Conductor is an extension for the Gemini CLI, introduced by Google in the December preceding its February 2026 update, whose stated aim is to bring context-driven development to the terminal. Its central design choice is where project awareness lives: rather than accumulating in ephemeral chat logs, it is shifted into persistent, version-controlled markdown files — plan.md and spec.md are the two the source names. Google credits this with helping developers plan before they build.
With the Automated Review feature announced in February 2026, the extension covers validation as well as planning and execution. Once a coding agent completes its tasks, Conductor can generate a post-implementation report on code quality and on compliance with the guidelines the developer has defined.
In July 2026 Google announced that Conductor was evolving from a Gemini CLI extension into the Conductor Plugin, a package that can bundle skills, rules, MCP servers and hooks (Evolving Spec-Driven Development: Conductor Now Supports Antigravity). The stated consequences are that it can be driven conversationally rather than through strict command sequences, and that it is no longer tied to Gemini CLI.
The extension is installed from its GitHub repository, either directly or with gemini extensions install https://github.com/gemini-cli-extensions/conductor; the plugin is installed into Antigravity CLI with agy plugins install https://github.com/gemini-cli-extensions/conductor.
Capabilities
On the planning side, project context is kept in version-controlled markdown files that the workflow then works from. As a plugin, Conductor is described as engaging with the developer conversationally and generating context, specs and plans as a feature is discussed, while still producing spec.md and plan.md as persistent artifacts; Google says it decides for itself when to update the project context or check off a completed task in the plan. The plugin is stated to be backwards compatible with the older Conductor commands and with existing plans and specs.
The Automated Review side is described in more detail, as five checks that make up a single post-implementation report:
- Code review — deep static and logic analysis of newly generated files, going beyond syntax to flag complex issues such as race conditions in asynchronous blocks, potential null pointer risks, and logic errors that could lead to runtime exceptions.
- Plan compliance — checking the new code against
plan.mdandspec.mdto confirm every phase of the roadmap was addressed and that no core requirements were omitted during coding. - Guideline enforcement — verifying that new contributions adhere to the project’s style guides and to any custom guideline files generated during the planning phase.
- Test-suite validation — running the relevant unit and integration tests as part of the review workflow and incorporating the results and coverage data into the report, instead of relying on manual execution.
- Basic security review — scanning for critical vulnerabilities before code is merged, flagging high-risk issues such as hardcoded API keys, potential personally identifiable information leaks, or unsafe input handling that could expose the application to injection attacks.
Findings are graded High, Medium or Low and carry the exact file path, and a track can be started within Conductor to work through them. No version number is stated for any of this behaviour, and the capabilities are Google’s own description rather than measured results.
Adoption & Ecosystem
Conductor began as an extension of the Gemini CLI rather than as a standalone tool, and so inherited that assistant’s terminal-based way of working. As a plugin it is presented as usable from whichever tool fits the task — Antigravity CLI and Claude are the examples Google names — with the project’s foundational documents, shared configuration and ongoing development tracks persisting so that work started in one tool can continue in another. Its plan.md / spec.md convention places it alongside the broader practice of Spec-driven development, and the review step it adds is an instance of Agentic Code Review run against those same specification files rather than against a reviewer’s own judgment alone.
Google frames the arrangement as keeping agentic development supervised: the AI supplies the labour while the developer supplies high-level architectural oversight, backed by automated verification. The announcement is reported in Conductor Update: Introducing Automated Reviews.