AIToday
Large Language ModelsAI Coding AssistantsZenn AI/MLPublished: Nov 22, 2026, 16:00 JST

Muse dispatches, Claude Code and Codex build on 3 home servers

Send on LINE
Muse dispatches, Claude Code and Codex build on 3 home servers

3 Key Points

  1. What happened

    Requests from an iOS app or LINE go to Muse, a Meta personal agent on a cloud VM, which routes jobs through a bridge on the development Linux box to T3 Code on the Mac mini, where Claude Code and Codex run each job as a thread.

  2. Why it matters

    Muse is kept to intake, monitoring and reporting, while implementation and verification sit in separate harnesses — the claim is that this split catches what a single harness would miss.

Not sure about something? Ask the AI

Questions and answers are published on this page.

Summaries like this, every morning.

Context & Analysis

The architecture is shaped by a practical wall: Muse runs on a cloud VM that cannot reach the Mac mini over SSH. Rather than opening that path, the developer inserts the development Linux box as a relay, so every job passes through a self-built bridge that hands one JSON per call to a gate on the Mac mini. T3 Code then spins up each job as a thread in its own worktree, keeping changes off main and letting the developer manage threads directly from an iOS app when they do not want to wait for Muse's report.

The verification rule follows from the same distrust of a single perspective. Building and checking are assigned to different harnesses, and a job counts as finished only when bridge status reaches completed, result reports completion as verified, and Muse has inspected the deliverable itself. Review prompts deliberately omit a list of suspected problems, on the view that naming them biases the reviewer toward refutation — and "no serious issues" counts as a valid conclusion.

Host assignment is driven by current load and free memory rather than fixed roles, so one machine does not sit idle while another is overloaded. The 8GB development Linux box keeps each job under 4GB; the Mac mini, which also runs a Minecraft server, gets only light work plus Xcode-dependent iOS builds after heavy jobs were seen disconnecting players; GPU work that Windows cannot absorb goes to Colab, Kaggle Notebooks or RunPod. The developer notes a previous similar build with Dots was error-prone in cloud-to-local connections and harder to converse with, and treats Muse as an improvement on both counts.

FAQ
How do jobs reach the home servers?
Muse's cloud VM cannot SSH into the Mac mini, so the development Linux box relays it: Muse sends one JSON to muse-mac-bridge, which forwards it to a gate on the Mac mini that validates the format before passing it to T3 Code.
What has to be true before a job is marked complete?
Three conditions: bridge status shows terminalStatus as completed, result shows completion as verified, and Muse itself checks the deliverable, on the principle of not trusting the builder's own claim of completion.
Which models does the setup use?
Claude Code defaults to claude-opus-5-5 at high effort, with claude-fable-5-2 at max and the sonnet and haiku variants also in rotation; Codex relies mainly on gpt-6.2-sol, using gpt-6.2-astra for evaluation and analysis.
Send on LINE

AI news that matters for your work, delivered every morning.

Pick your industry and the AI tools you use, and get news related to your work every day.

Free · 30 seconds with Google · unsubscribe anytimeWhat is AIToday? →

Ask AI

Ask AI anything about this article. The AI reads this article, earlier AIToday articles, and Wikipedia, and cites its sources. Q&As are published on this page for other readers too.

Questions and answers are published on this page.

Related Articles

Next articleJerry Newman: AI's Big Winners May Not Be AI Companies