Meta enters the terminal coding agent race with Muse Code

Three labs, one terminal\n\nMeta shipped Muse Code on August 5, 2026 — a terminal coding agent powered by a new model called Muse Spark 1.2. It installs with a single curl command on macOS and Linux, plans changes before writing them, and keeps multiple background subagents alive throughout your session instead of spawning a fresh one for each task.\n\nThat puts Meta in direct competition with Anthropic’s Claude Code and OpenAI’s Codex CLI. Three of the biggest AI labs are now building what is essentially the same product: an AI that lives in your terminal, edits your code, and runs tools on your behalf.\n\nThe interesting part isn’t that Meta showed up late. It’s how they showed up.\n\n## The event log is the real innovation\n\nMost coding agents are stateless between turns. If Claude Code crashes mid-refactor, you might lose context. If a long-running task gets interrupted, you start over.\n\nMuse Code takes a different approach. Every model call, tool execution, approval, and file edit gets appended to a local event log — an append-only record of everything that happened during the session. That single source of truth makes the agent replay-exact and restart-safe. Crash mid-task? It resumes precisely where it stopped.\n\nFor developers running multi-hour migrations or large refactors, this matters more than benchmark scores. A coding agent that can’t survive a crash isn’t an agent — it’s a demo.\n\nThe architecture has a second benefit. Because the event log captures every decision, you get an automatic audit trail. You can see exactly which model call produced which edit, which tool run failed, and where human approval was given. That’s useful for teams that need to review AI-generated changes before merging them.\n\n## Co-trained model and harness\n\nHere’s something most coverage glossed over: Muse Spark 1.2 wasn’t trained as a general coding model and then bolted onto Muse Code. The model and the harness were trained together. Meta collected rejection-sampled trajectories from real Muse Code sessions and fed them back into training, meaning the model learned the specific patterns of how Muse Code invokes tools, structures plans, and handles approvals.\n\nThis is the same bet Anthropic made with Claude Code and the Claude model family, and it’s the opposite of what most coding tool startups do — they take an off-the-shelf model and wrap a UI around it.\n\nThe result, in Meta’s testing, is fewer hallucinated tool calls and tighter plan-then-execute behavior. Whether that holds up in production use is another question, but the training methodology is sound.\n\n## The contributor tier changes the economics\n\nStandard Muse Spark 1.2 API pricing is $1.25 per million input tokens and $4.25 per million output tokens — comparable to Muse Spark 1.1 and competitive with Claude Sonnet-tier pricing.\n\nBut the contributor tier is where it gets aggressive. Developers who opt in to share telemetry and help improve the model get API access at more than 10x cheaper than pay-as-you-go rates. For cost-sensitive developers and small teams, that’s a meaningful difference — potentially the difference between using a coding agent daily versus only for high-stakes tasks.\n\nMeta can afford this because they’re playing a different game than Anthropic or OpenAI. They want adoption and training signal. The contributor tier is a data acquisition strategy dressed up as a pricing discount. That’s not a criticism — it’s a smart move for a late entrant trying to close a quality gap fast.\n\n## How it compares to Claude Code and Codex CLI\n\n| Feature | Muse Code (Meta) | Claude Code (Anthropic) | Codex CLI (OpenAI) |\n|---------|-----------------|------------------------|-------------------|\n| Interface | Terminal only | Terminal + desktop app | Terminal only |\n| Crash recovery | Append-only event log | Session state | Session state |\n| Background agents | Persistent across session | Per-task spawn | Per-task spawn |\n| Model-harness training | Co-trained | Co-trained | Co-trained |\n| Discount tier | Contributor (10x cheaper) | None | Batch API discount |\n| Bundled commands | /plan, /grill, /goal | /init, custom commands | /explain, custom commands |\n\nThe comparison reveals Meta’s differentiation: persistent background agents and the event log. Claude Code and Codex CLI are excellent tools, but neither solves the long-running-session-crash problem as explicitly. And neither offers a tier this cheap.\n\nWhere Meta is behind: ecosystem maturity. Claude Code has months of community-built custom commands, MCP integrations, and battle-tested workflows. Muse Code shipped yesterday. The first weeks will be rough.\n\n## What this means for developers\n\nIf you’re already using Claude Code or Codex CLI and it’s working, there’s no urgent reason to switch today. The event log and persistent background agents are real advantages, but they’re advantages you notice during long sessions — not during a 10-minute bug fix.\n\nThe contributor tier is worth testing if cost has been your barrier to using coding agents regularly. At 10x cheaper, the math changes for solo developers and small teams who couldn’t justify the spend before.\n\nThe real winner here is the category. Three well-funded labs competing on agent quality, crash resilience, and pricing means the tools get better fast — for everyone. A year ago, terminal coding agents were a novelty. Now they’re infrastructure.\n\nMeta entering this market late but with a meaningfully different architecture is exactly the kind of competition that pushes the field forward. The event log alone might force Anthropic and OpenAI to take crash recovery more seriously.\n\n## Try it\n\nMuse Code is in open beta. You can install it via the curl command on Meta’s research blog and start a session with a single prompt. If you’ve been holding off on terminal coding agents because of cost or reliability concerns, the contributor tier plus crash recovery makes this worth a weekend experiment.\n\nThe terminal coding agent race just got a third serious contender. Whether Meta catches up to Anthropic’s lead depends on execution over the next few months — but the opening move is stronger than expected.