unaffected; and per Google, “Gemini CLI will remain accessible via paid Gemini and Gemini Enterprise Agent Platform API keys.” Individuals are pushed to Antigravity CLI — closed-source, Go, no feature parity at launch.
• Bottom line for an operator today: the free “sign in with Google, 1,000 req/day” pitch that made Gemini CLI attractive is gone for individuals as of June 18, 2026. If you want a free open-source terminal agent, you now either run Gemini CLI with a paid API key, adopt Antigravity CLI (closed-source), or look at other open agents. Verify current state before relying on any specific tier — this is a fast-moving situation.

What Gemini CLI Is

Gemini CLI brings an agentic coding assistant into the terminal, the same form factor as Claude Code: it can search a codebase, read and edit files, run shell commands, fetch the web, and carry out multi-step tasks from natural-language instructions.

• Agent loop: a ReAct (reason + act) pattern with environmental feedback. DataCamp’s teardown notes Gemini CLI runs its shell through a PTY (pseudo-terminal) that “spawns a virtual terminal in the background, takes snapshots of the terminal state, and renders the output inline” — which lets interactive and input-blocking scripts run without hanging the terminal.
• Context files: GEMINI.md files tailor the agent’s behavior per project — Google’s equivalent of Claude Code’s CLAUDE.md convention. ^[inferred]
• Open source: Apache 2.0 licensed, developed in the open on GitHub. The Register-class coverage of the 2026 transition noted Google had accepted thousands of community pull requests into the codebase before the individual-access change — part of why the pivot was controversial.
• Auth (three paths): (1) Google OAuth “Sign in with Google,” (2) a Gemini API key (from Google AI Studio), (3) Vertex AI with a Google Cloud project.

Core Features

• Built-in tools — Google Search grounding (native web-grounded answers), file operations, shell command execution, and web fetching, out of the box.
• MCP support — Gemini CLI is an MCP client; it speaks the same Model Context Protocol that Claude Code and Cursor speak, so many MCP servers built for one work across all three.
• Extensions system — bundles MCP servers, context files, and slash commands into installable packages; “over 90 available as of late 2025,” with integrations including Figma, Stripe, Elastic, Postman, and Snyk.
• Slash commands — 40+, including /plan, /memory, /skills, /restore, and /theme. Plan Mode is default-on (read-only; auto-routes to Pro for planning).
• Multimodal @ input — reference files, images, PDFs, and audio inline with @ syntax.
• Checkpointing & rollback — /restore for file checkpoints (requires the --checkpointing flag); /rewind (Esc+Esc) for conversation state.
• Headless / scripting — non-interactive modes with --output-format json and --output-format stream-json for automation pipelines.
• Token caching — to reduce repeated-context cost.

Models, Free Tier, and the 2026 Restrictions

Gemini CLI’s defining feature at launch was its free tier, and its defining event in 2026 was that tier’s removal for individuals. The arc:

Date	Change
Jun 2025 (launch)	Free via personal Google account: 60 req/min, 1,000 req/day, no card. API-key path: 1,000 req/day with Gemini 3 (mix of Flash + Pro).
Mar 25, 2026	Gemini Pro models become paid-only; free-tier users limited to Flash models. Abuse detection added (e.g., flagging OAuth use via third-party software); traffic prioritized by “license type and account standing.” Community reports free-Flash access dropped to ~150 req/day.
Jun 18, 2026	Gemini CLI + Gemini Code Assist IDE extensions stop serving Google AI Pro, Google AI Ultra, and free-individual (Gemini Code Assist for individuals) accounts. Individuals steered to Antigravity CLI.

Documented quotas (2026 docs, before/around the June cutoff): Google-account individual = 1,000 req/day; unpaid Gemini API key = 250 req/day, Flash only; Google AI Pro = 1,500/day; Ultra = 2,000/day; Workspace Standard = 1,500/day; Enterprise = 2,000/day. Free-tier data “may be used for model improvement” (a privacy consideration Claude Code’s paid tiers don’t carry). ^[inferred]

Models: the Gemini 3 family — Gemini 3 Flash as the default with Gemini 3.1 Pro for demanding tasks (DataCamp); the README references “Gemini 3 models” and a selectable gemini-2.5-flash, and the -m flag lets you pick a model. 1M-token context, ~64K max output. Exact Flash point-versions vary across sources (3 Flash vs 3.5 Flash) — treat the specific number as a moving target. ^[ambiguous]

The June 18, 2026 Antigravity CLI Transition

The most consequential Gemini CLI fact as of this writing, corroborated by the Google Developers Blog, two GitHub discussions (#27274, #28017), and multiple secondary outlets.

• What happened: “On June 18, 2026, Gemini CLI and Gemini Code Assist IDE extensions will stop serving requests for Google AI Pro and Ultra, as well as those using it free of charge using Gemini Code Assist for individuals” (Google Developers Blog). Discussion #28017 confirms: “Gemini CLI will stop serving requests for Google AI Pro, Google AI Ultra, and free tier individual accounts.”
• What still works: “Enterprise users with Gemini Code Assist licenses and API key authentication remain completely unaffected.” Google adds that “Gemini CLI will remain accessible via paid Gemini and Gemini Enterprise Agent Platform API keys.” So the tool is not dead — its free individual on-ramp is.
• The replacement — Antigravity CLI: Google’s new terminal interface for its “agent-first development platform,” invoked as agy, written in Go (a rewrite of the Node.js codebase), adding Agent Skills, Hooks, Subagents, and plugins. It auto-imports existing config: skills, MCP servers, agent profiles, and gemini.md memory files.
• Closed-source, no parity: unlike Gemini CLI, Antigravity CLI is closed-source — “its public GitHub repo holds only a changelog, readme, and a GIF — no application code.” At launch it lacks feature parity (e.g., no ACP stdio mode gemini --acp for orchestration bridges; no custom-theme migration). The original repo “remains available under Apache 2.0, but the ‘still open’ original is a backend-dependent skeleton” — “you can read the code; you cannot run the tool without Google’s infrastructure.”
• Why it was controversial: Google built Gemini CLI in the open, merged thousands of community PRs, then closed individual access and shipped a proprietary successor — a reversal that drew criticism from contributors. ^[inferred]

Gemini CLI vs. Claude Code

The comparison this wiki’s readers care about. Drawn primarily from DataCamp’s 2026 head-to-head, plus the pricing/transition sources above.

	Gemini CLI	Claude Code
Vendor / license	Google · open source (Apache 2.0)	Anthropic · proprietary
Default model	Gemini 3 Flash (Gemini 3.1 Pro for hard tasks)	Sonnet 4.6 (Pro); Opus 4.6 (Max)
Context / output	1M / ~64K	1M / 128K (Opus 4.6, Max), 64K (Sonnet 4.6, Pro)
Free tier	Yes at launch (1,000 req/day) — removed for individuals Jun 18, 2026	None — $20/mo Pro minimum
Entry paid	Google AI Pro ~$20/mo	Claude Pro $20/mo
Heavy use	Google AI Ultra ~$250/mo	Claude Max 5x $100/mo
Model choice	Gemini family only	Anthropic models only
Agent loop	ReAct over a PTY shell	Multi-agent orchestration (v2.0, Dec 2025); diff-based edits
Config file	GEMINI.md	CLAUDE.md
MCP	Yes (client) + Extensions system (90+)	Yes (client)

Where each pulls ahead (DataCamp):

• Gemini CLI — free starting point (historically), open source you can read/fork, “responds faster” on simple tasks, strong for interactive scripts and multimodal inputs (images/PDFs), and a natural fit for Google Cloud / Workspace workflows.
• Claude Code — “more consistent output on complex codebases,” “stronger multi-file reasoning,” better at matching existing project conventions, a more structured permission model, and the Agent SDK for automation. Its cited weaknesses: no free tier, approval prompts before each write/command, and Anthropic-models-only.
• Efficiency note: DataCamp measured Gemini CLI using 432K input tokens versus Claude Code’s 261K for the same task, and flagged “tool call errors and retry loops on complex multi-file tasks” as a common Gemini CLI complaint.

The operator pattern most sources land on: run both — “Gemini CLI for quick exploration and planning, then Claude Code when the task needs more precision and consistency.” The catch as of mid-2026: the free-Gemini-CLI half of that pattern no longer exists for individuals, which materially weakens the “free exploration tool” argument. ^[inferred]

Open Questions

• Does the unpaid Gemini API-key tier still work with the open-source CLI after June 18, 2026? The docs listed a 250 req/day Flash-only unpaid API-key path, but the transition announcements only confirmed paid Gemini / Enterprise Agent Platform API keys and enterprise licenses as unaffected. The status of the free API-key route through the open CLI is unconfirmed in this research pass.
• Antigravity CLI licensing and roadmap. It launched closed-source with no feature parity; whether Google restores an open-source or free-individual path, and how fast parity closes, is unresolved. No dedicated Antigravity article exists in this wiki yet — a candidate for a future ingest.
• Exact current model lineup and quotas. Point-versions (Gemini 3 Flash vs 3.5 Flash) and per-tier request/token limits shifted twice in 2026 and vary across sources; confirm at the official docs before quoting a specific number.

Try It

The open-source CLI still installs and runs (with a valid API key / enterprise license). To evaluate it against Claude Code:

1. Install: npm install -g @google/gemini-cli (or run once with npx @google/gemini-cli; Homebrew: brew install gemini-cli).
2. First run: type gemini in a project directory to start the interactive agent. Given the June 18, 2026 change, prefer authenticating with a Gemini API key (Google AI Studio) or Vertex AI rather than the individual “Sign in with Google” path, which no longer serves free/Pro/Ultra individuals.
3. Add project memory: drop a GEMINI.md in your repo root to give the agent standing instructions — the direct analog to CLAUDE.md.
4. Try the differentiators: use @ to attach an image or PDF (multimodal), enable --checkpointing and test /restore, or wire up an MCP server you already use with Claude Code and confirm it works unchanged.
5. If you’re an individual who wants Google’s current terminal agent, note that Google now points you at Antigravity CLI (agy, closed-source) — install via curl -fsSL https://antigravity.google/cli/install.sh | bash; it auto-imports your existing skills, MCP servers, and gemini.md files. Evaluate the closed-source tradeoff deliberately.

Related

• Cursor — AI-Native Code Editor (Anysphere) — the sibling competing-tools entry; Cursor runs Gemini 3.1 Pro / 3.5 Flash among its model roster, so it’s one way to use Gemini models without the standalone CLI.
• Claude Code CLI Reference — the Anthropic terminal agent Gemini CLI is most directly compared against; read alongside this article’s comparison table.
• MCP — Model Context Protocol — the shared open protocol Gemini CLI, Claude Code, and Cursor all speak as clients; MCP servers built for one generally work across all three.
• GLM-5.2 (Z.ai) — another open-weight/open-source entrant at the agentic-coding frontier; relevant when weighing open vs. proprietary terminal agents.
• Agent Guardrails — Hooks, Permissions, and Sandboxing — Gemini CLI’s checkpointing and Plan Mode vs. Claude Code’s approval-first permission model are two points on the same guardrail spectrum.
• Claude Fable 5 and Claude Mythos 5 — the frontier Anthropic model that anchors the Claude Code side of the comparison.