Open Source Should be Non-Negotiable in the Agentic CLI

TLDR: The terminal is where developer workflows ossify into shell configs, CI pipelines, and dotfiles, and that makes the open source question for AI CLI agents a structural concern, not a preference. The Gemini CLI to Antigravity CLI transition is a clean case study in what "open at one layer, closed at three" actually costs developers. Real openness requires source, model choice, endpoint, and pricing to all be accessible.

The terminal is one of the most personal interfaces in software development. Your dotfiles live there. Your aliases, your Git muscle memory from three jobs ago, the small rituals that don't show up in any documentation. And in 2026, the AI lives there too. Claude Code, Codex CLI, Aider, Kilo CLI, the terminal has become one of the most interesting venues for AI coding agents, and it's where a lot of real engineering work is now happening.

That makes the design choices behind these tools worth thinking about more carefully than you'd think. A CLI agent gets woven through your shell config, your CI, your team's runbook scripts, your custom skills, your MCP servers, and your prompt patterns. When you evaluate an IDE plugin and decide you don't like it, you close the settings panel. When you evaluate a CLI agent that's been running in your pre-commit hooks for eight months, you migrate a workflow.

There are four layers in a modern AI CLI where openness actually matters. First, the source code of the agent harness itself. Second, the model choice. Third, the inference endpoint. Fourth, the pricing structure around that endpoint. A tool that's open at all four layers is one where the worst thing a vendor can do to you is stop maintaining it. You fork it, point it at a different model, hit a different endpoint, and keep going. A tool that's open at the harness but locked to a single proprietary endpoint looks open on the marketing page but structurally isn't.

The Gemini CLI situation that played out in May 2026 is the cleanest illustration of why this matters. Gemini CLI launched under Apache 2.0 with legitimate community traction: over 100,000 GitHub stars, 6,000 merged pull requests, hundreds of contributors. Then Google announced at I/O that Gemini CLI is being transitioned to Antigravity CLI, and free-tier users lose access on June 18, 2026. The successor hasn't been published as open source. The harness code was open the whole time, but the model and the inference endpoint were Google's, and when Google changed the terms, the open license on the harness gave the community almost nothing to work with. "Open source" at the harness layer, it turns out, is necessary but not sufficient.

Trust numbers make this more urgent. Stack Overflow's 2025 Developer Survey of over 49,000 respondents found 84% using or planning to use AI tools, up from 76% the year before. Trust in AI accuracy fell to 29%, down from 40%. Sonar's State of Code 2026 survey found 61% of developers agree AI often produces code that looks correct but isn't reliable. Adoption climbing, trust falling. That's the gap where lock-in does the most damage, because developers are increasingly dependent on tools they increasingly can't verify. When the tool is unreliable, the ability to see what it's doing and swap what's behind it isn't a nice-to-have anymore. It becomes part of the engineering process itself.

Key takeaways:

A CLI agent being open source at the harness layer is not enough if the model, endpoint, and pricing are all controlled by one vendor.
The Gemini CLI to Antigravity transition is a concrete example of how "open source" can be revoked in practice without touching the license, just by switching off the infrastructure.
Real openness in an agentic CLI means you can read the source, pick the model, pay provider-direct rates, and take your configs and skills elsewhere if the vendor changes direction.

Why do I care: I've been thinking about workflow portability for a long time, and the agentic CLI space is replaying a very familiar pattern. Cloud providers ran this playbook a decade ago: subsidize heavily until developers are fully woven in, then adjust pricing once migration cost is high enough to absorb the friction. The honest question to ask about any CLI agent you're adopting isn't "is it open source?" but "which of the four layers are actually open?" If the answer is just the harness, you're renting, not owning. For teams who are wiring agents into CI, pre-commit hooks, and shared MCP servers, that distinction is going to matter at a time that is very inconvenient to discover it.

Open Source Should be Non-Negotiable in the Agentic CLI