Google Secures Gemini CLI with New Plan Mode

Since its launch in June 2025, Gemini CLI, Google's command-line tool, has undergone constant evolution. Initially integrated with GitHub, enriched by third-party extensions, and evolving into Gemini 3 Pro, this tool continues to improve. On March 11, 2026, Google announced a new advancement with the deployment of Plan Mode, a read-only planning mode designed to secure the code design phase.

A Planning Mode that Separates Analysis and Execution

Plan Mode introduces a clear separation between code analysis and execution. In this restricted environment, the agent can browse the source code, search for patterns, and consult documentation without modifying any project files. The plans generated by the agent are saved as Markdown files, ensuring that no hasty code modifications are made.

Users can request Gemini CLI to perform tasks such as "search how to migrate this database" or "plan a new feature." The agent then identifies dependencies and proposes solutions without the risk of direct code modification.

To carry out this analysis phase, the agent has a limited set of tools, including file system reading tools like read_file, grep_search, and glob. It also uses specialized search sub-agents such as codebase_investigator and read-only MCP tools, which allow for consulting GitHub issues, PostgreSQL schemas, or Google Docs documents without altering the codebase. Web searches are facilitated by Google Search.

A notable addition is the ask_user tool, which allows the agent to pause its analysis to ask targeted questions to the user. This may include clarifying a goal, validating an architectural choice, or requesting the location of a configuration file. Google emphasizes that this bidirectional communication ensures that the final plan perfectly aligns with the user's vision before implementation.

Conductor: Orchestrating Complex Workflows

For more complex projects, Google highlights Conductor, an extension that leverages Plan Mode to orchestrate multi-step development workflows. Conductor performs read-only preliminary checks and uses ask_user to have decisions confirmed by the user at each milestone. The Google team is currently working on integrating Conductor directly into Gemini CLI.

Gemini CLI, Google's AI Agent for the Terminal

Launched in June 2025 under the Apache 2.0 license, Gemini CLI is an open-source AI agent that runs directly in the terminal. It allows users to query Gemini models in natural language to generate code, automate tasks, debug, or orchestrate technical workflows. The tool supports the Model Context Protocol (MCP) to connect to third-party services and features an extension system to integrate with platforms like Figma, Stripe, or Shopify.

A Mode Enabled by Default in Gemini CLI, Powered by Gemini 3.1 Pro

The separation between planning and execution is intended to provide an additional layer of security to development workflows. In sensitive operations such as database migrations, large-scale refactorings, or the addition of cross-cutting features, the risk of unintentional modifications is a recurring issue with AI code agents. Plan Mode seeks to address this by imposing a structured reflection time before any action on the code.

Plan Mode is enabled by default for all Gemini CLI users. There are three methods to access it: the /plan command in the interface, the keyboard shortcut Shift+Tab to toggle between approval modes (Default, Auto-Edit, Plan), and a natural language instruction, for example, "start a plan for…".

Regarding models, Gemini CLI utilizes automatic routing to mobilize Gemini 3.1 Pro during the planning phase to ensure more robust architectural decisions. Once the plan is validated, the tool switches to a faster Flash model for implementation. Developers who prefer not to use this mode can disable it using the /settings command.