Meta Just Shipped Its First AI Coding Agent

Meta released Muse Code on August 5, 2026, and it changes who Meta is competing with. Muse Code is a terminal-based AI coding agent that plans changes, writes code, and checks its own work across large repositories. It runs on Muse Spark 1.2, a new coding-focused model Meta trained specifically to pair with this agent. If you build or maintain websites for a living, this matters. Muse Code puts Meta directly against Anthropic’s Claude Code and OpenAI’s ChatGPT Codex, and it gives developers another serious option for offloading repetitive engineering work.

You can install Muse Code on macOS or Linux with one terminal command. There’s no Windows version yet, and unlike Claude Code or Codex, Muse Code doesn’t have a dedicated app interface, as 9to5Mac noted in its coverage of the launch. It lives in your terminal, which tells you a lot about who Meta built this for first: developers who already live in the command line.

What Muse Code Actually Does Differently

Most coding assistants spin up a helper, finish one task, and forget everything. Muse Code works differently. It runs a main agent loop alongside async background agents that stay active for your entire session instead of respawning for each subtask. That persistence means the agent doesn’t re-read your codebase from scratch every time you ask for a change, and it can push forward on the next step without you prompting it again.

Muse Code also keeps a local event log of every model call, tool run, approval, and edit. If the agent crashes mid-task, it resumes exactly where it left off instead of starting over. That’s a real fix for a common frustration with long-running AI coding sessions. On top of that, Muse Code ships with three built-in commands: /plan turns a request into an approval-gated plan before any code gets written, /grill stress-tests that plan, and /goal keeps the agent working toward a defined objective across many steps.

This kind of agent-driven auditing and building is becoming a bigger part of how sites get made and checked. If you’ve followed how Google’s Lighthouse tool now grades whether your website is agent-ready, you’ll recognize the pattern. AI agents aren’t just browsing the web anymore. They’re building and reviewing it too.

How Muse Code Compares on Price and Access

Muse Code is in beta, and Meta co-trained Muse Spark 1.2 with the agent itself so the two work well together instead of bolting a general-purpose model onto a coding harness. Pricing for the standard tier runs $1.25 per million input tokens and $4.25 per million output tokens, with cached input priced lower at $0.15 per million. Muse Spark 1.2 is also available through the Meta Model API with expanded global access, separate from the Muse Code agent itself.

Tools ship and change fast right now, and web professionals need to track more than just AI agents. You saw the same pace this week with WordPress 7.1 entering its release candidate stage ahead of its August 19 launch, and with Webflow retiring its legacy editor. None of these tools wait for you to catch up.

What This Means for Your Web Development Workflow

If you run an agency or build sites for clients, Muse Code’s biggest promise is handling whole-repository tasks instead of single functions. Meta says the agent can plan a change, write it, and validate the result without you babysitting every step. In one case study, Meta had the model iteratively optimize GPU kernels over more than 1,000 tool calls across 24 hours, a task that would take a human engineer far longer to attempt.

That said, this is a beta release. You should treat Muse Code the way you’d treat any new coding agent: useful for scaffolding, refactoring, and repetitive work, but not a replacement for reviewing what ships to production. The approval-gated planning step in /plan exists for exactly that reason. Use it.

Frequently Asked Questions

What is Muse Code?

Muse Code is Meta’s first AI coding agent, released in beta on August 5, 2026. It runs in your terminal on macOS or Linux, and it can plan, write, and validate code changes across large software repositories using Meta’s Muse Spark 1.2 model.

How much does Muse Code cost?

Muse Code’s standard tier costs $1.25 per million input tokens and $4.25 per million output tokens, with cached input priced at $0.15 per million. It’s currently free to install and try in beta, though usage against the model still consumes credits.

Is Muse Code available for Windows?

Not yet. Muse Code currently supports macOS and Linux only, and it runs entirely from the terminal with no dedicated app interface, unlike competitors such as Claude Code and ChatGPT Codex.

Muse Code Signals Where AI-Assisted Web Development Is Headed

Muse Code won’t replace your judgment about what belongs in a client’s codebase, but it does show where the tooling is going. Persistent background agents, crash-safe sessions, and approval-gated planning point toward AI coding tools that handle more of the grunt work while you focus on architecture and review. Whether you adopt Muse Code specifically or stick with Claude Code or Codex, the direction is the same: AI agents are moving from autocomplete to genuine engineering collaborators. Keep an eye on how Meta develops this beta, because the pricing and capabilities will likely shift fast.