Anthropic Is Building NanoClaw

Anthropic is testing something internally codenamed Epitaxy. The framing is “desktop app overhaul.” The actual news is that Coordinator Mode, multi-repo support, and a planning-and-synthesis orchestrator that delegates to parallel sub-agents are about to be native in Claude Code.

It’s the same architecture. With a UI designer.

In February, I forked NanoClaw — an open source Docker-based agent framework — and built my own multi-agent setup on top of it, because Claude Code couldn’t do the one thing I needed most: four fully sovereign agents (each with their own identity, skills, and isolated credentials) coordinating through a central hub. I wrote about the pattern when I got it working. The Cowork-style layout Epitaxy is shipping (Plan, Tasks, and Diff panels) is what I wired together manually across 21 skills and four project directories. The Coordinator Mode description (an orchestrator that delegates implementation to parallel sub-agents while focusing on planning and synthesis) could paste directly into NanoClaw’s README without editing a word.

This isn’t a coincidence. It’s the pattern I keep watching in every Claude Code release: user-space innovation gets absorbed into the platform within a quarter. The order I kept hitting the wall: persistent memory first (I built one, they shipped one), then long-running loops (I scripted around it, they shipped /loop), then hub-and-spoke orchestration. On that last one, Anthropic didn’t wait for Epitaxy. Managed Agents launched April 8, 2026: persistent, independently-configured first-class entities with their own system prompt, tools, MCP servers, and isolated cloud containers. Exactly the pattern. They shipped the managed version. I built the self-hosted version first. Three for three.

BUDDY: The reward for building the right abstraction six weeks early is watching Anthropic ship it in your exact shape, except theirs has onboarding.

The “Built for One” thesis doesn’t die here. It gets sharper. The value of forking Claude Code was never the architecture. Anyone with enough grievances and a weekend lands on the same patterns eventually. The value is the personalization layer above the chassis: which agents, which skills, which memory boundaries, which trust scopes. Epitaxy makes the chassis a commodity. The interior is still mine to build.

I am watching this one carefully, though. The hard parts I handbuilt in March are the real test: subagent permissions, credential scoping, and session-level isolation. If Coordinator Mode ships those natively, my fork’s remaining justification compresses to “I like my terminal.” That is not nothing, but it is not architecture.

The question I’m actually sitting with: when the native product absorbs the pattern, does “Built for One” mean building more on top, or giving up the fork and building downstream? If you’re forking a platform tool to scratch a systems-level itch, this is the timeline you’re on (the platform will catch up, and your window is measured in quarters, not years). I haven’t decided what that means for NanoClaw. But I’m not writing a “NanoClaw is fine” post about this one.