Meta Muse Code & Muse Spark Course – Build AI Agents, APIs, and Full-Stack Apps

Summary
In this comprehensive technical course, instructor Andrew Brown (ExamPro) demonstrates how to develop applications and AI agents using Meta’s Muse models (Muse Spark 1.1 and 1.2, and Muse Glimmer 30B) alongside Meta’s agentic CLI harness, Muse Code. He guides viewers through the developer playground, testing vision transcription, structured JSON schema outputs, and grounded web search, before integrating the models into code via the OpenAI, Anthropic, and LangChain SDKs. The course culminates in an in-depth exploration of the Muse Code CLI, covering autonomous multi-step development, AGENTS.md, settings, session memory, headless mode, sandboxing with Bubblewrap, and Model Context Protocol (MCP) servers.

---

What is shown

• Overview & Architecture Slides [00:00 - 12:10]: Breakdown of Meta’s model offerings (Muse Spark 1.1/1.2, Muse Glimmer 30B), standard vs. contributor pricing tiers, benchmark rankings, and OpenAI/Anthropic wire compatibility.
• Account Setup & Billing [12:10 - 13:45]: Navigating dev.meta.ai, configuring spend limits, viewing API credits, and exploring documentation.
• Playground & Prompt Testing [13:45 - 20:53]: Testing Muse Spark 1.2 on reasoning tasks, including JLPT Japanese grammar checklists and Japanese OCR text transcription from NHK Web Easy screenshots.
• Full-Stack HTML/CSS Prototyping & Mermaid Diagrams [20:54 - 36:40]: Generating a retro PHP-Nuke-style three-column forum website (“TechNuke”) and generating complex Mermaid mindmaps representing community engagement architecture.
• Search Grounding Evaluation [36:45 - 45:56]: Testing live web search grounding against ungrounded generation for automotive parts pricing and Canadian shipping rates across RockAuto and PartsAvatar.
• Structured Output via JSON Schema [45:57 - 59:55]: Defining strict JSON schemas to evaluate Japanese translation attempts, generating structured grading scores, and creating an HTML report card viewer.
• Low-Level Python SDK Integration [59:56 - 74:40]: Pointing the official openai and anthropic Python SDKs directly to https://api.meta.ai/v1 to execute chat completions and message APIs.
• Agent Frameworks [74:41 - 96:45]: Running the Anthropic claude_agent_sdk to scaffold a command-line Ruby game, setting up LangChain (@langchain/openai) in TypeScript, and configuring a bash wrapper to power Claude Code via Muse Spark.
• Muse Code CLI Installation & Full Project Build [104:20 - 156:01]: Installing Muse Code via curl -fsSL https://dev.meta.ai/install.sh | bash, generating AGENTS.md via muse init, and using the /goal command to build a Go backend, SQLite schema with WAL mode, and Docker Compose environment.
• Advanced Muse Code Configurations [156:02 - 180:32]: Switching models (/model), toggling reasoning effort (/effort), setting permissions with --yolo and --disable-approval, managing persistent project memory in .agents/memory/, installing bubblewrap for OS sandboxing, and configuring DuckDuckGo MCP servers.

---

Claims & numbers

• Model specifications: The presenter states Muse Spark 1.1 and 1.2 support text, image, video, and PDF inputs with text outputs and feature a 1,048,576-token context window; Muse Glimmer 30B is an open-weights 30-billion-parameter causal language model released under an Apache 2.0 license designed to run on a single 32 GB GPU (such as an RTX 5090).
• Pricing tiers: The presenter displays Meta Model API pricing:
  • Standard tier: $1.25 / 1M input tokens, $0.15 / 1M cached input tokens, and $4.25 / 1M output tokens (rate limits: 3,000 RPM, 4,000,000 TPM).
  • Contributor tier (data-sharing discount): $0.10 / 1M input tokens, $0.002 / 1M cached input tokens, and $0.20 / 1M output tokens (rate limits: 100 RPM, 3,000,000 TPM).
  • Web search grounding: $2.50 per 1,000 search queries.
• Benchmarks: The presenter claims Muse Spark 1.2 sits in the upper-middle “Goldilocks zone” on GDPval-AA v2 and Vals Index leaderboards, often performing as a top-three cost-effective alternative across tasks.
• API Compatibility: The presenter states Meta Model API endpoints (https://api.meta.ai/v1) are drop-in compatible with standard OpenAI and Anthropic SDKs without requiring dedicated Meta-specific client libraries.

---

Notable quotes

• [00:19]: “When you have a first-class company that has their own models that they’ve trained, and they have a coding harness and they’re entirely vertically integrated, they are a plausible option that we could utilize in the AI space.” — Andrew Brown
• [06:56]: “In terms of benchmarks, I would describe Meta’s as a Goldilocks model. It’s not at the highest, but it’s not the lowest, it’s somewhere in the middle... fast, cost-effective, and you can broadly use everywhere.” — Andrew Brown
• [104:36]: “Now this is a very minimal—not in a bad way—but like it’s minimal in the sense that there’s not so much configuration nonsense or constant features being pushed out like Claude Code is...” — Andrew Brown

---

Assessment

A genuine, unedited instructional video course. Andrew Brown conducts live coding, genuine command-line interactions, environment debugging, and browser verifications in real-time without misleading cuts, pre-rendered mockups, or exaggerated claims.

Described by gemini-3.8-flash on 2026-10-13 from the video’s audio and frames.