Google's Conductor Now Plugs Into Claude Code

Google's Conductor started last December as a Gemini CLI extension with one strict rule: the agent writes a spec and a plan before it writes code, and both live as markdown in your repo. Google has now rebuilt it as a portable plugin. It installs into the Antigravity CLI, and it installs into Claude Code.

A Google team now ships and maintains first-party tooling for Anthropic's coding agent. That's a first for Google's agent tooling, and it tells you where this layer of the stack is going: the process artifacts move into your repo, and the agent underneath becomes swappable.

What Conductor puts in your repo

Conductor's premise is that chat history is a terrible system of record. Instead of letting project context evaporate at the end of a session, it maintains a conductor/ directory in the repository: product context, tech stack notes, workflow preferences, and a "track" per feature or bugfix containing a spec.md and plan.md. The agent reads these files at the start of every session and updates them as work progresses. Commands cover the lifecycle you'd expect (setup, new-track, implement, status, review) plus a revert that undoes a logical unit of work, a track or a phase, rather than making you hunt for commit hashes.

Because the artifacts are plain markdown under version control, they show up in pull requests. Your reviewer can push back on the spec before the agent generates a thousand lines against it. That's the strongest practical argument for spec-driven development in team settings, and it has nothing to do with which model does the generating.

The migration behind the rebrand

The plugin architecture isn't purely a philosophical evolution. Google announced at I/O in May that Gemini CLI sunsets for individual users on June 18, 2026, with Antigravity CLI (agy) as its replacement. Every Gemini CLI extension needed a new home, and Conductor's team used the forced move to build something portable instead of porting one runtime dependency to another. Smart, but the "evolving" framing in Google's announcement papers over the churn: if you had Conductor wired into Gemini CLI scripts, you're migrating whether you wanted to or not.

The Claude Code support is the more deliberate choice, and the community forced Google's hand a little here too. Within days of Conductor's original release, a developer had already ported the whole workflow to Claude Code and posted it on Hacker News. When your process layer is markdown files plus prompting conventions, anyone can reimplement it in a weekend. Google shipping official cross-agent support is an acknowledgment that the moat was never there.

Specs are becoming the commodity layer

Conductor joins a crowded field. AWS built Kiro as an entire IDE around spec-driven development. GitHub's Spec Kit took the opposite approach: a thin CLI and slash commands designed from day one to work across Copilot, Claude Code, Gemini CLI, and others. Conductor has now converged on Spec Kit's position. The workflow lives in the repo; the agent is an implementation detail.

This is good for developers and clarifying for vendors. If your team's specs, plans, and conventions are agent-agnostic files, switching from Claude Code to Antigravity (or back) costs you nothing at the process level. Lock-in moves up the stack to the models themselves, which is where the actual differentiation is anyway. Google isn't being altruistic: once conductor/ conventions are in your repo, trying Antigravity is frictionless. But the same door swings both ways, and Anthropic users get a well-maintained SDD plugin for free.

The conversational bet is the risky part

The other change in this release is tonal. The original Conductor, like Spec Kit, marched you through a command sequence. The plugin drops the rigidity: you talk to the agent naturally, and it decides when to create or update the spec and plan in the background. The old commands still work. Google says internal Terminal-Bench testing showed better task completion with the new approach, though those are their numbers and I haven't seen an independent run.

I'm skeptical of this direction, and not because conversation is a worse interface. The ceremony was doing work. A mandatory spec step functions like a PR template: annoying, and effective precisely because you can't skip it. When artifact maintenance becomes something the agent handles "organically," you invite drift, a spec.md that no longer matches the code, plans backfilled to match whatever got built. The failure mode of conversational SDD is vibe coding with paperwork.

Trying it

Installation is one command per platform:

# Antigravity CLI
agy plugins install https://github.com/gemini-cli-extensions/conductor

# Claude Code
/plugin marketplace add gemini-cli-extensions/conductor
/plugin install conductor

It's Apache 2.0, so forking or auditing the prompts is fair game. If you trial it, watch three things. Token overhead, because keeping artifacts current means the agent re-reads and rewrites context on every session. Artifact drift, by diffing spec.md against what actually shipped after a couple of tracks. And cross-agent consistency, because a plugin that behaves differently on Claude Code than on Antigravity gives you portability in name only. The prompt-level protocol is identical, but different models follow protocols with different fidelity, and nobody has published a comparison yet.

Adopt the portability even if you throw the conversational mode away and keep typing the commands. Process tooling that survives an agent switch is the only kind worth building on, because if the last year proved anything, it's that you will be switching.