Meta Launches Muse Code: AI Coding Agent for Large Codebases Enters Beta

ProductsTop News8 sources·Aug 5

coding-tools agents meta muse-spark-1-2 muse-code

Summary

• Meta releases Muse Code beta powered by Muse Spark 1.2, a terminal coding agent targeting complex large-scale repos.
• Agent runs 1,000+ tool calls autonomously per session and fans out parallel sub-agents in isolated worktrees.
• Crash-safe local logs enable full session resumption; video input lets agents turn walkthroughs into fully functional web apps.
• Scores 82.9% on Terminal-Bench 2.1; available on macOS and Linux via the Meta Model API.

Adjust signal

Updates

Aug 6

Type	Key Point	Context
Stat	82.9% on Terminal-Bench 2.1	Muse Code scored 82.9% on Terminal-Bench 2.1, a benchmark measuring agent performance on real-world terminal-based coding tasks.
Stat	1,000+ Autonomous Tool Calls Per Session	Each Muse Code session can execute more than 1,000 tool calls without human intervention, enabling sustained long-horizon coding workflows.
Tech Info	Crash-Safe Session Resumption	Local logs record every agent action; if a session crashes or is interrupted, it can be fully resumed from where it left off without losing progress.
Tech Info	Video Input Capability	Muse Code accepts video input — for example, processing a home walkthrough MP4 and producing a fully functional booking website.
Product Launch	macOS and Linux Availability	Muse Code is available on macOS and Linux through the Meta Model API, targeting the widest possible developer platform base.
Tech Info	Muse Spark 1.2 Underlying Model	The model powering Muse Code is Muse Spark 1.2, an updated version of Meta's specialized coding model with replayable event logs for auditability.

Stat

82.9% on Terminal-Bench 2.1

Muse Code scored 82.9% on Terminal-Bench 2.1, a benchmark measuring agent performance on real-world terminal-based coding tasks.

Stat

1,000+ Autonomous Tool Calls Per Session

Each Muse Code session can execute more than 1,000 tool calls without human intervention, enabling sustained long-horizon coding workflows.

Tech Info

Crash-Safe Session Resumption

Local logs record every agent action; if a session crashes or is interrupted, it can be fully resumed from where it left off without losing progress.

Tech Info

Video Input Capability

Muse Code accepts video input — for example, processing a home walkthrough MP4 and producing a fully functional booking website.

Product Launch

macOS and Linux Availability

Muse Code is available on macOS and Linux through the Meta Model API, targeting the widest possible developer platform base.

Tech Info

Muse Spark 1.2 Underlying Model

The model powering Muse Code is Muse Spark 1.2, an updated version of Meta's specialized coding model with replayable event logs for auditability.

Details

Type	Key Point	Context
Product Launch	Muse Code Beta Released	Meta's terminal coding agent available in beta; installed via a single command, powered by the Muse Spark 1.2 coding model.
Tech Info	Parallel Sub-Agent Architecture	For large jobs, Muse Code fans out to sub-agents working simultaneously in isolated worktrees — the developer's working copy is never touched.
Stat	Six Features Built Simultaneously	In testing, Muse Code built six game features at the same time with no collisions between parallel sub-agents.
Market Impact	Takes On Claude Code & Codex	Meta positions Muse Code as a cost-competitive alternative to Anthropic's Claude Code and OpenAI's Codex for developer workflows.
Strategy	Meta Catching Up in AI Harnesses	Meta was considered a straggler in AI coding tools; Muse Code is a direct catch-up move under Meta Superintelligence Labs led by Alexandr Wang.
Industry Update	Enterprise AI Expansion	Meta entered enterprise AI in June 2026 with a customer service agent; Muse Code signals a broader enterprise developer tools strategy.

Product Launch

Muse Code Beta Released

Meta's terminal coding agent available in beta; installed via a single command, powered by the Muse Spark 1.2 coding model.

Tech Info

Parallel Sub-Agent Architecture

For large jobs, Muse Code fans out to sub-agents working simultaneously in isolated worktrees — the developer's working copy is never touched.

Stat

Six Features Built Simultaneously

In testing, Muse Code built six game features at the same time with no collisions between parallel sub-agents.

Market Impact

Takes On Claude Code & Codex

Meta positions Muse Code as a cost-competitive alternative to Anthropic's Claude Code and OpenAI's Codex for developer workflows.

Strategy

Meta Catching Up in AI Harnesses

Meta was considered a straggler in AI coding tools; Muse Code is a direct catch-up move under Meta Superintelligence Labs led by Alexandr Wang.

Industry Update

Enterprise AI Expansion

Meta entered enterprise AI in June 2026 with a customer service agent; Muse Code signals a broader enterprise developer tools strategy.

Muse Code is Meta's first terminal coding agent, entering a competitive market dominated by Anthropic and OpenAI. Updated with new technical benchmarks, autonomous capacity, and multi-modal capabilities.

What This Means

Meta is making a serious play for the AI developer tools market with Muse Code, directly challenging the dominant positions held by Anthropic's Claude Code and OpenAI's Codex. New technical disclosures sharpen the picture considerably: with 1,000+ autonomous tool calls per session, an 82.9% Terminal-Bench 2.1 score, crash-safe session resumption, and video-to-web-app capability, Muse Code is positioned as a robust long-horizon coding agent rather than a simple code-completion wrapper. The parallel sub-agent architecture and video input are technically distinctive features that could appeal to enterprise engineering teams working on large codebases. With Muse Spark 1.2 powering the agent and macOS/Linux availability via the Meta Model API, Meta has a credible developer platform — and the cost-efficiency positioning may attract price-sensitive teams currently paying premium rates for competing products.

Sentiment

Cautiously optimistic among AI practitioners, focused on technical features and pricing trade-offs

@plumberbutt97former plumber, current ai fan · Meta Muse Spark 1.1/1.2 enjoyer and AI practitionerView post

Supportive

“Muse Code isn’t $META releasing another model. It’s Meta building an AI software engineer that can work on real codebases, recover from interruptions, coordinate agents, and execute long-running dev tasks. This is a strategic move.”

@ruchishshahRuchish (Ru) Shah · AI EngineerView post

Mixed

“Meta’s Muse Code contributor tier: give up your prompts and completions as training data, get a 12x discount on tokens. ... Would you take that trade?”

@MarcusSchulerMarcus Schuler · AI and tech writerView post

Skeptical

“Meta released Muse Code, a coding agent priced 21x cheaper if developers let Meta train on their code. Claude still leads Meta's own benchmarks.”

@GEN_TECHCAMPGEN (げん) · AI trend watcher at テックキャンプView post

Impressed

“Metaが、コーディングを任せられるAI「Muse Code」をベータ公開しました。あわせて、これに合わせて鍛えた新モデル「Muse Spark 1.2」も登場です。... 長く走らせるほど、途中で崩れないことが価値になります。「一発できれいに書かせる」から「任せて放っておける」へ。”

Split

~60/40 positive-to-mixed on technical promise vs. data-sharing costs and whether Meta can close the gap with Claude/OpenAI.

Sources

Update history (1)

Aug 6Added new technical details from Alpha Signal and Grok Sweep sources: Muse Code scores 82.9% on Terminal-Bench 2.1; runs 1,000+ tool calls autonomously per session; crash-safe with local logs for session resumption; accepts video input (e.g. walkthrough MP4 → booking website); available on macOS and Linux via Meta Model API; powered by Muse Spark 1.2. Updated tier1_scan, added tier3 rows 7–12, and expanded key_facts.

Similar Events

Meta Launches Muse Personal AI Agent with Free, $20/mo, and $100/mo Plans

Sep 3

Meta's Muse Shopping Agent Coincides With Slide in 'Consumer Inertia' Stocks

6d ago