Meta’s Muse Code Is Here: A Terminal AI Coding Agent You Can Actually Watch Work
For the last two years, the AI coding story has mostly happened inside a chat box or a little sidebar in your editor. You type, it suggests, you tab-complete. Useful, but junior. Meta just took a swing at the next rung: on August 5 it launched Muse Code, a coding agent that lives in your terminal, reads across your whole repository, and works a task from start to finish while you watch. It runs on a new model called Muse Spark 1.2, and the most interesting thing about it isn’t raw horsepower. It’s that you can see everything it does.
Here’s what actually shipped, why the “watch it work” part matters more than another benchmark score, and how to try it today.
What Meta actually launched
Two things, released together. Muse Spark 1.2 is the engine: a coding-focused update to Meta’s Spark 1.1 model, retrained with more compute and tuned specifically for real development work rather than general chat. Meta points to gains in code generation, complex debugging, understanding an existing codebase, and running long, multi-step workflows without losing the plot. Muse Code is the tool built to drive it: a terminal agent, currently in beta, aimed at engineers working across large repositories.
The pairing is the point. Meta trained the agent and the model alongside each other, so Spark 1.2 isn’t a general model bolted onto a wrapper — it was shaped by the exact job Muse Code does. That’s the same lesson the whole industry has been learning: the coding tools that feel good are the ones where the model and the harness grew up together.
The engine: a 1M-token window and steadier hands
Muse Spark 1.2’s headline number is a one-million-token context window. In plain terms, that’s how much the model can hold in its head at once — and a million tokens is enough to load large chunks of a real codebase, its history, and the task description in a single session, instead of squinting at one file at a time. For agent work, that’s not a vanity spec. An agent that can “see” more of your project makes fewer dumb mistakes about how your code actually fits together.
Meta also claims higher first-attempt accuracy and more reliable tool calling — the model’s ability to correctly use the commands, file edits, and tests you give it — at lower latency. Reliable tool calling is the unglamorous thing that separates an agent that quietly gets work done from one that fumbles every third step and needs babysitting. If you’ve been burned by AI coding tools before, you’re not alone: in one recent survey, 84% of developers now use AI coding tools but only 29% actually trust them. Reliability, not cleverness, is the gap Meta is trying to close.
The real headline: an agent you can watch and rewind
Most coding agents are a black box. You give them a task, they churn, and you get a pile of changes with only a vague sense of how they got there. Muse Code’s pitch is the opposite. It’s built for long-horizon, multi-agent workflows — meaning it can spawn helper sub-agents to tackle parts of a big job in parallel — and every one of those moves is recorded in an event log.
Every subagent it spawns, every tool call it makes, every time you steer it or cancel a step: all of it is observable and replayable. You can watch the agent reason in real time, rewind to the moment it went sideways, and understand why it did what it did instead of just accepting the diff. For anyone who has to review AI-written code before it ships — which should be everyone — that auditability is the feature that makes trusting an autonomous agent even thinkable.
How to try Muse Code today
Meta made this deliberately low-friction. Muse Code installs with a single shell command on macOS or Linux, and then it runs right inside your terminal — no new IDE, no separate app to babysit. The rough shape of getting started looks like this:
- Install it with the one-line command from Meta’s developer site, on a Mac or Linux machine.
- Point it at a repo — open your project folder in the terminal, the same place you already run git and your build.
- Give it a real task, not a toy one: “add pagination to the users endpoint and update the tests,” or “track down why this build fails on CI but not locally.”
- Watch the event log as it plans, edits, and runs tests — and steer or cancel the moment it drifts.
- Review before you commit. Read the diff, run your own tests, and only then ship.
Because it’s an early beta, expect rough edges and expect it to get better fast — that’s rather the point of putting it in developers’ hands now.
How it stacks up against the rest
Muse Code isn’t landing in an empty field. GitHub Copilot, Cursor, and a growing wave of terminal agents from other labs are all chasing the same “agent that ships code” prize, and the frontier chat models keep getting better at it too — OpenAI’s latest flagship and Alibaba’s Qwen line both code well out of the box. What Meta is betting on to stand out is the combination: a coding-tuned model, a huge context window, native multi-agent orchestration, and that fully replayable event log. The auditability angle in particular is something most rivals don’t foreground.
There’s also a business-model wrinkle worth naming. Meta has a long history of offering powerful things for zero dollars and learning from how you use them, and a coding agent sitting inside professional developers’ terminals is an extraordinarily rich source of signal. That’s not a reason to avoid it — it’s a reason to be thoughtful about what repositories you turn it loose on, especially anything proprietary or under a client NDA.
The honest caveats
A few things to keep your feet on the ground. It’s beta, so it will occasionally do something baffling. It’s macOS and Linux only for now, so Windows users are watching from the sidelines (WSL aside). And the oldest rule of AI-assisted coding still applies with force: the agent writes, but you’re accountable. A million-token window and a tidy event log make it easier to review the work — they don’t remove the need to. Read the diff. Run the tests. If you wouldn’t have merged it from a human junior without a look, don’t merge it from an agent either. If you’d rather keep your code on your own hardware entirely, that’s a real trade-off worth weighing — there’s a whole wave of small models you can run locally for exactly that reason.
Strip away the launch noise and Muse Code is a genuinely interesting step: not a smarter autocomplete, but a coworker in your terminal that shows its work. The tools that win the next year won’t be the ones that hide the most — they’ll be the ones that let a human stay in the loop without slowing down. Point one at a real task this week, keep your hand on the wheel, and let it take the grind off your plate. That’s the whole game: not AI replacing the developer, but amplifying one.
Sources & further reading:
- Introducing Muse Code and Muse Spark 1.2 (Meta AI Research)
- Meta launches new AI coding tool powered by Muse Spark 1.2 (Reuters)
Related Reading
- GitHub Copilot vs Cursor vs Lovable: Best AI Coding Tools for 2026
- Top AI Coding Assistants in 2026: Beyond GitHub Copilot
- Spec-First AI Coding: Why Writing the Plan Beats Firing Off the Prompt