Muse Code: Meta's Free CLI With a Token Meter

Muse Code is Meta Superintelligence Labs' terminal coding agent, shipped in beta on August 5 2026 alongside the Muse Spark 1.2 model. You install it with one command (curl -fsSL https://dev.meta.ai/install.sh | bash), run it in a project directory with muse, and it plans, edits, and validates multi-file changes across large repositories. Its distinguishing features are persistent background agents that stay alive across the session instead of spawning per subtask, automatic fan-out of parallel subagents into isolated git worktrees, an append-only local event log that records every model call, tool run, approval, and edit (replay-exact and restart-safe), and /plan, /grill, and /goal built-in workflows. Headless mode is muse exec for scripts and CI.

The CLI binary is free to install - but there is no free tier, trial, or bundled allowance anywhere on Meta's pricing, product, or billing pages, and a payment method is required before your first run. Meta bills per token through the Model API with no seat price, no subscription, and no hard spend cap (email alerts only). Two tiers: muse-spark-1.2 Standard at \$1.25 per 1M input / \$0.15 cached / \$4.25 output, or muse-spark-1.2-contributor at \$0.10 / \$0.002 / \$0.20 - roughly 5-10% of standard - in exchange for letting Meta train on your prompts and completions. The contributor tier is rate-limited at 100 requests per minute vs Standard's 3,000, and is only available in select countries.

There is no monthly price - billing is per token and your cost depends on agent volume. Budget math from published rates: a heavy agent day can push 5-10M tokens, so Standard-tier output alone can run \$21-\$42 per day of heavy use. The contributor tier cuts the same usage by 12.5x on input and 21x on output. Compare Claude Code Pro at \$20/mo (subscription with 5-hour windows), Codex included with ChatGPT Plus at \$20, and Cursor Pro at \$20 - Meta is the only major lab selling agent tokens with no plan wrapper, which makes a quiet month almost free and a heavy month uncapped.

It is the same Muse Spark 1.2 model on a data-for-discount trade: you grant Meta training rights over your prompts and completions, and pay \$0.10 per 1M input / \$0.002 cached / \$0.20 output instead of \$1.25 / \$0.15 / \$4.25 - 12.5x cheaper on input, 21x on output, 75x on cache. The trade-off besides privacy: contributor runs at 100 RPM / 3M TPM versus Standard's 3,000 RPM / 4M TPM, is limited to select countries, and the discount only applies to the muse-spark-1.2-contributor model ID. Meta labels the tiers bluntly - 'Used to improve our products' vs 'Not used to improve our products.'

Not natively. Muse Code ships for macOS and Linux only; Windows runs it through WSL, and Meta has announced no native Windows installer or PowerShell-first workflow - one platform behind Claude Code, which installs natively via PowerShell/WinGet, and Codex, whose Windows path needs no Linux layer. Community forks document the WSL route, and a community Claude-Code-commands bridge exists for driving muse from Claude Code. For a native Windows agent experience today, Claude Code, Codex, or a Windows-first harness like VibeFuse are the practical paths.

Meta reports Terminal-Bench 2.1 at 82.9% and DeepSWE 1.1 at 59.3% with the Muse Code harness - but both are vendor-run, and the Terminal-Bench figure was not on the public verified leaderboard when independent reviewers checked after launch. In Meta's own comparison panels Muse Spark 1.2 finished second to Claude Opus 5, and Meta itself calls 1.2 'a moderate improvement over 1.1'. Hands-on reviews (t3.gg's Theo Browne) found it exceptionally fast and cheap for read-heavy work - a 222-PR audit for about \$0.10 on the contributor tier - but unreliable for long autonomous builds: hallucinated plans, broken integrations, and recovery failures. The model is heavily RL-trained on Meta's own harness, so expect degraded behavior through third-party harnesses like OpenCode.

Different meters for different jobs. Claude Code is a mature, Windows-native agent behind a \$20/mo Pro subscription with verified frontier benchmarks and the deepest ecosystem; Muse Code is a beta, macOS/Linux-only agent with no subscription at all - per-token billing that is either 12.5x-21x cheaper (contributor, Meta trains on your data) or comparable-to-pricey (Standard \$1.25/\$4.25). Meta's own benchmarks put Muse Spark 1.2 second to Claude Opus 5. Muse Code's genuinely different engineering: persistent background agents, automatic worktree fan-out, and the replay-exact event log. Claude Code wins on verified quality, platform coverage, and ecosystem; Muse Code wins on token price and the auditable event log.

Muse Code is another one-vendor CLI with one meter - VibeFuse is the free Windows AI harness that runs Claude Code, Codex, Gemini CLI, Cursor, and any Model-API agent as widgets over your real repositories. VibeFuse costs \$0 with no harness meter: your vendor accounts do the billing, and you switch models per task instead of being locked to muse-spark-1.2. When a quota wall or rate limit hits one widget, route the task to another vendor's CLI on the same canvas. VibeFuse also runs an open-source marketplace where creators sell widgets, skills, and styling packs for revenue share - pair it with VocalFuse for local voice dictation while you build.