AGENT HARNESS · TERMINAL AI · 2026
Gemini CLI is discontinued: what actually still works
Most “Gemini CLI vs Claude Code” roundups still lead with a free tier that no longer exists. The facts: Google announced the Antigravity CLI transition on May 19, 2026, and on June 18, 2026 Gemini CLI stopped serving requests for free, Google AI Pro, and Google AI Ultra accounts. The repo is still open source, the tool still runs — but the consumer quota behind it is gone. Here is the state of play: what still works, what the successor changes, and how to think about agent choice when agents this popular can be switched off.
The timeline, from Google's own announcement
- June 2025 — Gemini CLI launches: Apache 2.0, 1,000 model requests/day and 60/minute free on a personal Google account, 1M-token context. It becomes the most-starred AI coding CLI on GitHub (~107K stars, ahead of Codex CLI's ~73K) largely because of that free tier.
- March 2026 — the free tier narrows: a repository service update limits free users to Gemini Flash models; Pro models move behind paid plans. Plan Mode becomes the default around v0.34.
- May 19, 2026 — Google announces the transition to Antigravity CLI (Go-based, async background workflows, unified with the Antigravity 2.0 desktop app) and sets the date.
- June 18, 2026 — Gemini CLI stops serving requests for free, Pro, and Ultra accounts. Enterprise access via Gemini Code Assist Standard/Enterprise licenses is explicitly unaffected, and the CLI keeps working with paid Gemini API keys.
- Ongoing — the repository remains Apache 2.0 and maintained (model releases, bugs, security fixes for enterprise customers); the community keeps forking it, Qwen Code being the most notable.
If you are reading an article that still advertises “1,000 requests per day, no credit card” as a current Gemini CLI feature — that article predates June 2026. The 1,000/day tier was real, and it is the single biggest thing the transition took away.
What still works after June 18, 2026
| Route | What you get | Catch |
|---|---|---|
| Paid Gemini API key | Gemini CLI keeps working against your paid key — the documented continuation path | Pay-per-use billing; no more hosted free quota |
| Unpaid AI Studio key | The CLI still runs: ~250 model requests/day, Flash models only, 10/minute | A tenth of the old free tier, and a model request is one API call — one prompt can trigger dozens |
| Enterprise licenses | Gemini Code Assist Standard/Enterprise and Google Cloud customers: access unchanged | Organization-only; procurement path |
| Build from source | The Apache 2.0 repo is intact — fork it, patch it, point it at any Gemini-capable endpoint | You own the maintenance; Qwen Code already took this road |
| Antigravity CLI | Google's successor: free consumer access lives here now | Closed source, tighter limits, new privacy terms — see below |
The counterintuitive summary: the binary still installs and runs; what was discontinued is the free ride. Open code, closed access. Whether that still counts as “the free AI coding agent” depends entirely on which of the five rows above you can actually use.
Antigravity CLI: the successor, and its strings
Antigravity CLI is built in Go (snappier than the Node-based original — a long-standing community complaint), supports asynchronous background workflows for long tasks, shares an architecture with the Antigravity 2.0 desktop app, and carries over the features that mattered: Agent Skills, Hooks, Subagents, and Extensions (now “plugins”). Migration tooling imports existing setups.
The strings are worth reading before you move, because they are the reason the transition stung in the community threads:
| Aspect | Gemini CLI (legacy) | Antigravity CLI |
|---|---|---|
| License | Apache 2.0, source on GitHub | Closed source |
| Consumer free access | Was 1,000 req/day (ended June 18, 2026) | Yes, with tighter limits — community reports quota burn an order of magnitude faster, a shared token pool across models |
| Data handling | Consumer Gemini terms | Consumer tiers: interactions recorded and used to improve Google services including ML, with no independent do-not-train toggle (telemetry-off is not a documented guarantee). Workspace/GCP sessions are explicitly excluded from collection — the clearly documented non-training path, under Cloud terms with abuse-monitoring logging |
| History controls | Session resume, checkpoints | Community reports no chat-history compression; long sessions drain the shared pool faster |
| Skills / hooks / subagents | Yes | Yes (Extensions renamed to plugins) |
If you stay on Google's rails, the decision is really enterprise-vs-consumer: a Workspace or GCP binding gives you the documented no-training boundary (and consumption billing); consumer Antigravity tiers trade openness and predictability for continued free access. Neither is wrong — but decide with the terms in front of you, not the launch blog post.
Gemini CLI vs Claude Code, after the free tier died
The honest 2026 comparison is narrower than the roundups suggest, because the pricing column that decided most matchups is gone:
| Dimension | Gemini CLI (2026) | Claude Code |
|---|---|---|
| Benchmark accuracy | Gemini 3 Flash ~75.4% SWE-bench Verified (Mar 2026) | Claude Sonnet 4.6 ~75.2% — effectively tied |
| Hands-on speed | ~2 min 36 s on a standard build task (3-run average) | ~1 min 45 s on the same task, more production-ready output per independent tests |
| Context window | 1M tokens | 1M tokens (GA at standard pricing since March 2026) |
| Free tier | Ended for individuals June 18, 2026 | None — $20/mo Pro entry |
| License | Apache 2.0 (code), consumer access discontinued | Proprietary |
| Extensibility | MCP servers, extensions, skills, TOML custom commands | MCP, skills, hooks, subagents, Agent Teams |
Model quality is a coin flip; polish and paid reliability favor Claude Code; openness favors the Gemini CLI repo — but only if you bring your own key or your own build. The 2025-era advice “start free with Gemini CLI, upgrade to Claude Code when you outgrow it” no longer parses as written, because the starting position changed under everyone's feet.
The durable lesson: agents are perishable, harnesses persist
Step back from the vendor news and the pattern is stark. In under eighteen months, a top-three terminal agent — the most-starred AI coding CLI on GitHub, with the industry's largest free allowance — had its consumer access switched off by its own maintainer. Every capability built inside that agent (the model routing, the quota, the hosted conveniences) evaporated overnight. The things that survived were the things users owned: their scripts, their MCP servers, their GEMINI.md files, their skills, their forks.
That is the argument for keeping your investment in the harness rather than the agent. VibeFuse is built on that bet: the FIRST EVER free widget-based AI harness, where agents are swappable components on a desktop canvas — if an agent dies, is discontinued, or changes its terms, the widgets, skills, and workflows around it keep running and you plug in a different agent. It runs locally and offline on Windows, works in any app, and its open-source marketplace pays creators revenue share for widgets, skills, and styling packs — community assets no single vendor can switch off.
Concrete migration notes for displaced Gemini CLI users: your GEMINI.md project files and TOML custom commands translate conceptually to any harness with project-context files and custom commands; your MCP server configs carry over as-is (MCP is protocol-level, not vendor-level); and your skills are portable markdown-plus-scripts. The one thing you cannot migrate is the quota you were relying on — price your new setup by requests, not by months, until you know your real usage.
Related reading
What an agent harness actually is
Harness vs framework, the Microsoft Agent Framework harness, and why the runtime layer outlives agents.
Harness vs frameworkAgent harnesses compared
The 2026 field of agent harnesses evaluated — including which are open source.
Best agent harnessRun agents in parallel
Worktree isolation, orchestrators, and the multi-agent workflow on Windows.
Parallel coding agentsSkills and MCP, compared
When a skill (portable procedure) beats an MCP server (live access) — and how they compose.
Skills vs MCPExplore related AI note taking guides
Gemini CLI — FAQ
Is Gemini CLI discontinued?
Partially — the consumer access is, the code is not. On June 18, 2026, Gemini CLI stopped serving requests for free, Google AI Pro, and Google AI Ultra accounts, per Google's own transition announcement. Enterprise access via Gemini Code Assist Standard/Enterprise licenses is unchanged, the CLI keeps working with paid Gemini API keys, and the Apache 2.0 repository remains open and maintained. The successor for individual users is the closed-source Antigravity CLI.
What happened to the free 1,000 requests per day?
It ended. The 1,000 requests/day and 60/minute free tier on a personal Google account was Gemini CLI's headline feature from June 2025 until the transition. A March 2026 update first limited free users to Flash models, and the June 18, 2026 cutoff removed consumer access entirely — the free allowance now lives in the Antigravity CLI instead, with tighter limits. Many roundups still describe the old tier as current; they predate the change.
Does Gemini CLI still work at all?
Yes, on four routes: with a paid Gemini API key (the documented continuation path), with an unpaid AI Studio key (~250 model requests/day, Flash only, 10/minute), on an enterprise Gemini Code Assist or Google Cloud license, and as a from-source build you point at your own endpoint. The binary still installs and runs — what was discontinued is the hosted consumer quota, not the tool.
What is Antigravity CLI?
Google's successor to Gemini CLI, announced May 19, 2026: a Go-based terminal agent with asynchronous background workflows, unified with the Antigravity 2.0 desktop app, carrying over Agent Skills, Hooks, Subagents, and Extensions (renamed plugins). It is closed source and takes over the free, Pro, and Ultra tiers Gemini CLI served. Community reports note tighter shared token limits and no chat-history compression; Workspace/GCP-bound sessions get the documented no-training data boundary.
Gemini CLI or Claude Code — which should I use in 2026?
The pricing column no longer decides it. On accuracy the models are effectively tied (~75.4% vs ~75.2% SWE-bench Verified in March 2026); independent hands-on tests favor Claude Code for speed and production-ready output; the Gemini CLI repo stays Apache 2.0 if you bring your own key or build from source. Claude Code has no free tier ($20/mo Pro entry). If openness matters, fork or self-host; if paid reliability and polish matter, Claude Code; if you want Google's current consumer path, that is Antigravity CLI.
Can I still self-host or fork Gemini CLI?
Yes — the repository remains Apache 2.0 with no changes, and Google states it continues model-release, bug, and security updates for enterprise customers. Forking is a real path (Qwen Code is the most notable fork), and a from-source build can be pointed at any Gemini-capable endpoint including paid API keys. What you inherit is the maintenance burden: model migrations, breakage fixes, and security patches are now yours.
Is Gemini CLI better than Claude Code because it is open source?
They are open in different senses, and neither is fully open end to end. Gemini CLI: open code (Apache 2.0, forkable, the most-starred AI coding CLI at ~107K stars), but its consumer access is controlled by Google and its successor is closed source. Claude Code: proprietary code throughout. Openness of the code does not guarantee openness of the service — the June 2026 cutoff proved a hosted free tier can vanish while the repo stays up.
How do I migrate off Gemini CLI?
Inventory what you actually own: GEMINI.md project files and TOML custom commands translate to any harness with project-context and custom-command features; MCP server configs carry over as-is because MCP is protocol-level, not vendor-level; skills are portable markdown-plus-scripts. Run Google's migration guide if you are moving to Antigravity. The one thing you cannot migrate is the quota you were relying on — budget your new setup by requests until you know your real usage.