Meta released Muse Code on August 5, 2026, a terminal-based coding agent aimed at engineers working across large software repositories. The beta agent runs from a single install command and is powered by Muse Spark, Meta's previously released coding model. It marks Meta's most direct move yet against OpenAI's Codex and Claude Code from Anthropic, the two agents that currently define the category.
CEO Mark Zuckerberg announced Muse Code on Wednesday, saying the tool can handle "complete software engineering tasks across large repos," including planning changes, writing code, and validating the results. That framing puts Muse Code in the same bucket as the incumbents rather than the lightweight autocomplete tier of AI coding tools.
“complete software engineering tasks across large repos”— Mark Zuckerberg, Meta CEO
The design choice that separates Muse Code from single-threaded agents is parallelism. When a task is large enough, Muse Code spawns sub-agents that each run in their own isolated worktree, leaving the developer's working copy untouched. Meta says one internal test had the agent build six features for a game at the same time without collisions between the parallel branches.
Key facts
- 01Meta launched Muse Code in beta on August 5, 2026, a terminal coding agent installable with a single command.
- 02The agent is powered by Muse Spark, Meta's previously released coding model.
- 03In internal testing, Muse Code built six features for a game simultaneously by fanning out to parallel sub-agents in isolated worktrees.
- 04Meta is positioning Muse Code on cost against OpenAI's Codex and Anthropic's Claude Code, according to AI chief Alexandr Wang.
- 05In June 2026, Meta entered the enterprise AI market with an agent for customer service and support.
That architecture matters because long-horizon coding tasks are where most agents currently fail — context drifts, edits stomp on each other, and a single broken step can derail a multi-hour run. Isolated worktrees give each sub-agent a clean surface to work against, which in principle should reduce the merge chaos that hits when an agent tries to touch a dozen files at once.
“When a job is big enough, it fans out to separate sub-agents working in parallel in isolated worktrees”— Mark Zuckerberg, Meta CEO
Pricing is the other front. Meta's AI chief Alexandr Wang, who leads Meta Superintelligence Labs, told the Wall Street Journal the goal is to be a cost-competitive option for a broad set of workflows. He did not disclose pricing tiers for the beta, but the message to buyers is explicit: OpenAI and Anthropic have the mindshare, and Meta intends to compete on the invoice.
That framing lands in a market where Codex and Claude Code sit at the premium end. Claude Code in particular has become the reference implementation for agentic coding at large enterprises, and Anthropic just secured up to $5 billion in AMD compute commitments to keep scaling. Meta is betting that a meaningful slice of developers will trade some capability for a lower run cost — a familiar Meta playbook from the Llama rollout.
“We think that for a lot of workflows and a lot of use cases, this can be an incredibly good option, especially from a cost perspective”— Alexandr Wang, Meta AI chief, Meta Superintelligence Labs
Muse Code is the latest step in Meta's aggressive push to catch up in applied AI. In June 2026, the company moved beyond its core advertising use of AI and entered the enterprise market with a customer service and support agent. Meta Superintelligence Labs, the unit Wang runs, has been the vehicle for that expansion, and the group has been on a hiring and product tear through the summer.
The competitive gap is still real. Codex has been deployed inside OpenAI's enterprise customers for months, and Claude Code has become the default agent inside firms that standardized on Anthropic. Muse Code arrives without published benchmark scores, without a public pricing sheet, and in beta — three things that will need to change before large engineering orgs adopt it as a primary tool.
There are also open questions about how Muse Spark, the underlying model, compares on the benchmarks that matter for agentic coding — SWE-bench Verified, long-context recall, and tool-use reliability. Meta has not published those numbers alongside the launch. Until it does, the six-features-in-parallel demo is a promise rather than a comparison, and reviewers will run their own tests.
For Meta, Muse Code is more strategically important than a single product release suggests. Coding agents are the highest-revenue application of frontier models today, and every lab that ships a credible one converts developer usage into infrastructure lock-in. If Muse Code can undercut Codex and Claude Code on cost without sacrificing reliability on real repos, Meta closes a gap that has defined its position in applied AI for the last two years. If it can't, the launch becomes another item on the list of tools engineers tried once and left behind — and the price argument stops mattering.
Working on something we should cover, or seeing a story we missed? Send leads, documents, or feedback to hello@aichatdaily.com. For sensitive tips, see our secure tips page for Signal and PGP options.
Spotted an error? Email hello@aichatdaily.com with the URL and the issue, or read our full corrections policy.
