Epitaxy Unifies Claude Code: Local + Web in One Interface

Video description

In this video, I'll be talking about the latest Anthropic leaks, including the rumored Epitaxy interface, the Mythos or Capybara model story, and why this may be a much bigger Claude product shift than most people realize. I also break down what is actually visible in the leaked clips and why a unified Claude Code, Claude Desktop, and web workflow could matter more than the model leak itself. -- Key Takeaways: 🔥 Epitaxy appears to be a rumored new Claude interface that could unify local and web-based task execution in one place. 🖥️ The leaked UI shows Claude Code-style controls like Local, Select folder, worktree, auto-accept edits, and a model picker. 🌐 One leaked clip reportedly shows a claude dot ai slash epitaxy route, making the feature look more like a real internal build than a mockup. 🧠 Mythos and Capybara may be two names for the same unreleased Anthropic frontier model rather than separate launches. ⚙️ The leaked model details suggest stronger coding, reasoning, and cybersecurity performance, while still having believable flaws like over-commenting and false claims. 🎨 The LET CLAUDE COOK design gives the leaked interface a more distinct long-running agent feel than a normal chatbot UI. 🚀 If Anthropic ships this, the bigger story may be product integration across Claude Code, Claude Desktop, and the web, not just another model update.

Epitaxy Interface Solves Claude's Workflow Fragmentation

Anthropic's leaked Epitaxy unifies Claude Code, Desktop, and web into one surface for task execution. Controls include 'Local', 'Select folder', 'Worktree', 'Auto-accept edits', and a model picker (showing Opus 4.6, Sonnet 4.6, Haiku 4.5, Sonnet 4.5 in footers). Browser dev tools reveal claude.ai/epitaxy route, confirming it's a real internal build, not mockup. Reverse-engineering of Claude Desktop points to Epitaxy as a mode alongside chat/code/task sections.

Switch execution targets without product-switching: run repo tasks locally for file/terminal context, or remotely for background persistence. 'Let Claude Cook' design—with dark CRT aesthetic, pixel claw icon, flame animation—signals long-running agent mode, distinct from chat UIs. This reduces mess from CLI/Desktop/web silos, making Claude feel like one agent system. Trade-offs: UI details may evolve; web execution could be sandbox, hosted env, or remote agents.

Mythos/Capybara: Frontier Model Powers the Shift

Mythos and Capybara are likely the same unreleased frontier model (Anthropic confirmed testing one with reasoning/coding/cybersecurity gains; early access for select customers). Internally on Capybara v8 for Claude Code dev, with 1M token context and 'FAST mode' for speed/depth trade-offs.

Strengths: Step beyond Opus in coding, reasoning, cybersecurity. Flaws preserved for realism—over-commenting, false claims—matching real dev notes over hype. Other code names: Numbat (placeholder for launch removal), Fenex (Opus 4.6 mapping). Model pairs perfectly with Epitaxy: massive context/fast modes enhance unified local/web tasks.

Product Integration > Model Benchmarks Alone

Leaks prioritize product evolution over raw model specs. Frontier models excel on benchmarks but falter in fragmented workflows; Epitaxy delivers cohesive experience, turning capability into daily utility. Anthropic iterates fast internally (v8 dogfooding), signaling imminent rollout. Direction: Intentional agent identity via distinct modes, not bolted-on chat features. If shipped, this reshapes Claude as task launcher across envs—huge for AI coding where integration trumps isolated upgrades.

Video description

Epitaxy Interface Solves Claude's Workflow Fragmentation

Mythos/Capybara: Frontier Model Powers the Shift

Product Integration > Model Benchmarks Alone

More on Edge

Use Claude Code + Codex Together for Best AI Coding

DeepSeek V4 + Claude Code Proxy for 76% Cheaper Coding

Free NVIDIA NIM API Unlocks Kimi K2.6 for Agentic Coding

Codex CLI Beats Claude Code on Cost and Autonomy