Gemini CLI — Google's Terminal Coding Agent

https://github.com/google-gemini/gemini-cli

Why This Matters

Gemini CLI is Google's open-source terminal coding agent — the same shape as Claude Code: natural-language input, planning, tool use, filesystem reads, command execution, git awareness. It launched June 25, 2025 under Apache 2.0, sits at ~103k GitHub stars, and as of May 12, 2026 is on stable v0.42.0 with a weekly release cadence (Tuesdays, stable + preview).

The reason to write this up now, eleven months after the initial launch: the field has consolidated around three serious terminal agents — Claude Code, Gemini CLI, and OpenAI Codex CLI — and the workflow patterns for using more than one have stabilized enough to be worth documenting.

What It Actually Is

Open source. Apache 2.0. Source at google-gemini/gemini-cli. The full agent loop is auditable, unlike Claude Code (which is closed-source despite being a CLI).
Free tier that is actually usable. 1,000 requests/day on a personal Google account via OAuth, 60 req/min throttle. No credit card. This is the part that makes it credible as a daily-driver alternative for individual developers.
Default model: Gemini 3 Pro. Gemini 3.1 Pro available in preview. Scores 80.6% on SWE-bench Verified vs. Claude Opus 4.7's 87.6% — competitive but not at the top of the leaderboard.
MCP support. Yes, the same Model Context Protocol that Claude Code uses. Our existing MCP servers (chroma, tavily, etc.) work in Gemini CLI without changes.
Multimodal via @filename syntax. Drop images or PDFs into a prompt with @design-spec.pdf or @screenshot.png. This is genuinely useful for design-doc work and is something Claude Code does not natively support in the same ergonomic way.
OS-level sandboxing by default. gVisor / LXC. Stronger isolation than Claude Code's permission-prompt model out of the box.
VS Code integration. Via Gemini Code Assist. GitHub Actions integration ships in the box.

What It Doesn't Have

No sub-agents. Single-agent loop only. Claude Code's Agent tool with specialized subagents (Explore, Plan, general-purpose) has no analog. For the kind of work where we routinely fan out to Explore for parallel codebase search, this is a meaningful gap.
No skills system. No /loop, /schedule, /ship, no skill auto-loading based on directory context. The Lossless Group's heavy use of skills (pseudomonorepos, context-vigilance, astro-knots) does not have a Gemini-CLI equivalent.
No 1M-context UI/UX yet. The model supports it; the CLI's session-management surface around long contexts is less polished than Claude Code's.
Closed gap on tool-use quality. Per multiple 2026 comparisons, Claude Code "consistently produces cleaner, more maintainable code, especially for complex multi-file refactors." Gemini CLI is competitive for simpler tasks but not yet for the hardest agent workloads.

The Emerging Multi-Agent Workflow

The pattern that has stabilized across the comparison reviews:

Claude Code for production work — multi-file refactors, anything that touches astro-knots/, anything that needs the Lossless skill suite.
Gemini CLI for free spikes, large-context greps, multimodal tasks (screenshots, PDFs), and prototyping where the 1,000 req/day free tier removes the cost-anxiety of exploration.
Codex CLI as the third option for OpenAI-stack-specific work.

Worth noting: all three accept MCP servers. That means the Chroma corpus we built for Claude Code is also available from Gemini CLI without re-wiring — the corpus pays off across the whole agent fleet.

Recent Release Cadence

Date	Release	Notable
2025-06-25	Initial launch	Apache 2.0, ~103k stars within a year
2026-01-13	Notable updates landed (per release notes)	—
2026-05-12	v0.42.0 stable	Voice mode polish, session management (`/exit --delete`), reduced API timeouts
2026-05-12	v0.43.0-preview.0	Weekly preview cadence

New preview every Tuesday 23 UTC; new stable every Tuesday 20 UTC. Weekly release cadence on an OSS terminal agent is unusual — most comparable tools ship monthly at best. It's the most visible signal that Google is treating this as a strategic surface, not a side project.

Where It Fits in Our Workflow

Worth installing alongside Claude Code, not instead of it. The free tier alone makes it useful for exploration that would otherwise hit our Anthropic API budget.
Multimodal handling is the wedge. When the next deck-iteration or crawl-fetch-ingest job involves screenshots, partner logos, or PDF decks, Gemini CLI's @filename syntax is a faster path than the Claude Code workflow of base64-encoding and pasting.
Don't expect skill parity. The Lossless skill ecosystem (auto-loaded pseudomonorepos, context-vigilance, etc.) is Claude Code-specific. Gemini CLI sessions will need explicit prompting for the conventions that Claude Code picks up from CLAUDE.md automatically.
MCP-server compatibility means tooling investment pays off twice. Anything we build as an MCP server (and we have several) is usable from both agents.

How Significant Is This Compared to Other Terminal Agents?

Agent	Launch	License	Free Tier	MCP	Sub-agents	SWE-bench Verified
Claude Code	Feb 2025 (preview) → May 2025 GA	Closed	Limited	Yes	Yes	87.6% (Opus 4.7)
Gemini CLI	Jun 2025	Apache 2.0	1,000 req/day	Yes	No	80.6% (3.1 Pro)
Codex CLI	2025	Open source	Limited	Yes	Partial	~78% (GPT-5.5)

The Gemini CLI value proposition is: open source + generous free tier + multimodal-first. The Claude Code value proposition is: best code quality + richest agent harness (sub-agents, skills, hooks) + tightest IDE/MCP integration. They are competing on different axes; the practical answer is to use both.