← Back to Harnesses
Meta AINotable
Muse Code
Released Aug 5, 2026
Meta's terminal coding agent, launched in beta with Muse Spark 1.2; it runs a main agent alongside persistent background subagents.
- Interface
- CLI
- Pricing
- Subscription
- Models
- Vendor's own models
- Best Terminal-Bench
- 14.55% (v4.0) · Muse Spark 1.3
- Open source
- No
- Licence
- Proprietary
Terminal-Benchv4.0 · per harness + model pairing
| Muse Spark 1.3 | 14.55% |
Strengths & weaknesses
Strengths
- Approval modes and an OS sandbox on by default
- Subagents, skills, hooks and MCP; event log lets sessions resume after a crash
- Headless `muse exec`, session protocol and TypeScript SDK
Weaknesses
- Muse Spark 1.3 scores 14.6% on Terminal-Bench 4.0
- Workflows unavailable on some builds/platforms; Windows lacks voice input and session messaging
- Default contributor mode lets Meta train on your content (can be turned off)
Notable: appears on the official Terminal-Bench leaderboard (tbench.ai, Terminal-Bench 4.0 — the only version tbench.ai currently shows), best rank #32; only five distinct agents are listed there, so every one is top-10. Scores are TB 4.0 resolution rates (best per model across reasoning efforts), fetched 2026-10-14 — not comparable with older Terminal-Bench 2.x figures. Default model muse-spark-1.2; the changelog names Muse Spark 1.2 and 1.3 and allows any model id 'on the same provider', hence 'vendor'. No licence or repo stated, so openSource=false.