Muse Delegation

Table of Contents

• Overview
• When To Use
• When NOT To Use
• Prerequisites
• Quick Start
• Smart Delegation
• Muse-Specific Details
• Exit Criteria

Muse Code Delegation

Overview

Muse Code is Meta's terminal coding agent, released August 2026 and built on Muse Spark 1.2. It runs either as an interactive TUI or as a one-shot headless command, and the headless form is what delegation uses.

The binary is muse. Meta publishes no npm or PyPI package for it: the documented install is a script served from Meta's own domain.

When To Use

• delegation-core selected muse for the task
• Repository-scale reading where Muse Spark 1.2's 1M-token context helps
• Work that should run unattended, since muse exec is designed for CI

When NOT To Use

• META_API_KEY is unset and the run cannot be interactive
• A specific model must be selected. Meta documents no --model flag for exec, so the CLI default stands
• Structured output is required. No output-format flag is documented for exec

Prerequisites

Installation

curl -fsSL https://dev.meta.ai/install.sh | sh

This pipes a remote script into a shell. Read https://dev.meta.ai/docs/muse-code before running it. The guided installer offers the same command and shows its source first:

make -C plugins/conjure delegate-setup

Authentication

Meta documents the environment variable as the CI path:

export META_API_KEY="<your-key>"
muse exec "Run the test suite and summarize failures."

A bare muse prompts for browser sign-in on first run, which is not usable unattended. There is no muse auth status command, so the service declares an empty auth probe and verification checks META_API_KEY instead.

Quick Start

Using the shared delegation executor

uv run python scripts/delegation_executor.py muse "Summarize this module" \
  --files src/

Through the Makefile

make -C plugins/conjure delegate-muse PROMPT='Explain this design'

Direct CLI usage

muse exec "Explain the retry logic in this package"

Smart Delegation

muse declares the strengths code_execution and large_context, and ranks behind gemini, qwen, minimax and glm in the candidate order. It declares no model ids, so smart_delegate leaves model selection to the CLI rather than passing a flag Meta does not document.

Muse-Specific Details

The prompt is positional. A command ending in a flag would consume the prompt as that flag's value, which is why the service config sets prompt_flag=None and the War Room expert's command ends at muse exec.

Two claims circulating in third-party guides are not in Meta's own documentation and are not relied on here: a --json flag streaming JSONL events to stdout, and muse auth set --api-key-stdin. The confirmed JSONL reference is to on-disk session logs. Treat both as unverified until Meta documents them.

Exit Criteria

• muse resolves on PATH and muse --version answers
• META_API_KEY is set, or the operator accepted an interactive sign-in
• The built command is muse exec <prompt> with the prompt as the final argument and no prompt flag before it
• File context was inlined into the prompt rather than passed as @path references
• Output saved to delegations/muse/YYYYMMDD_HHMMSS.md
• This skill ran because delegation-core selected muse, not by default