Gemini CLI review, pricing and verdict

If you're evaluating a terminal coding agent today, Gemini CLI (107k+ GitHub stars) is not an option unless your organization already holds a Gemini Code Assist Enterprise license. Individual access ended, and Google's own successor, Antigravity CLI, is the practical next step for everyone else.

Gemini CLI was Google's open-source terminal AI agent, launched June 25, 2025 with a 1 million token context window. It ran a reason-and-act loop to fix bugs and ship features from the command line until Google ended free and Pro/Ultra consumer access on June 18, 2026 in favor of Antigravity CLI.

Maker: Google · Autonomy: semi autonomous · Maturity: ENTERPRISE

Underlying models: Gemini 2.5 Pro, Gemini 2.5 Flash, Gemini 3 Pro, Gemini 3 Flash

About Gemini CLI

Gemini CLI is Google's open-source AI agent for the terminal, launched June 25, 2025 under an Apache 2.0 license. It runs a reason-and-act loop against your local files and shell, using Gemini's 1 million token context window to read and edit whole codebases, fix bugs, and ship features while a human approves each change by default. On June 18, 2026, Google stopped serving Gemini CLI requests from free accounts and Google AI Pro/Ultra subscribers, redirecting individual users to its successor, Antigravity CLI; only organizations already holding Google's enterprise Code Assist contract, or developers paying per token through the Gemini API, still have working access.

At launch it defaulted to Gemini 2.5 Pro, and Google later added Gemini 3 Pro and, in December 2025, Gemini 3 Flash, which the company's own developer blog reported scoring 78% on SWE-bench Verified for agentic coding, ahead of both the 2.5 series and Gemini 3 Pro at the time. Three approval modes controlled how much it could do without asking: the default mode confirmed every file edit and shell command, auto_edit approved edits automatically while still asking about shell commands, and yolo mode (--yolo) approved everything with no prompts. It also supported Model Context Protocol servers, Google Search grounding for up-to-date answers, session checkpointing, and a headless scripting mode for CI pipelines.

Before the shutdown, individuals used it for the same reasons any terminal coding agent gets adopted: fixing bugs across a large repository, generating apps from a sketch or PDF using Gemini's multimodal input, and automating routine pull-request and rebase work. A 2026-04-24 security advisory (GHSA-wpqr-6v78-jr5g) rated CVSS 10.0 after researchers showed a GitHub issue could trigger indirect prompt injection in headless CI mode, exfiltrating a workflow's GITHUB_TOKEN; Google patched it in @google/gemini-cli 0.39.1 and the companion google-github-actions/run-gemini-cli GitHub Action.

The project remains open source and still actively maintained on GitHub as of today. Development continued past the consumer cutoff: Google shipped a preview build on September 1, 2026 hardening MCP OAuth discovery against SSRF, and the latest stable release is v0.61.0. For anyone without an existing enterprise contract, Google's own guidance is to move to Antigravity CLI; reviewers otherwise point migrating users toward Claude Code, OpenAI Codex, or Google's own async agent Jules.

Pricing

No consumer free tier remains. Continued access requires either a Gemini Code Assist Standard or Enterprise license (custom pricing) or a pay-per-token Gemini API key; individuals had a generous daily free allowance before the 2026 shutdown.

Key Features

• 1M token context window: Reads and edits an entire codebase in one session instead of working file-by-file, the feature most reviewers cited as its standout trait.
• ReAct loop over local tools and MCP: Reasons about a task, calls a local tool or an MCP server, observes the result, and repeats until the task is done or it needs your approval.
• Google Search grounding: Pulls current web results into its reasoning instead of relying only on training data, useful for library APIs that changed after the model's cutoff.
• Session checkpointing: Automatically snapshots a session's state so a long-running task can be resumed instead of restarted from scratch after an interruption.
• Headless scripting mode: Runs non-interactively from a script or CI pipeline, the same mode a critical security advisory found unsafe until Google patched it in v0.39.1.
• Three-tier approval control: default confirms every edit and command, auto_edit approves file edits only, and yolo mode approves everything with --yolo and no prompts.

Strengths

• Its GitHub repository carries 107.2k stars and 14.6k forks under Apache 2.0, making it one of the most-starred open-source coding agents anyone can still inspect or fork.
• A 1 million token context window let it hold entire small-to-mid-size codebases in a single session, well beyond what most rival agents offered at its 2025 launch.
• Gemini 3 Flash scored 78% on SWE-bench Verified for agentic coding per Google's own December 2025 announcement, ahead of Gemini 3 Pro at the time.

Weaknesses

• Individual, Google AI Pro and Google AI Ultra access ended completely on June 18, 2026; anyone without an enterprise Code Assist license or a paid API key cannot use it at all.
• A CVSS 10.0 remote-code-execution flaw in its headless CI mode (GHSA-wpqr-6v78-jr5g, disclosed 2026-04-24) let a GitHub issue exfiltrate a workflow's secrets before Google patched it.
• It required constant internet connectivity, since all reasoning ran on Google's servers rather than locally, which also raised data-handling concerns for some organizations.
• Free-tier users before the shutdown reported the agent silently downgrading from Pro to Flash under quota pressure, changing response quality mid-session.

Visit Gemini CLI Official Site