Key Notes
- Meta released Muse Code in beta as a terminal coding agent for macOS and Linux powered by Muse Spark 1.2.
- Persistent background agents and a restart-safe local event log are designed to support complex, long-running repository work.
- Muse Spark 1.2 is available through Muse Code and the Meta Model API with expanded global access.
Meta has released Muse Code in beta, giving developers a terminal-based coding agent designed to plan, implement and verify complex changes across large software repositories.
The tool runs on macOS and Linux and is powered by Muse Spark 1.2, a coding-focused update to the model Meta introduced earlier this year. Its arrival places Meta more directly against OpenAI‘s Codex and Anthropic‘s Claude Code in the increasingly competitive market for autonomous developer tools.
Meta described Muse Code as its next step toward frontier systems, while acknowledging that larger and more capable models are still in development. Muse Spark 1.2 is available immediately through Muse Code and the Meta Model API with expanded global access.
Meta Enters the Terminal-Agent Race
Muse Code is built for repository-scale work rather than isolated code completion. According to Meta’s official announcement, the agent can inspect a codebase, plan a sequence of changes, write the implementation and validate the result.
That workflow follows the same broad direction as Codex and Claude Code: developers describe an outcome, while an agentic AI system navigates files, runs tools and iterates toward completion. AIstify previously covered Xiaomi’s MiMo Code terminal agent, illustrating how quickly this product category is expanding beyond the leading U.S. AI labs.
Muse Code ships with three default skills. /plan converts a request into an approval-gated plan, /grill challenges that plan before execution, and /goal keeps the system working toward a specified result. The structure gives users explicit control points without requiring them to direct every intermediate action.
Persistent Agents and a Restart-Safe Runtime
One of Muse Code’s distinguishing design choices is its use of persistent asynchronous background agents. These specialized agents remain active throughout a session instead of being created for individual subtasks. Meta says the approach reduces repeated information gathering, latency and the amount of steering required for difficult multi-step work.
The main agent and its background workers communicate through a local event log. Every model call, tool execution, approval and edit is appended to that log, creating what Meta calls a single source of truth. If the program crashes, it can replay the recorded state and resume from the same point rather than restarting the task.
That architecture matters because coding agents increasingly operate for hours rather than minutes. A tool that loses its reasoning state or repeats completed work after an interruption becomes expensive and difficult to trust on production repositories.
Muse Spark 1.2 Targets Long-Horizon Coding
Muse Spark 1.2 was co-trained with the Muse Code environment. Meta increased the training compute devoted to programming, expanded the variety of coding environments and optimized training around planning, context compaction, subagents and the agent’s toolset.
The model is intended to handle whole-repository generation, end-to-end projects, debugging and automated research. It also retains the multimodal AI capabilities of the wider Muse family. In one demonstration, Muse Code interpreted a video tour of a home and produced a working vacation-rental marketing and booking website.
Meta also tested the system on GPU kernel optimization over more than 1,000 tool calls, with runs lasting as long as 24 hours. The agent repeatedly wrote, compiled, profiled and refined kernels for NVIDIA Hopper hardware. The test was presented as evidence of sustained execution, although the results remain Meta’s own evaluation rather than independent verification.
The release follows the original Muse Spark launch in April, which marked Meta’s effort to rebuild its model strategy around agent-based reasoning. Muse Code turns that model family into a dedicated developer product rather than another capability inside the Meta AI assistant.
How Muse Code Challenges Codex and Claude Code
The competitive question is no longer limited to which model writes the best code snippet. Coding agents are increasingly judged on how well they understand a repository, use external tools, recover from failures and remain aligned with a goal over long sessions.
OpenAI has been extending Codex with the ability to learn from developer workflows, while Anthropic has added real-time security reviews to Claude Code. Meta is answering with persistent subagents, replayable execution and a model trained together with its terminal harness.
Muse Code therefore gives Meta a credible entry point, but a beta launch is not the same as proven adoption. Developers will need evidence on reliability, permission controls, security, model behavior in unfamiliar repositories and the cost of long-running sessions before treating it as a replacement for established tools.
The product nevertheless changes the shape of the contest. Meta now has a model API, a dedicated coding agent and an integrated training strategy aimed at long-horizon work. OpenAI and Anthropic no longer face only model competition from Meta; they face a directly comparable developer workflow.
Disclaimer: AIstify is an independent media brand owned and operated by NuvexMedia LLC, publishing news, research, and insights on artificial intelligence, emerging technologies, automation, and related industries. NuvexMedia LLC invests in and collaborates with companies across the AI, technology, software, and digital innovation sectors. These relationships do not influence AIstify’s editorial coverage, and the publication maintains full editorial independence to provide accurate, timely, and objective information. © 2026 NuvexMedia LLC. All rights reserved. This content is for informational purposes only and should not be considered legal, tax, investment, financial, or other professional advice.
