Meta AI Research announced Muse Code on August 5, a terminal-based coding agent built by Meta Superintelligence Labs to operate across entire codebases rather than single files. The tool runs on Muse Spark 1.2, a coding-focused update to Meta’s Spark model line, and lets developers hand it planning, implementation and validation work inside real repositories, from a beta release available on macOS and Linux.
The launch places Meta squarely inside a category it had mostly ignored. Claude Code, Codex and Gemini CLI already compete for developers who want an agent living in their terminal instead of a chat window. Meta had built its reputation on Llama, an open-weight foundation other companies build agents on top of. Muse Code is Meta building the agent itself, and pricing that agent against rivals with a year or more of head start in developer habits.
Muse Code’s technical pitch centers on persistence. Rather than spinning up new subagents for each request, the tool keeps background agents running for an entire session, letting them carry context forward and decide on their own when to report results back to the main loop. Meta says this design cuts down on repeated fact-finding during multi-step coding jobs and reduces how often a developer has to step in and redirect it.
The system also keeps a running ledger of everything the agent does: which model was called, which tools ran, what got approved, and what changed in the code. Meta describes the runtime as replay-exact and restart-safe, meaning a crashed session can pick back up from its last logged step instead of starting over. That claim matters for anyone running unattended, multi-hour jobs, though Meta has not published data on how often those sessions actually fail.
Three bundled commands shape how Meta wants the tool used. /plan produces an approval-gated plan before any code changes happen, /grill pressure-tests that plan for weaknesses, and /goal drives the agent toward a stated objective until it is done. The structure follows the plan-then-execute pattern that Claude Code and Codex have already made familiar to engineering teams, rather than introducing a new way of working with an agent.
Meta’s headline proof point is a kernel-optimization exercise: Muse Spark 1.2 ran for more than 1,000 tool calls across sessions as long as 24 hours, tuning GPU kernels for Nvidia Hopper hardware while barred from importing existing kernel libraries. According to Meta, the model produced a hand-built Triton implementation with meaningful speed gains over the baseline. That is an internal benchmark, on an internal task, graded by Meta on its own infrastructure. No independent lab has replicated the result.
The strategic signal here outweighs the benchmark. Meta has spent two years positioning itself as the open-weight alternative to Anthropic, OpenAI and Google, mostly by giving away Llama for others to build on. A proprietary terminal agent, gated behind Meta’s own model rather than released as open weights, is Meta competing directly for the coding subscriptions and API spend that Claude Code and Codex already capture. It reads as a bet that owning the workflow developers touch daily pays off faster than owning the weights underneath it.
Muse Spark 1.2 is live now inside Muse Code and through the Meta Model API with wider global availability. Engineering teams already standardized on Claude Code or Codex should treat this release as a prompt to benchmark, not a reason to switch: run the same long-horizon, multi-hour coding tasks that matter to your codebase before taking Meta’s kernel numbers at face value.
Announced by Meta AI Research on August 5, 2026.
