Meta’s Muse Code Exits Beta at $5 a Month

Meta has taken Muse Code out of beta roughly four weeks after the terminal coding agent first appeared, and paired the stable release with the two things the beta was missing: monthly pricing and a way for other software to build on top of it. Three subscription tiers now start at $5 a month, and a TypeScript SDK in developer preview lets a program drive the agent directly.

The pricing is the part that will travel. Meta is selling Muse Code as Everyday Usage at $5 a month, High Usage at $15, and Power Usage at $50 — figures that sit well under the $20 floor most rival coding assistants treat as their entry plan. Meta’s plan page puts the base tier at “10–50 requests every 5 hours”, including image and video uploads, with the middle and top tiers multiplying that allowance roughly threefold and tenfold. All three run on Muse Spark 1.2 and its million-token context window.

Three feature additions come with the stable release. Sessions can now message each other over local Unix sockets, so a change in one terminal can warn another that is building against it — Meta stresses those messages never leave the machine. A workflow engine fans a job out across parallel sub-agents, with progress visible at /workflows and successful runs saveable for reuse. And a rewind control, triggered by a double press of Escape, rolls conversation history back to what Meta calls safe rewind points derived from the session event log.

The SDK is the more consequential piece for anyone shipping a product. Rather than exposing a hosted API, it spawns a local muse host and talks to it over the Muse Session Protocol, an open protocol running over standard input and output — no server, no network hop. Developers can wire in custom tools, stream progress, embed the agent inside their own applications, and resume sessions later. Installation remains a single shell command, and macOS and Linux are the supported platforms.

What Meta still has not published is a benchmark. When Muse Code launched in beta, AI chief Alexandr Wang pitched it as “an incredibly good option, especially from a cost perspective” — a cost argument rather than a capability one, and nothing in the general-availability release changes that framing. The wager is that agentic coding tools have converged closely enough on routine work that price and extensibility, not raw model quality, decide which one an engineering organization standardizes on.

That wager is testable in a way a launch demo is not. A $5 tier with a real SDK invites exactly the long-tail integration that turns a command-line tool into a platform; an allowance measured in tens of requests per five-hour window is also the kind of ceiling a heavy user hits before lunch. Which of those two facts ends up dominating will show up in a few weeks of developer complaints or developer plugins, not in a blog post.