Google's Test: Claude + Android CLI Fixed Memory Leaks in 46 Minutes, Not 88

Google gave Claude one prompt: "find and fix memory leaks in this app." With Google's Android CLI, the run took 45 minutes and 57 seconds. Without it, 88 minutes and 17 seconds. Same model, same prompt, same 45 user journeys. Google published the result at Android Dev Summit on October 28, 2026, alongside a stable Android Studio that runs over 40 coding agents, Claude and Codex included.

What happened

The leak test. In the Android Dev Summit keynote, Google gave Claude Code a list of critical user journeys and that single prompt. According to Google's post, Claude ran each journey on a separate device in parallel, found leaks in many of them, followed each reference chain to fix the underlying code, and re-ran every journey to confirm the leaks were fixed.

Google's chart compares two runs of Claude Opus 5.5 on the same open-source Android app (about 40,000 lines of code), 45 journeys on 6 emulators, on an Apple M4 Pro:

• With Android CLI: 45:57, 3.5M tokens (input plus output)
• Without Android CLI: 88:17, 4.6M tokens

That's 1.9x faster and 1.3x fewer tokens, by Google's numbers. Both runs used LeakCanary 2.14; the faster one added the Android Studio LeakCanary integration through Android CLI (android profiler leakcanary), which analyzes heap dumps on your computer instead of on the device.

Any agent in Android Studio. Android Studio Rabbit 2 went stable the same day. Its Bring Your Own Agent feature lets you choose from over 40 coding agents, including Claude Agent, Codex from OpenAI, Junie and Google Antigravity, through the open Agent Client Protocol. Open Settings > Tools > AI > Agents to install them. Google first showed the feature in preview on September 24.

Android CLI as a Claude Code plugin. Android CLI now installs as an official agent plugin for Claude and Codex. From a Claude Code session:

/plugin marketplace add android/android-cli-plugin
/plugin install android-cli@android-cli-plugin

The plugin repo is Apache-2.0. It adds an android-cli skill that teaches your agent to create projects, manage emulators, inspect live UI layouts, run Journeys and search the Android docs.

The catch

• It's Google's own test. One run per setup, dated October 12, 2026, on one app. The chart's footnote adds that emulator setup time is excluded, and that counting cache reads the token gap shrinks to 131M vs 157M (1.2x). Treat 46 vs 88 minutes as a promising data point, not a benchmark.
• The wording on availability is mixed. The same post says LeakCanary "is now integrated directly into Android CLI" and later that "the memory leak subcommand is coming soon to Android CLI," and it calls Fix with AI for memory leaks a canary feature. Google's Rabbit 2 post, meanwhile, lists agentic LeakCanary fixes with the built-in agent among its stable features. Check what your installed version actually has.
• Claude Code isn't free. It isn't on Claude's free plan. Pro includes it at $20 a month, or $17 a month billed yearly.

Try it yourself

1. Update to the latest stable Android Studio and open Settings > Tools > AI > Agents. Install Claude Agent or Codex.
2. Install Android CLI and run android init to set up your agent.
3. In Claude Code, add the plugin with the two commands above.
4. Add LeakCanary to a debug build of an app you own.
5. Ask Claude to run one key user journey, list any leaks LeakCanary reports, and propose a fix for each one.
6. Read every proposed change yourself, or start Android Studio's Change Review (/review in the agent window) on the pending changes.

Here's how you get paid

My opinion (no income guaranteed): plenty of Android apps leak memory, and the owners never find out until users complain about slowdowns or out-of-memory crashes. Very few small teams have time to hunt leaks journey by journey.

Sell a fixed-scope Android leak audit:

• Journeys. Agree on the 10 to 20 user journeys that matter most to the business: sign-in, checkout, the main feed.
• Find and fix. Claude runs each journey with LeakCanary, follows the reference chains, and proposes fixes. Every journey is re-run to confirm the leak is gone.
• Human sign-off. You review every change before it ships and hand over a short report: what leaked, why, what changed.

Then offer a monthly re-check after each release. Work only on apps the client owns, in a debug build, and never ship a change you haven't read.