Meta has launched a new AI coding agent called Muse Code, a terminal‑based tool designed to handle full‑scale software engineering work rather than just suggest snippets, the clearest sign yet that the company wants its own answer to OpenAI’s Codex and Anthropic’s Claude Code.
What Muse Code is
Meta announced Muse Code this week as its first dedicated AI coding agent, entering a crowded market already populated by OpenAI’s Codex‑based tools and Anthropic’s Claude Code. In a blog post and media briefings, the company described Muse Code as an “agentic coding tool” capable of performing “complex software engineering tasks” across large, real‑world codebases: planning changes, writing code, and validating the results.
The agent runs entirely from the terminal and launches in beta for macOS and Linux. As Apple‑focused outlets have noted, there is no graphical IDE plug‑in or standalone app yet, developers interact with Muse Code via a single install command and then drive it through text prompts and built‑in commands on the command line.
Under the hood, Muse Code is powered by Muse Spark 1.2, Meta’s latest large language model tuned specifically for coding and agentic workflows. Spark 1.2 is described as delivering improvements in code generation, complex debugging, codebase understanding and “end‑to‑end developer workflows,” and it is available both through Muse Code and via Meta’s Model API.
How it works inside the terminal
Meta and early technical coverage paint Muse Code as more than a glorified autocomplete. Notebookcheck, and The Register describe it as an “agent orchestrator” that can:
- Read and learn large repositories – ingesting multi‑file projects, building an internal representation of the codebase.
- Plan changes – generating structured plans for features or refactors rather than line‑by‑line edits.
- Write and modify code – editing files in isolated work trees so the developer’s working copy is never touched.
- Debug and run tests – executing test suites, inspecting failures, and iterating on fixes.
- Coordinate multiple sub‑agents – spinning up background agents that tackle different parts of a task in parallel without colliding.
Mark Zuckerberg, in a series of posts cited by The Register, said that when a developer starts a task, Muse Code fires up background agents that share a context file. Each agent works on its own isolated branch of the code, while the main orchestrator logs “every action before execution” so no work is lost, and changes must be explicitly approved. In one internal test, he said, the system was able to build six features for a single game simultaneously “with no collisions among the agents.”
Muse Code also keeps an event history, a detailed log of model calls, tool usage, approvals, and edits, allowing long‑running tasks to resume after a crash or interruption without starting from scratch, an important capability for multi‑hour refactors or migrations.
From the user’s side, installation is a single curl command. 9to5Mac quotes Meta’s recommended macOS command as:
bash
curl -fsSL https://dev.meta.ai/install.sh | bashOnce installed, developers can issue built‑in commands to generate a plan, stress‑test that plan, or continue progressing toward a specified goal, all from the terminal.
Pricing and access
Unlike many competing AI tools that rely on fixed monthly subscriptions, Muse Code uses a pay‑as‑you‑go token model. Notebookcheck and Engadget both report the standard pricing as:
- $1.25 per million input tokens
- $4.25 per million output tokens
Engadget adds that Meta will offer a cheaper “contributor tier” for developers who agree to provide feedback to improve the agent. According to that report, contributor pricing drops to $0.10 per million input tokens and $0.20 per million output tokens, in exchange for sharing usage data and evaluations back with Meta.
Muse Spark 1.2 itself is a closed‑weight proprietary model hosted in Meta’s cloud, a break from the open‑weight approaches the company championed with its earlier Llama releases. The Register notes that Meta “has Muse Spark locked away as a proprietary, closed‑weight model,” though Zuckerberg did not rule out opening it up in future. Developers can access Spark through Muse Code, via Meta’s own API, or through aggregators like OpenRouter.
How it compares to Codex and Claude Code
Meta stacks Muse Code directly against OpenAI’s Codex and Anthropic’s Claude Code. CNBC, TechXplore and other outlets stress that the new tool is designed to “take on” those incumbents by offering similar capabilities, complex planning, code generation and validation, while differentiating on depth of agent orchestration and pricing.
Key points of comparison:
- Interface – Muse Code is terminal‑only at launch, unlike many Codex‑based tools that ship as IDE plug‑ins, and Claude Code’s browser‑based environment. Meta is betting that serious developers are comfortable living at the command line.
- Agent orchestration – While Codex and Claude Code can run tools and keep context, Meta is heavily marketing Muse Code’s ability to coordinate multiple parallel sub‑agents with isolated work trees, plus its event log and crash‑resilience.
- Model openness – Codex and Claude models are closed‑weight; Meta had previously leaned into open weights with Llama but is keeping Muse Spark proprietary for now.
- Pricing – At $1.25/$4.25 per million tokens and an even cheaper contributor tier, Muse Code is positioned as an aggressively priced alternative to competitors’ enterprise plans.
Meta’s announcement highlighted demo projects — including an interactive physical simulation and a Plants vs. Zombies‑style game — built end‑to‑end with Muse Code, meant to show that the agent can handle full projects, not just toy examples. Engadget notes a demo where Muse Code builds a webpage directly from an MP4 video, combining multimodal input with code generation.
Part of a broader AI push inside Meta
The public launch of Muse Code reflects changes that have already been underway inside Meta. Business Insider reported in April that the company is reorganizing engineering work around AI tools, with internal expectations that some teams produce 50–80% of their code with AI assistance. Zuckerberg has spoken about wanting “100x engineers who command armies of AI agents” rather than large numbers of junior developers.
At the same time, Meta has rolled out AI‑assisted coding interviews that give candidates access to an AI assistant in the coding environment and evaluate how effectively they use it, signaling that fluency with AI tools is now part of the hiring bar.
In that context, Muse Code is both a product for external developers and a reflection of Meta’s internal bet: that the future of software engineering will revolve around orchestrating capable coding agents across massive codebases, with humans increasingly focused on directing, reviewing, and integrating those agents’ work.
What to watch next
Muse Code remains in beta, with no Windows support yet and significant questions unanswered about real‑world reliability across diverse stacks. But with Meta Superintelligence Labs now fielding a direct challenger to Codex and Claude Code, competition in the AI developer‑tool market is likely to intensify on:
- Context window size and repository understanding
- Agent orchestration features and safety
- Pricing and usage tiers for individual and enterprise teams
- Openness of underlying models and tooling
For developers, the arrival of Muse Code adds another powerful, and potentially cheaper, option to the growing toolbox of AI coding assistants. For Meta, it marks a new stage in an AI strategy that aims to get inside not just users’ social feeds, but engineers’ terminals.
