Meta has launched Muse Code, a terminal-based artificial intelligence coding agent designed to work across large software repositories rather than respond only to isolated code prompts.
Released in beta on August 5, the tool is powered by Muse Spark 1.2, Meta’s latest coding-focused model. Muse Code can inspect a repository, plan changes, write or modify code, use development tools and check whether its work produces the intended result. It is available for macOS and Linux through Meta’s developer platform.
The release moves Meta into direct competition with coding agents from Anthropic, OpenAI and other AI developers. It also reflects a wider shift from autocomplete tools toward agents expected to carry out longer assignments with less continuous direction.
Built for Repository-Scale Work
Muse Code targets tasks that require an understanding of relationships across files, dependencies and services. A developer can give it a broader objective, such as tracing a bug, changing an application feature or refactoring a system, instead of manually supplying every relevant code fragment.
Meta said the agent can handle “planning changes, writing code, and validating the results” across large repositories. The company described the release as its “next step toward the frontier,” with larger and more capable models planned.
That focus addresses a persistent weakness in AI coding tools. Producing a plausible function is different from safely changing an established application, where one edit can affect tests, interfaces, build processes and code maintained by other teams. Muse Code attempts to manage that wider context by pairing a main agent with specialized workers.
Persistent Agents Work in Parallel
The system’s central design feature is its use of asynchronous background agents. These subagents remain active throughout a session instead of being created again for each small assignment.
While the main agent manages the overall task, background agents can investigate different sections of a codebase, gather information and complete follow-up steps. They decide when to report findings to the main process. Meta says this reduces repeated investigation, latency and the manual steering needed for difficult work.
The architecture could help when a problem spans several components. One agent might inspect test failures while another searches for related implementations or configuration files. Parallel work does not guarantee accuracy, but it may shorten the time needed to build a useful view of a large repository.
Muse Code also includes three default skills. The /plan command creates a developer-approved plan, /grill challenges that plan to expose weaknesses, and /goal continues working toward a defined result. The approval step lets developers review the proposed approach before code is changed.
A Runtime Designed to Recover
Long-running agents create a practical problem: a crash or interrupted session can erase progress and force repeated work. Meta addresses that risk with a local event log.
Every model request, tool execution, approval and code edit is appended to the log. Meta calls it a “single source of truth” that makes the runtime restart-safe, allowing Muse Code to continue from the point where it stopped.
The record may also make agent behavior easier to review by showing which actions were taken and in what order. Meta’s announcement did not specify how long logs are retained, whether organizations can export them or what administrative controls will be offered to enterprise teams.
Muse Spark 1.2 Powers the Agent
Muse Spark 1.2 updates Muse Spark 1.1, which Meta released for developer testing in July. The company says the new version improves code generation, complex debugging, codebase comprehension and end-to-end development workflows.
The model and agent were co-trained rather than built as separate products. Training covered tool-use trajectories, subagent behavior and context-compaction methods intended to keep the system focused during long assignments.
Meta also used Muse Spark 1.1 to create difficult coding environments and instruction-following exercises. The earlier model evaluated candidate solutions, helping the company produce additional training material intended to improve Muse Spark 1.2’s ability to follow complex requirements.
Meta says Muse Spark 1.2 was trained for whole-repository generation, extended development projects and automated research. In one case study, the model completed more than 1,000 tool calls over a period of up to 24 hours while writing, compiling, profiling and improving GPU kernels.
Benchmark Claims Need Context
Meta evaluated the Muse Code and Muse Spark 1.2 pairing on several agent and software-engineering benchmarks. Its Terminal-Bench 2.1 evaluation covered 89 terminal tasks and measured the average success rate across five attempts.
The DeepSWE 1.1 evaluation included 113 tasks drawn from 91 repositories across TypeScript, Go, Python, JavaScript and Rust. Each proposed patch was applied to a clean copy of the repository and checked using functional tests and regression controls. Internet access was blocked during the evaluation, apart from access to the model endpoint.
Meta also tested the system on an internal collection of 440 tasks based on real company pull requests. The work covered bug fixes, feature development, refactoring, code cleanup and other engineering assignments. Submissions were compiled and assessed against unit tests in dedicated grading environments.
The results require some caution. Meta acknowledged that comparisons with rival models were based on best-effort evaluations and were not always conducted with identical agent software. Its methodology says the tools and system prompts may not have been optimized for third-party models, meaning the comparisons may not show those systems at their strongest.
Two Pricing Routes
Developers can use Muse Spark 1.2 through standard pay-as-you-go pricing. Meta lists the model at $1.25 per million input tokens, $0.15 per million cached input tokens and $4.25 per million output tokens.
A contributor tier costs $0.10 per million input tokens, $0.002 per million cached input tokens and $0.20 per million output tokens. The discount comes with permission for Meta to use prompts and completions to improve its products. Teams working with proprietary code will therefore need to examine the data terms before selecting the cheaper option.
Actual costs will depend on the amount of repository context sent to the model, the length of a session and the volume of generated code. A lengthy investigation across a large project may consume considerably more tokens than a brief coding question.
Meta Enters a Crowded Market
Muse Code arrives as coding agents become an important route for AI companies seeking developer adoption and recurring revenue. Anthropic’s Claude Code and OpenAI’s Codex already let developers delegate multi-step work from terminals and development environments.
Meta is differentiating its product through relatively low API pricing, persistent parallel agents and an agent runtime developed alongside the underlying model. Sustained use will depend less on demonstrations than reliability inside unfamiliar repositories, where a wrong assumption can produce changes that pass basic checks but introduce deeper problems.
The beta is available now, although questions remain around security controls, repository permissions, organizational administration and performance in production systems. Meta says more powerful models and additional agent features are planned.
Muse Code represents another step away from AI as a passive code-suggestion box. Meta is betting that coding tools will be expected to understand projects, coordinate work, survive interruptions and verify what they build. The beta will show whether that ambition becomes dependable help on the large codebases it was designed to handle.
Comments