Meta AI
Free
~6 min read
New · August 6, 2026
Meta just walked into the coding-agent race: meet Muse Code.
On August 5, 2026, Mark Zuckerberg announced Muse Code — Meta's first AI coding agent, now in beta. It lives in the terminal, takes on complete software-engineering tasks across large codebases, and is powered by Muse Spark 1.2, a coding-focused update to the model behind Meta AI. The launch puts Meta in direct competition with Anthropic's Claude Code and OpenAI's Codex. And it comes with the most interesting pricing decision of the year: a tier that costs almost nothing — if you let Meta train on your code. Here's what shipped, and the trade to understand before anyone on your team types the install command.
First, what's a "coding agent"? (60 seconds)
A coding agent is an AI you hand a task, not a question. Instead of suggesting a line of code, it plans the change, edits the files, runs the checks, and reports back — working across an entire project the way a contractor works across a whole job site, not one brick. Claude Code and Codex defined the category; Muse Code is Meta's entry. If nobody at your business writes software, you'll never touch it — but the pricing model it introduces is something every AI buyer should understand, which is why this lesson is free.
What Muse Code actually does
Meta says
In plain English
"Complete software engineering tasks across large repos"
Give it a goal — fix this bug, add this feature — and it plans the changes, writes the code, and validates the results across a big, real-world codebase.
Persistent background agents
Specialized helpers keep studying your codebase between tasks, so the agent starts every job already knowing the project instead of re-learning it from scratch.
Local event log
Everything it does is journaled on your machine. If it crashes mid-task, it picks up where it left off instead of starting over.
One-line install
A single terminal command sets it up, on macOS and Linux. There's no app or website — it's a tool that lives where developers already work.
Under the hood is Muse Spark 1.2 — Meta's first coding-focused revision of the Muse Spark model family. That matters beyond this one tool: model improvements tend to flow back into the Meta AI assistant everyone else uses. The model story so far is in Muse Spark 1.1, explained.
The 30-cent trade
The headline price is the lesson. Muse Code's "contributor tier" charges just
$0.30 per million total tokens — a fraction of what comparable agents cost — in exchange for
permission to use your code to train Meta's future models. That's not a hidden clause; it's the deal, stated plainly. It's also the clearest example yet of a pattern worth teaching your whole team:
when an AI service is dramatically cheaper than its competitors, check what you're paying with instead. For a hobby project, the trade may be fine. For your employer's codebase or a client's, it almost certainly is not yours to make — that code is confidential, and "the AI discount tier" is not an approved disclosure channel. Same instinct as
the Meta AI privacy lesson: Meta's products are inexpensive because data is the business.
How to try it — or watch it
Developers: Muse Code is in beta on macOS and Linux, installed from the terminal with one line — details are in Meta's announcement and developer docs. Everyone else: the thing to watch is whether the contributor-tier model spreads. If "discount for training data" works here, expect other AI vendors to try it — and expect it to show up in tools your team uses. Beta caveat: features and pricing details can shift while Meta is still rolling this out.
Muse Code — quick answers
What is Muse Code?
Meta's first AI coding agent, announced in beta by Mark Zuckerberg on August 5, 2026. It runs in the terminal and handles complete software-engineering tasks — planning changes, writing code, validating results — across large codebases. It's powered by Muse Spark 1.2.
How much does it cost?
The launch headline is the "contributor tier": $0.30 per million total tokens, in exchange for permission to use your code to train future Meta models. Meta hadn't published full standard-tier pricing at launch, and beta pricing can change.
What platforms does it run on?
macOS and Linux, installed with a single terminal command. It's a terminal tool, not an app — the same territory as Claude Code and OpenAI's Codex CLI.
How is it different from Claude Code or Codex?
Same category, three differentiators: persistent background agents that keep building knowledge of your codebase between tasks, a local event log that lets it resume after a crash, and the data-for-discount contributor pricing — which none of its rivals offer.
Should I use the contributor tier?
Only for code you're genuinely free to share — personal projects, open source you own. Never for employer or client code without explicit permission: the discount is paid for with your codebase.