You spin up Gemini CLI on a Monday, wire it into your shell, get 1,000 free daily requests against a 1M-token context window, and call it your new pairing tool. By mid-June, the free tier is gone. The replacement - Antigravity CLI - has no public source code.
Google announced it was retiring the 104k-star Gemini CLI in June and replacing it with a closed-source successor. That switch happened on June 18, 2026. Before the shutdown, the CLI was the only agentic coding tool running Gemini models at zero cost with the largest context window in the category. The repo still exists. The Apache 2.0 license still applies to the code. But the serving is gone, and the replacement is not open. That gap - between "open source" and "actually free to use with the model attached" - is the real story of the terminal coding agent space right now.
What "open source" actually means for a coding agent
A coding agent has two separable parts: the harness (the code that drives the loop, reads files, calls tools) and the model (the thing doing the reasoning). An Apache 2.0 harness does not make the model free. Most coverage collapses this distinction, which leads teams to pick tools based on star counts and then hit an inference bill they did not plan for.
Gemini CLI offered 60 requests per minute and 1,000 requests per day on a personal Google account, with access to Gemini 3 models, a 1M-token context window, built-in Google Search grounding, and MCP support. That was genuinely rare: an open harness and free model access bundled together. Gemini CLI's free tier was the headline - OAuth login on a personal Google account gave 60 requests per minute and 1,000 requests per day. When the free serving ended, the harness remained open but the actual value proposition vanished.
OpenCode is free as software under MIT. Your only cost is the model provider you connect. Point it at a cheap local model through Ollama and you pay nothing for inference; point it at a frontier API and you pay that provider's per-token rate. That is a cleaner separation - but it means the "free" framing still depends on what you plug in.
The governance fractures nobody is writing about
Open source coding agents had a wild first half of 2026. The licensing and governance picture shifted faster than most teams noticed:
- Claw Code (195k stars) is a clean-room Python/Rust rewrite of Claude Code architecture - born from the March 2026 Claude Code source leak, MIT licensed.
- Goose was donated by Block to the Linux Foundation's new Agentic AI Foundation in April, making it vendor-neutral and foundation-governed.
- OpenCode lost Claude Pro/Max subscription login after an Anthropic dispute in January 2026; you can still use Claude with a raw API key, and ChatGPT subscriptions work natively.
- Roo Code, at 24k stars, archived itself.
- Pi - from Armin Ronacher, the person behind Flask and Jinja2 - crossed 54k stars by running on a system prompt under 1,000 tokens.
The non-obvious read: open-source coding agents are bifurcating. One branch is vendor-controlled open source (Apache from Google or OpenAI, with the model access terms set by the same vendor). The other branch is genuinely community-governed - Linux Foundation Goose, MIT-licensed OpenCode, Pi. Which branch you're on matters when a vendor changes the free tier.
How the benchmark numbers actually compare
Terminal-Bench 2.1 measures an agent driving a real terminal to complete development tasks: editing files, running commands, fixing failures. It tests the agent and model together - the same model scores differently inside different agents. That's the right framing. A harness is not a model, and star counts are not benchmark scores.
| Agent | Model (default) | Terminal-Bench 2.1 | License | Free inference? |
|---|---|---|---|---|
| Codex CLI | GPT-5.6 Sol | 89.5% | Apache 2.0 | No (API key) |
| Claude Code | Opus 5 | 89.1% | Proprietary | No (subscription) |
| Gemini CLI | Gemini 3.1 Pro | 70.7% | Apache 2.0 | Ended June 18 |
| OpenCode | Model-agnostic | Ceiling = backend | MIT | No (BYOK) |
| Claw Code | Any | Not yet published | MIT | No (BYOK) |
Gemini CLI with Gemini 3.1 Pro scores 70.7% (±2.9) on Terminal-Bench 2.1.
OpenCode can connect Opus 4.8 or GPT-5.5 and reach the higher entries, but you pay those providers' API rates. The practical ceiling for a zero-cost setup is now lower than it was before June - local models through Ollama are the only genuinely free path, and they trail frontier models by a wide margin on this benchmark.
What to actually run today
If you are picking a terminal coding agent right now, the decision is not really about stars or licenses - it is about what model access you already have and what governance risk you can tolerate.
Google Search grounding let Gemini CLI pull live documentation and API references during execution without extra configuration. A GEMINI.md file at the project root told the agent how the codebase was structured and which test commands to run. That pattern - a project-root config file telling the agent the local context - is now standard. It's the same pattern OpenAI Codex uses with AGENTS.md. Whatever harness you land on, that file is worth writing.
For teams that want genuine source access and no single-vendor risk: OpenCode (MIT, BYOK) or Linux Foundation Goose. For teams on Azure or .NET, the Microsoft Agent Framework, which reached 1.0 GA on April 2, 2026, unifies AutoGen and Semantic Kernel into a single supported platform
- though that is a framework for building agents, not a CLI coding agent itself.
A teammate like Beagle can sit inside the Slack channels where your engineering team discusses which tools to adopt - surfacing the relevant changelog, benchmarks, or license change when someone asks - so the decision lands on the right desk rather than dying in a thread.
Beagle in action#platform-eng, 10:42amOpen source terminal coding agents: common questions
What is the best free open source terminal coding agent right now?
As of August 2026, there is no single agent that is both genuinely free and frontier-model-quality. Gemini CLI's free tier ended June 18, 2026. OpenCode (MIT) and Goose (Linux Foundation) are the most open harnesses, but both require a paid model API key or a local model to run. Local models via Ollama are free but trail frontier scores on Terminal-Bench 2.1 by 20+ points.
Is Gemini CLI still open source?
The Gemini CLI code remains Apache 2.0 licensed on GitHub. The free model-serving tier ended June 18, 2026, replaced by Antigravity CLI, which is closed source. You can still use the Gemini CLI harness with an API key at Gemini's standard rates - you just lose the 1,000 free daily requests.
How does OpenCode compare to Gemini CLI?
OpenCode is model-agnostic - its benchmark ceiling is set by the backend you connect, not the agent itself. It is the most-starred open-source coding agent, at 172,198 stars, MIT licensed. Gemini CLI is tied to Gemini models; OpenCode routes to any provider. OpenCode's trade-off: no bundled free inference.
What happened to Roo Code and Goose?
Block donated Goose to the Linux Foundation's new Agentic AI Foundation in April, making it vendor-neutral. Roo Code, at 24k stars, archived itself. Goose is still active under foundation governance; Roo Code is not.
What is Claw Code?
Claw Code is a clean-room Python/Rust rewrite of the Claude Code architecture, born from the March 2026 Claude Code source leak, and carries a 195k-star GitHub count - reportedly the fastest repo in GitHub history to reach 100k stars. It is MIT licensed and model-agnostic, but has no published Terminal-Bench score yet.