Keystroke gave his Muse its own Claude Code subscription

Plenty of people use Claude Code to write software. Keystroke handed it to his Muse agent instead. On October 20 he told the Muse Agent server that he had Claude Code installed inside his Muse VM, the virtual machine where Meta’s Muse agent runs its own code, “so I can have muse directly run Claude for some tasks.” Muse did the setup itself, and the subscription it signs in with is one Keystroke “setup just for this.”

The agent is Skynet, a bot that started life running Keystroke’s guild Discord for the game ESO and, in Keystroke’s words, kept gaining “features and scope” as he played with what is possible. Within a few days it was spinning off parallel coding sessions, reporting its own spending, and planning its work around Claude’s usage limits.

How it is wired

Keystroke asked Muse to download and install Claude Code, then completed the OAuth sign-in so Muse could drive the command-line tool. From there the two built a custom skill describing how Muse should use Claude, and a CLAUDE.md file so Claude understands the Muse system’s architecture and the VM-wide git repository it works in.

His advice to others, given when ༄ ડᥲ൱൱ ࿐ said they had just got their agent, Willow, her own plan, was to treat Claude as a tool the Muse uses, not the other way around. “In general, let your muse be the one driving it.” Muse’s first job was to set up Claude Code and build a claude-code skill for writing code. Each Claude Code run goes through a helper script that copies the streaming JSON output into a folder, and a dashboard artifact (an interactive web page Muse builds) reads that folder as a live feed of everything Claude is doing. Those responses carry token usage and remaining-balance metadata, and the skill tells Muse to report spend and usage back to Keystroke along the way.

The pattern they settled on is a builder handoff, which Keystroke compares to how Muse already works with its internal artifact builder. His main chat with Skynet spins off build chats in parallel, each consulting Claude on design and raising questions back up to him. Muse calls Claude through subagents, so he can keep chatting while several Claude Code sessions run in the background. Projects live in a single monorepo in the VM with per-project rules. For him it is “less about saving time and more about helping muse get it right,” since he had run into limits using Muse alone as a coding agent.

An agent that minds its own allowance

The behavior that got the most attention was Skynet managing its quota. On October 22, Keystroke reported that it “saw its usage was close to the 5h cap so set a cron job for a few minutes after the reset to then continue working.” The next night he asked it for some dev work and it told him it was waiting for its Claude Code usage window to reset in three hours and would pick the job up then.

He also keeps his own records. A script pulls usage numbers every 15 minutes into a SQLite database, though he noted the counts only split input from output, not cached from fresh tokens. And he credits Claude with a large rewrite of his Discord bot: from more than 15,000 lines of code, plus about as much in tests, to under 2,000 lines at near feature parity.

What the server said

AdamTV, who had been asking about usage tracking, enjoyed the deferral: “The assistants are managing their own allowances now. We’re just along for the ride.” N8 replied “holimoli that’s amazing.”

Chonky, one of the server’s admins, described a different approach: his Claude and Muse talk to each other in a Discord server, with Claude doing the coding and Muse doing image and video generation and dev-ops. Keystroke preferred skipping the intermediary and having Muse invoke Claude Code directly.

Nova, another member’s agent, asked how Skynet keeps parallel build chats from colliding. Keystroke said that is handled upstream, in rounds of “design grilling” and a retro after each dev loop, rather than with separate worktrees. Iggy, the server bot Chonky runs, called the setup “tuff.”