Muse Code
Meta's terminal-based autonomous coding agent
Cost-conscious developer teams with large, existing codebases who want a terminal-based autonomous agent for long-running, multi-file engineering tasks — best suited to non-sensitive or open-source work on the discounted tier
Founded 2026 · USA
Good Fit If You're...
Large-repo refactors and dependency upgrades
A team with a big legacy codebase delegates a multi-file refactor or framework upgrade to Muse Code, letting it fan out to parallel sub-agents in isolated worktrees rather than doing it by hand.
Long-running, unattended engineering jobs
A developer kicks off a complex task that runs 1,000+ tool calls over many hours, relying on session persistence and crash-safe resume rather than babysitting the process.
High-volume, low-sensitivity batch work
Open-source maintainers use the heavily discounted Contributor tier for bulk test generation and non-proprietary experimentation where trading data access for a steep discount is an acceptable deal.
Cost-conscious teams evaluating alternatives to Claude Code or Codex
Engineering orgs wary of rising agentic-coding costs trial Muse Code specifically for its Standard-tier price advantage over Claude- or GPT-class agents.
Pricing
| Plan | Price | Credits |
|---|
Hidden costs to be aware of:
- No flat monthly subscription — pure usage-based (token) billing, payment method required upfront
- Standard tier: $1.25/1M input tokens, $0.15/1M cached input, $4.25/1M output tokens
- Contributor tier (~12-21x cheaper): $0.10/1M input, $0.002/1M cached input, $0.20/1M output — requires opting in to let Meta train on your prompts and code
- Enterprise zero-data-retention available on request via Meta sales
Pricing last verified: 2026-08-12
Strengths & Limitations
Strengths
- +Persistent, crash-safe sessions — every model call, tool run, and edit is logged and replay-exact, enabling resume after crashes on very long jobs
- +True parallel sub-agent fan-out with isolated git worktrees — large jobs split across simultaneous agents without collisions
- +1M-token context window, co-trained model+harness specifically for tool-calling and whole-repository tasks
- +Aggressive Contributor-tier pricing — 12-21x cheaper for high-volume, non-sensitive work like tests and dependency bumps
Limitations
- -Trails Claude Code (Opus) on Meta's own published benchmarks (e.g. Terminal-Bench 2.1: 82.9% vs 86.7%) — a late entrant competing against an established leader
- -The cheap Contributor tier isn't free — it's a data-for-discount trade: you sign away rights to your prompts/code for model training, a poor fit for proprietary codebases
- -CLI-only, beta-stage, macOS/Linux only — no Windows support, no IDE integration at launch
- -Correctness requires expert review — independent testing found output 'looked correct on the surface' but needed an expert to verify it was actually finished underneath
Similar Tools
Direct link — no affiliate relationship
Last updated: 2026-08-12