Meta ships Muse Code, a terminal coding agent built for multi-agent workflows

Meta shipped Muse Spark 1.2 and Muse Code on August 5, 2026. Muse Spark 1.2 is a coding-optimised model. Muse Code is the agent that runs on it, and it is available today in early beta. Meta describes the combination in its announcement as a coding system for long-running, multi-agent work with an auditable record of what happened.

What Muse Code actually does

Muse Code is a terminal coding agent. Meta calls it "a purpose-built coding agent optimized for long-horizon, multi-agentic coding workflows and transparent auditability." Its subagents are observable. Its tool calls can be inspected, and its event logs can be replayed. These details distinguish the product's stated focus: the work is not presented as one hidden model response, but as an agent process whose activity can be followed.

The native CLI can be installed with:

curl -fsSL https://dev.meta.ai/install.sh | bash

Muse Spark 1.2 can be reached in three stated ways: through the Muse Code native CLI, the Meta Model API, and OpenRouter. Meta's product-page summary is concise: "Multi-agent coding in your terminal. Built for Muse Spark. Now in beta." The model identifier used for pricing is muse-spark-1.2.

The standard token prices are:

• $0.15 per 1 million cached input tokens
• $1.25 per 1 million input tokens
• $4.25 per 1 million output tokens

Muse Spark 1.2 is a reasoning model. Thinking tokens are billed separately, so the listed input and output rates are not the complete cost description for every run. The context window has a maximum of 1 million tokens. Meta also offers zero data retention on request through sales.

There is a separate muse-spark-1.2-contributor tier. It is rate-limited by tokens in a rolling five-hour window rather than by request count. Meta says this tier is available in select countries.

What it means when picking a coding agent

Muse Code now joins the terminal and coding-agent options in the directory. The verified differentiators to evaluate are specific: it is terminal-first, built around multi-agent workflows, exposes subagents and tool calls, keeps replayable event logs, publishes token pricing, and supports a context window up to 1 million tokens. Those are the useful criteria to test during the beta.

The relevant alternatives include Claude Code, Cursor, GitHub Copilot, and Gemini CLI. The facts available for this launch do not support a feature-by-feature ranking against those products. Instead, the practical comparison is whether Muse Code's terminal workflow, multi-agent structure, and audit trail fit the way a team reviews agent activity. Pricing can be compared using the published Muse Spark 1.2 token rates, while remembering that thinking tokens are separate.

The 1 million-token maximum is also a concrete selection constraint, not a promise that every job should use the full window. The contributor tier's rolling five-hour token limit is another distinct operating rule. Buyers considering that tier need to account for both its token-based limit and its availability only in select countries.

For a wider view of products in this area, the Coding category collects the directory's coding tools. Muse Code should be evaluated as a beta entrant with a clearly stated architecture and billing model, rather than through unverified benchmark or discount claims.

Limits and what is not available yet

Muse Code is available today, but only as an early beta. Meta has shipped the native CLI and provides model access through the Meta Model API and OpenRouter. The announcement does not provide a later general-availability date. It also does not state an operating-system list, so Windows support should not be assumed from this release.

The contributor tier is not described as globally available; Meta limits it to select countries. Zero data retention is not presented as the default. It is available on request through sales. The 1 million-token context is a maximum, and reasoning adds separately billed thinking tokens.

Meta also lists Muse Glimmer, a 30-billion-parameter local agent model, on the same product hub. The Muse Code announcement does not establish additional availability details for Muse Glimmer, so it should not be treated as part of the Muse Code beta without further information.

The next points to watch are a move beyond early beta, broader country access for the contributor tier, and any published operating-system support. Clearer availability details for Muse Glimmer would also show how Meta intends the local model and terminal agent to relate.