Leaked Screenshots Reveal Claude's New Desktop App: Sessions, Multi-Repo, and a 'Let Claude Cook' Mode

Screenshots of an unreleased Claude Code desktop experience started circulating on X today.

The images show something very different from the terminal-based Claude Code that developers use today. This is a full graphical interface with session management, pull request tracking, tabs for multiple repositories, and a custom agent builder, all under the internal codename “Epitaxy.”

The screenshots first surfaced in pieces during the Claude Code source leak on March 31, when Anthropic accidentally shipped 512,000 lines of TypeScript source in an npm package. But the images circulating now show what appears to be a more recent, more complete build.

What the screenshots show

The main dashboard greets users by name with “Ready to cook?” and presents two panels: Sessions and Pull requests.

The Sessions panel shows active coding sessions with status indicators. One session reads “Awaiting input” next to a user identifier and timestamps showing how long ago it was opened. Pull requests appear with review status badges, PR numbers, size labels (like “XL”), and a button to “Open session” directly from the PR.

At the bottom of the window, a row of tabs shows multiple repositories and branches open simultaneously. In the leaked screenshot, three separate repos are visible: tc-ghost-theme on main, tc-newsletter-agent on main, and a third project. Each tab has its own close button, just like browser tabs.

The prompt input bar sits at the bottom with placeholder text: “Find a small todo in the codebase and do it.” Below it, “Plan mode” is toggled on and the model selector shows “Sonnet 4.6.”

A small pixel-art Claude mascot sits in the bottom-right corner of the interface.

Custom agent creation

Additional screenshots reveal an Agents panel where users can define specialized sub-agents with custom prompts and tool access. The page includes a “New agent” button and states: “Specialized subagents with custom prompts and tool access. Agents defined here also appear in the CLI via /agents.”

The agent creation form has three fields: a name (restricted to lowercase letters, numbers, and hyphens, with “code-reviewer” as the example), a description (“Reviews code for best practices and bugs”), and a full system prompt field (“You are an expert code reviewer…”). This bridges the desktop app and the CLI: agents you create in the GUI show up in the terminal too.

”Let Claude Cook”

One of the most-shared screenshots from the earlier leak shows a dark-themed loading screen with the Claude mascot rendered in pixel art above the text “LET CLAUDE COOK” in retro arcade-style lettering. Below that, a stylized cityscape burns in pixel-art flames.

TestingCatalog described this as a “playful animation featuring the Claude mascot at the top of the screen launching fireballs downward, triggering a burning effect.” Whether this is a startup screen, a loading animation, or an easter egg for power users isn’t clear from the screenshots alone.

From CLI to IDE

Claude Code launched in February 2025 as a command-line tool. You type prompts into your terminal, Claude reads your code, edits files, and runs commands. It works, and it’s popular, but the interface is entirely text-based.

Epitaxy appears to be Anthropic’s answer to developers who want that same agentic capability wrapped in a visual interface. The session management alone changes the workflow significantly: instead of running Claude in a single terminal window, you’d manage multiple parallel sessions across different repositories from one app.

The Coordinator Mode described in the source leak goes further. According to analysis of the leaked code, it would let Claude “act as an orchestrator, delegating implementation work across parallel sub-agents while focusing on planning.” The dedicated Plan, Tasks, and Diffs panels in the UI appear designed for exactly this kind of multi-agent workflow.

The competitive context

Anthropic isn’t building this in a vacuum. OpenAI launched the Codex desktop app for macOS (and later Windows), with a task list view for managing parallel agents. Reports from MacRumors and others indicate OpenAI is developing a “superapp” merging ChatGPT, Codex, and the Atlas browser into one desktop application. Cursor 3.0 introduced its own multi-agent window. The race to own the developer desktop is getting crowded.

Epitaxy takes a different approach from Codex’s cloud-first architecture. The leaked code suggests it runs locally on the user’s machine, reflecting what TestingCatalog called Anthropic’s “desktop-first approach to agentic coding.”

What we don’t know

Anthropic hasn’t acknowledged Epitaxy publicly. The name, the UI, and the feature set all come from leaked source code and screenshots, not official announcements.

There’s no confirmed release date. TestingCatalog reported on April 13 that the upgrade could ship “as early as next week,” but Anthropic hasn’t confirmed any timeline. The original source leak also contained dates in code comments suggesting a May 2026 launch window for several features, though code comments and shipping dates are very different things.

The Claude Code source leak on March 31 also revealed KAIROS, an always-on background agent referenced over 150 times in the source. Whether KAIROS will ship as part of Epitaxy or as a separate feature is unknown. The proactive tick engine, the autoDream memory consolidation, and the daemon mode described in the source analysis would fit naturally inside a persistent desktop app.

We also don’t know how Epitaxy will be priced, whether it will require a Max or Team subscription, or whether it will replace or complement the existing Claude Code CLI and Claude Code Desktop experiences.

The bigger picture

Two of the features revealed in the March 31 source leak have already shipped: Ultraplan (remote cloud planning sessions) and Buddy (the terminal pet). That track record suggests Epitaxy is real engineering work, not a prototype sitting in a branch somewhere.

Anthropic told Axios after the leak: “This was a release packaging issue caused by human error, not a security breach.” They didn’t comment on the unreleased features themselves.

If these screenshots represent something close to the final product, Anthropic is about to make a significant move beyond the terminal. The question is when.