🎧 Prefer to listen?
One sentence in an install doc is doing more damage to your excuses than any course ever will: “an AI agent that reads your entire codebase.” Not a chatbot that guesses at your project. An agent that opens every file, holds a million tokens of context, plans the work, then hands pieces of it to sub-agents.
That tool is real. It is called Muse Code, Meta shipped it on October 12 alongside its Muse Wonder frontier models, and it is Meta’s most aggressive move into coding AI since Muse Spark 1.1 launched earlier this year. Back then my advice was “watch it, don’t switch.” This time Meta shipped an actual product, not just an API — so I spent a weekend pointing it at a real repository to see if the claim survives contact with a non-coder.
What Meta’s Muse Code actually is
Muse Code is a terminal tool — you type commands, it edits your project. Meta describes it as “multi-agent coding in your terminal,” enabled by two models: Muse Wonder and Muse Wonder Full. According to Meta’s model page, both carry a 1-million-token context window, which is where the “reads your entire codebase” promise comes from. A million tokens is roughly 30,000 lines of code, so the agent can hold most small-to-medium projects in its head at once instead of sampling a few files like older tools do.
Three capabilities matter for you:
It plans and delegates. Give it a big task and it breaks the work down, spins up sub-agents to handle pieces in parallel, and recovers on its own when a step fails. That is the difference between a tool that writes snippets and one that finishes a job.
It remembers. Muse Wonder ships with Active Memory — it carries context across conversations. When you fix something on Monday and ask about it Friday, it knows what you did.
It runs long. Meta built it for multi-hour jobs, not quick one-liners. In my tests it handled a 40-minute refactor without hand-holding, which matches what independent reviewers found on long-running agentic tasks.
The pricing tells you where Meta thinks the value is. Muse Wonder costs $4 per million input tokens and $20 per million output tokens on the API — right in the range covered in the AI coding price war breakdown. But the flagship Muse Wonder Full is not on the API at all. You get it only through a Muse Code subscription, at $50 or $200 a month.
What I actually did
Full disclosure on where I sit: I run this site on Hugo, GitHub, and Vercel without writing code myself — the same setup described in start here. Every change to the repository is made by AI agents under my direction. That makes Muse Code’s pitch a direct test of my daily workflow.
Here is what the weekend looked like:
I created an account at developer.meta.com, subscribed to the $50 Muse Code plan, and ran the install script in a terminal. The script took about two minutes. Then I pointed the agent at a live repository and gave it three jobs, smallest to scariest: fix every dead internal link, add a new content section matching the existing page structure, and audit the whole build config for anything that would break a deploy.
The link fixes came back in under five minutes, all eleven of them, each with a one-line explanation of what pointed where. The new section took about twenty minutes including a self-check where the agent verified its own output against the site’s templates — the behavior Meta markets as Active Focus. The build audit found two misconfigured settings that had been silently degrading image optimization for months. Nothing in my previous tooling had ever flagged them.
Do this yourself
If you maintain a site or project built with AI assistance, here is the version of this worth twenty minutes of your evening:
- Decide if you have a codebase worth reading. If your project lives in a GitHub repo with real history, yes. If it is a Notion page and a Zapier account, this is not your tool yet — start with a simpler automation.
- Create the account and pick your plan. developer.meta.com for the subscription; the $50 tier includes Muse Wonder Full.
- Install Muse Code and run it on a copy first. Clone your repo, run the agent against the clone, never the live project.
- Give it one small, boring task. Broken links, outdated dependencies, a missing page. Judge it on something cheap to verify.
- Read every diff before you accept it. This is the whole discipline. The agent proposes changes as diffs; you approve them line by line. If you cannot tell what a change does, reject it and ask for a plain-English explanation first.
- Only then merge to your live repo. Small, reviewed, reversible steps — the same rules from the agent context budget guide apply here.
What failed
Three things, and they matter more than the wins.
Design work is genuinely bad. The same independent reviewer who found Muse Wonder excellent at agentic work called its frontend output “about 18 months behind” the top models — generic layouts, broken buttons, ignored style guidelines. I saw the same pattern when I asked it to restyle a page template. The logic of the change was correct; the taste was not.
The setup assumes you are a developer. Everything happens in a terminal. There is no polished app for this yet, and Meta’s docs read like they were written for engineers. If the word “terminal” makes you close the tab, that instinct is currently correct.
The trust math is uncomfortable. Meta’s own safety documentation reports the model mishandled personal data in 3.4% of test cases — roughly one failure in thirty. Meta also broke from its open-source Llama tradition here: neither Wonder model has downloadable weights, so everything runs on Meta’s servers. Anthropic has had its own agent security incidents, so nobody is clean — but you should assume anything in the repo can end up in Meta’s cloud.
Who should skip this
Skip it if you have no repository to point it at — there is no consumer-friendly version of Muse Code, and the Muse personal agent app covers the non-coding use cases instead. Skip it if your current coding agent already works; the benchmarks have Muse Wonder Full at 55.8 on the Artificial Analysis index, still behind Claude Opus 5.5’s 57.6, so switching is churn, not progress — diversify instead. Skip it for anything visual. And skip it if you are unwilling to read diffs, because an agent this capable running unreviewed is how small mistakes become outages.
The honest target: you maintain a repo, you are comfortable approving changes you understand, and you want a second strong agent in your stack. That is who Muse Code is for today, and for that group it is the cheapest path to a frontier model — $50 a month gets you the model that everyone else has to rent token by token.
The deeper story is that Meta stopped giving its best work away. The open-weights era paid for your cheap backups; the subscription era pays for a gigawatt data center in Ohio. Watch what that does to pricing across the industry over the next quarter — the tools I actually use every day will tell you whether it earns a permanent slot.
The bottom line
Muse Code is the first Meta coding product you can actually use rather than read about, and the “reads your entire codebase” claim held up on a real repo for link fixes, structure work, and config audits. It fails at design, assumes a terminal, and asks for a level of review discipline most beginners do not have yet. If you qualify, start it on a clone with one boring task tonight.
If you are earlier in the journey, start here with the tools that require zero terminal time — Muse Code will still be here when your project grows into it.