Meta has taken its terminal coding agent Muse Code out of beta, roughly four weeks after the August 5 preview launch. The general-availability release adds inter-session messaging, a workflows engine for multi-agent task decomposition, a TypeScript SDK in developer preview, and three monthly subscription tiers priced at $5, $15, and $50 per month.
The underlying model remains Muse Spark 1.2, unchanged since the beta. Meta is positioning the release as a maturity upgrade to the harness rather than a model capability leap, putting it in direct competition with Anthropic's Claude Code, OpenAI's Codex, and Cursor as a subscription-based coding product.
What's new
- Inter-session messaging: Sessions running locally on the same machine can now exchange context over a Unix socket. Two built-in tools handle peer discovery and message transmission, letting developers hand off review requests, status updates, and code changes between related sessions without manual copy-paste.
- Workflows engine: A task-decomposition feature that breaks a large engineering task across multiple specialized agents working in parallel and staged sequences. Workflows can be triggered explicitly or by setting effort to "ultra," and are visible through a /workflows command that shows progress, allows stop/restart/cancel, and returns a consolidated result. Meta's documentation notes the workflow engine is gated behind the installed build and rollout enablement, so availability may vary by platform at launch.
- TypeScript SDK (developer preview): A library that spawns a local Muse host process and communicates over stdio using the Muse Session Protocol (MSP). It covers session management, tool registration, and permission handling with types auto-generated from the schema. The schema is machine-readable for potential clients in other languages, though only TypeScript ships in this preview.
- Subscription plans: Three monthly tiers sit alongside the existing per-token API pricing:
- Everyday Usage — $5/month: Access to Muse Spark 1.2, 10–50 requests per 5-hour window including image/video uploads, voice mode, and web search.
- High Usage — $15/month: Everything in Everyday plus 3× usage, more multimodal inputs, and access to the latest models.
- Power Usage — $50/month: Everything in High Usage plus 10× usage versus Everyday, expanded file uploads, and early access to new features.
Context: from beta to general availability
Muse Code entered beta on August 5, 2026, as Meta's first dedicated terminal coding agent under AI chief Alexandr Wang. The beta already featured multi-agent coordination, async background agents, a local event log for replay-exact restarts, and bundled skills like /plan, /grill, and /goal. For benchmarks, contributor-tier API pricing, and harness architecture from that launch, see our August 5 coverage.
Today's release keeps the same model and core runtime. Muse Code remains a macOS and Linux terminal agent installed via
curl -fsSL https://dev.meta.ai/install.sh | bash
The new subscription tiers are additive — the per-token Standard and Contributor tiers on Meta Model API continue, with the Contributor tier offering materially lower rates in exchange for allowing Meta to use prompts and outputs for model improvement. Meta also says it is accepting requests for zero-data-retention, an enterprise feature Wang highlighted in August.
Why it matters
Meta becomes the fourth major closed-source lab to ship a dedicated terminal coding agent with a subscription model, joining Anthropic, OpenAI, and xAI. The emphasis on harness maturity — session messaging, workflow orchestration, and an embedding SDK — signals that differentiation in this category is shifting from raw model benchmarks to the developer experience around multi-agent coordination. For teams evaluating coding agents, the practical question becomes whether Muse Code's workflow engine and inter-session messaging reduce the manual coordination overhead that still dominates multi-agent work in competing products.
Our take
Meta is betting that a polished harness with first-class workflow tooling can close the gap on Claude Code and Codex despite Muse Spark 1.2 trailing on Meta's own August benchmarks. The SDK preview and subscription tiers suggest a platform play: if third-party developers build on MSP, Muse Code becomes an infrastructure layer rather than just a CLI. The workflow engine's gated rollout and the absence of new benchmark data are the honest constraints to watch.
