June 2025, Revisited: Gemini CLI and the Month the Terminal Agent Went Mainstream

This post is part of our Three-Year Retrospective series: thirty-six posts, one per month, looking back at what actually mattered in software engineering. This one covers June 2025.

On June 25, 2025, Google released Gemini CLI: an open-source AI agent that ran in your terminal, talked to Gemini 2.5 Pro with its million-token context window, and shipped under the Apache 2.0 license. Claude Code had already proven that developers would happily live in a terminal agent; OpenAI's Codex CLI had followed in the spring. But Gemini CLI arrived with two things neither competitor had offered together: a fully open-source codebase and a free tier — 60 requests per minute, 1,000 requests per day with a personal Google account — generous enough that most individual developers would never hit it.

Free changed the adoption math

We want to be careful here, because we generally avoid talking about pricing in these retrospectives. But the structure of the offer mattered in a way the numbers didn't. Every previous terminal agent had required an API key, a billing relationship, or a subscription before you could type your first prompt. Gemini CLI required a Google account. That removed the single biggest barrier we saw in our trainings at the time: the developer who was curious about agentic coding but couldn't justify a procurement conversation to find out whether it was useful.

By July, we were watching classrooms full of people run their first real agentic session — explain this codebase, write this feature, fix this failing test — on tooling they'd installed five minutes earlier with npx. The conversion from skeptic to practitioner usually took about twenty minutes. The objection that AI coding tools were "something other companies use" died quietly that summer, one terminal session at a time.

Open source as a strategy, not a gesture

The Apache 2.0 license was the other half of the story. Gemini CLI's code was on GitHub from day one, and the community treated it accordingly: issues, pull requests, forks, and — within months — extensions and MCP server integrations the core team had never planned. Built-in Google Search grounding, file operations, shell execution, and web fetching gave it a respectable toolkit out of the box, and Model Context Protocol support meant the toolkit wasn't a ceiling.

For the industry, this was Google declaring that the terminal agent was now table stakes — infrastructure to be commoditized rather than a product to be defended. Anyone watching could read the implication: the durable differentiator would be the model and the ecosystem, not the loop that calls it. That framing held up remarkably well over the following year.

What it meant for how teams worked

June 2025 is the month we date the shift in our own curriculum from "AI-assisted coding" to "agentic coding" as the default framing. With three credible terminal agents available — Claude Code, Codex CLI, and now Gemini CLI — the interesting questions stopped being about which autocomplete was smartest and became operational: how do you scope a task for an agent, how do you review what it did, how do you keep it from wandering, what context do you hand it and how.

The terminal turned out to be the right venue for learning those skills. There's no UI magic in a CLI session — you see every file read, every command run, every diff proposed. Teams that learned agentic habits in the terminal in mid-2025 found that the habits transferred everywhere else.

Looking back from June 2026

A year on, Gemini CLI is a fixture. The project kept shipping — Gemini 3-era models eventually landed in the same tool, and the extension ecosystem grew into something the original announcement only hinted at. The bigger legacy is the category itself: the terminal agent went from novelty in early 2025 to the assumed baseline of professional development by 2026, and the free, open-source entry point Google built in June 2025 did a lot of the work of getting the long tail of developers through the door. The race it accelerated — every major lab shipping and iterating its own agent loop — defined the year that followed.

If your team is still figuring out how to put terminal agents to real work, Building with Coding Agents covers scoping, context, and review workflows across Gemini CLI, Claude Code, and Codex, and AI-Assisted Development with Python puts those habits to work in a language-specific, hands-on setting.