What is Gemini CLI now?
Gemini CLI is Google's open-source AI coding agent for a terminal. It can inspect and edit repositories, execute shell commands, search and fetch web content, maintain project instructions, plan work, resume sessions, and connect to MCP servers, extensions, hooks, skills, and subagents. Headless mode supports scripts and structured output for automation.
Its 2026 access story is more important than the feature list. Google announced that, from June 18, Gemini CLI would stop serving requests for free-tier, Google AI Pro, and Google AI Ultra individual accounts as the individual terminal product moved to Antigravity CLI. Enterprise users with Gemini Code Assist licenses and API-key authentication were said to remain unaffected. The current quota page still presents the earlier individual tiers, creating an official documentation mismatch. Do not buy a personal plan or design onboarding around the old table without testing the exact account.
Coding, planning, and extensibility
The agent can read many files, search a codebase, propose patches, write files, run tests and shell commands, access the web, and track tasks. Plan Mode restricts the normal tool set to research and read-oriented actions, produces an implementation plan, and requests approval before switching to execution. Session resume, rewind, and optional checkpointing help users recover context or restore state after AI edits.
Gemini CLI is also a platform. MCP servers expose external tools and data; extensions can package MCP, commands, skills, hooks, themes, and configuration; hooks can inspect or change behavior; subagents can isolate specialized work. Headless mode accepts prompts in scripts and can return text, JSON, or streaming JSON.
Every extension point expands the trust boundary. A project can contain settings, MCP definitions, hooks, skills, commands, and environment files. An MCP tool can read data or take actions outside the repository. A hook inherits the CLI process environment and may see secrets. Review the code and publisher, pin versions, minimize environment variables, and use separate credentials for development and automation.
Pricing, quotas, and total cost
The CLI client itself is licensed under Apache 2.0. Model access is a separate service. Official quota documentation lists an unpaid Gemini API key allowance of up to 250 requests per user per day, restricted to Flash, while paid Gemini API and regular Vertex AI access vary by model, tokens, pricing tier, and dynamic or provisioned quota.
For organizations, the page lists 1,500 requests per user per day for Gemini Code Assist Standard and 2,000 for Enterprise. It also still lists old personal quotas of 1,000 for Code Assist Individual, 1,500 for AI Pro, and 2,000 for AI Ultra, despite the later transition announcement. Request limits are not guaranteed throughput and can be subject to per-minute limits and service availability.
Use /stats model to inspect current-session usage and quota. For paid API use, set budgets and alerts and test long-context costs, cache behavior, tool-call loops, retries, subagents, and CI concurrency. Count developer review, failed changes, security scanning, test infrastructure, and incident handling—not only tokens.
Tool approvals, sandboxing, and repository trust
Default policies generally allow read-only tools and ask before file writes or shell commands. The CLI shows a diff or exact command for confirmation. Auto Edit can approve edit tools automatically, while YOLO approves all tools. Plan Mode is intended to remain read-only. Broad persistent rules or allowed tools can silently weaken those defaults.
Sandboxing can isolate commands with macOS Seatbelt, containers, Windows controls, gVisor, or experimental LXC depending on the platform. A container normally mounts the active workspace, so the agent can still modify every writable file inside that mount. Sandbox expansion can request more filesystem or network access. Isolation reduces blast radius; it does not prove a command, dependency, download, or generated patch is safe.
Folder Trust prevents project-specific settings from loading before a repository is approved. In an untrusted workspace it ignores local settings and environment files, blocks MCP connections and custom commands, disables auto-acceptance, and restricts extensions. However, the feature is disabled by default. Enable it before opening unfamiliar repositories, inspect everything discovered, and avoid trusting a broad parent directory.
Start in Plan or default approval mode, enable sandboxing and Folder Trust, keep secrets outside the workspace, run with a low-privilege account, and inspect git diffs. Never approve a command merely because its description looks safe; check the exact paths, flags, redirects, package scripts, and network targets.
Data, telemetry, and governing terms
Open-source licensing applies to the client, not to model processing. Depending on authentication, requests are governed by Gemini Code Assist, Gemini API, or Google Cloud terms and privacy notices. Prompts, selected source files, retrieved content, tool results, and conversation context can be sent to the chosen service. Unpaid and paid API services may have different terms, so sensitive teams must choose the correct route and document it.
Gemini CLI usage statistics are enabled by default but can be disabled with privacy.usageStatisticsEnabled. Google's configuration documentation says these statistics record model, duration, success, configuration, tools, and approval mode, not prompt, response, file content, PII, or API keys. This client statistic stream is distinct from model-service processing.
Optional OpenTelemetry is disabled by default, but if enabled its logPrompts default is true and detailed traces can contain inputs, outputs, system instructions, tool definitions, and other sensitive context. Send it only to an approved destination, disable prompt logging where needed, set retention and access controls, and never assume “telemetry” is one uniform channel.
Verdict
Gemini CLI remains a sophisticated, extensible terminal agent and a credible option for organizations already using Google Cloud or Gemini Code Assist. Its open client, planning flow, automation formats, MCP ecosystem, and layered policy controls are useful for serious development work.
For individuals, product continuity changed materially in June 2026. Confirm authentication before evaluating anything else. For every team, safe adoption requires more than accepting default prompts: enable Folder Trust and an appropriate sandbox, restrict integrations and secrets, keep human review over commands and diffs, establish API budgets, and align the selected Google service with the sensitivity and contractual status of the code.
