Meta releases Muse Code beta, its first AI coding agent

Meta Muse Code beta is now available, giving developers a new terminal-based AI agent for handling software work. For investors following Meta’s AI push, the release marks a move into the coding-agent market alongside products from OpenAI and Anthropic, although the available evidence does not establish how Muse Code performs against those rivals.

Meta Chief Executive Mark Zuckerberg announced the beta release on Aug. 5, according to Bloomberg. The company describes Muse Code as its first AI coding agent, powered by the newly released Muse Spark 1.2 model.

A coding agent is software that uses an AI model to carry out multi-step programming work rather than only answer a question or produce an isolated snippet of code. Meta says Muse Code can plan modifications, write code and validate the results across large code repositories. Developers can install the beta on macOS or Linux, according to Meta’s product announcement.

What can Meta Muse Code do?

Meta says the agent can coordinate persistent asynchronous background subagents during a session. The company says those agents can pursue parts of a task while the main agent continues its work, with the goal of reducing repeated information gathering on longer, more complicated assignments.

The company also disclosed a local event log that records model calls, tool runs, approvals and edits. Meta says the design allows a job to be replayed and resumed after a crash, a feature aimed at tasks that may run for extended periods.

Muse Code comes with built-in commands called skills. Meta says /plan creates a plan subject to approval, /grill tests that plan, and /goal directs work toward a specified objective. These are product claims from Meta, rather than independent measures of the tool’s coding quality.

Muse Code and Muse Spark 1.2 are different products

Muse Code is the agent, or coding harness, that works with Muse Spark 1.2. Yahoo Finance described Spark 1.2 as a coding-focused update to Meta’s earlier Muse Spark model. Meta says the new model is available through Muse Code and the Meta Model API, its developer access service.

Meta says it co-trained Spark 1.2 with Muse Code and trained the model on long-running coding assignments. The company also published benchmark results and performance claims, but the supplied reporting does not independently verify comparisons with OpenAI’s Codex or Anthropic’s Claude Code.

How much does Muse Code cost and what happens to developer data?

Alexandr Wang, who leads Meta Superintelligence Labs, told CNBC that developers can use Muse Code through pay-as-you-go access on Meta’s developer page. Wang also said Muse Code has a contributor tier that is more than 10 times cheaper than pay-as-you-go. Users of that tier must opt in to help improve the model, he said.

For companies that do not want their developer data retained for model improvement, Wang told CNBC that Meta was starting to accept requests for zero-data retention. CNBC reported that Wang declined to provide user statistics for the Muse Spark models. The reports supplied no Muse Code user figures.

The launch gives Meta another product to sell to developers as it builds out its AI services. Wang told CNBC that Meta was seeking to differentiate Muse Code and its Muse Spark models on price rather than claimed capabilities relative to Anthropic and OpenAI.

This story draws on original reporting from CNBC.