Meta's Muse Code: what it means for custom software projects

On August 5, 2026, Meta launched Muse Code, a terminal-based AI agent that can plan, write, and validate changes across entire codebases, alongside the model behind it, Muse Spark 1.2, as reported by TechCrunch and VentureBeat. It’s Meta’s official entry into the coding-agent race, a space previously dominated by Anthropic’s Claude Code and OpenAI’s Codex. At Evicron, an AI and custom software studio based in Barcelona, we track these tools closely because they change the timeline and cost of any custom build — but this launch also surfaces a detail any business hiring a developer should check before signing.

What Muse Code actually does differently

Muse Code installs with a single command on macOS or Linux and runs as a persistent agent: it doesn’t just autocomplete lines, it plans changes, coordinates multiple subagents per task, and validates the result across the whole repository, not just the open file. Meta trained it together with the Muse Spark 1.2 model specifically for this pairing, with gains in large-codebase comprehension, complex debugging, and end-to-end project generation. In practice, it’s the same category of tool as Claude Code or Codex CLI: an agent that works for hours unsupervised, not a glorified autocomplete.

The detail that didn’t make the headline: the “Contributor” tier

Here’s the part that actually matters to a business. Muse Code ships with two very different pricing tiers, according to several independent breakdowns of the launch: a standard tier ($1.25 per million input tokens, $4.25 output) and a noticeably cheaper “Contributor” tier ($0.10 input, $0.20 output) in exchange for letting Meta use your prompts and generated code to train its own models. That tier is also limited to select countries and explicitly discouraged by Meta for proprietary code, trade secrets, or anything under an NDA. Meta has started accepting zero-data-retention requests for enterprise customers who don’t want to share anything — which itself confirms that Meta draws a line between “cheap” and “safe for professional use.”

What this means if you’re about to commission custom software

The speed at which these agents now produce full projects isn’t a lab promise anymore — it’s why a custom build today can have a clickable demo by week two instead of a month in. But the catch in Muse Code’s free tier illustrates a real risk that isn’t unique to Meta: any vendor coding your project with an AI model set to its cheapest configuration could, without telling you, be training a third party’s model on your business logic, API keys, or customer data buried in that code. It’s the same kind of leak we already documented when we looked at the security risks of vibe coding: AI speeds up development, but the configuration it runs under decides whether that’s an advantage or a leak.

What to ask your vendor before you start

1. Which AI models they use to write code, and in what configuration — standard, or one of the tiers that share data for training.
2. Whether your code, keys, or customer data ever leave their environment, even to “improve the model.”
3. Who owns the final code — yours should be yours, with neither the vendor nor the model provider retaining rights to it.
4. What happens if they switch AI models or providers mid-project — the contract should protect you from that shift, not leave you exposed to it.

How we handle this at Evicron

We use state-of-the-art coding agents — from Anthropic, OpenAI, and other providers in our stack — to speed up delivery in our custom software service, but never on tiers that share a client’s code to train third-party models: that’s a line we don’t cross, discount or not. If you’re not sure what to check before signing with a development vendor, our AI consulting covers that alongside the rest of the technical stack.

The bottom line

Meta has entered the coding-agent race with Muse Code, a genuinely capable tool for large codebases — but its cheapest tier requires sharing your code to train its models, a detail any business commissioning custom software should ask about explicitly, whichever vendor or tool they end up using.

Want us to check what confidentiality guarantees your current or upcoming project actually has? Get in touch: the first discovery session is free, and we reply within 24 hours.