Meta Muse Wonder: Build Apps and Agents with Muse Code

Build a real Python tool with Muse Code, call Muse Wonder from the API, and measure effort, caching and cost

Created byJosé Calderón

Last updated 10/2026

English

What you'll learn

Choose between Muse Wonder, Muse Wonder Full and Muse Spark, and know where each one is available and what it costs
Install and configure Muse Code safely: model, Contributor mode, imported rules, permissions and the sandbox
Plan, build and test a real Python tool with Muse Code, its sub-agents and an AGENTS rules file
Make a small Muse Code plan last
Call Muse Wonder from Python with the OpenAI SDK and read what each call really cost
Tune reasoning effort and prompt caching, and compare models with an automatic quality test
Use structured output and screenshots with Muse Wonder, and know what each one guarantees

Course content

4 sections • 15 lectures • 32m total length

Welcome: what Muse Wonder is and what we'll build1:41
What Meta released on October 12, 2026, the project you will build (ReleaseNotes Bot, which turns git commits into release notes for real users), the three parts of the course and what you need: basic Python and a budget of about five to ten dollars.

Muse Wonder, Wonder Full and Muse Spark2:30
Where each model is available (Model API or Muse Code, and which plans), the API prices including Contributor mode, and what Meta's own benchmark table says, including the rows where Claude Opus 5.5 is ahead.
Install Muse Code and pick your plan2:09
The one-line install on Windows, macOS and Linux, signing in with a browser code, the four plans, and three settings to check before your first task: the default model (Spark 1.2, not Wonder), Contributor mode, and the import of your Claude Code rules.
Your first session in Muse Code2:24
A read-only first task on the PocketBudget repository: interactive and headless mode, what Muse Wonder ran in 15 seconds, how the approval policy and the sandbox handled each command, and the four permission profiles.
Quiz 1 · Muse Wonder fundamentals

Plan the project with Muse Wonder1:49
Ask for a plan before any code: the prompt, the seven-file plan Muse Wonder proposed in 17 seconds, how it planned to test without calling the API, and the five questions it asked before starting.
Sub-agents and parallel work2:22
Answer the questions and say go. What really happened: the $5 plan ran out mid-build, an API key could not continue the same session, and a new session finished the work with a workflow of three parallel agents. 67 tests in under a second, verified with a real run.
Tests, debugging and recovery2:10
A new requirement, breaking changes, done test-first: the new tests fail (8 failed), the code changes, the whole suite passes (85 passed). Then a real run that shows what tests cannot catch.
Making your usage last1:58
The real numbers from the usage page after one afternoon (100% of the 5-hour window, 35% of the weekly limit) and six habits that make a small plan last, plus the trade-offs of Contributor mode, API keys and bigger plans.
Project rules with AGENTS.md2:19
What muse init generates, the four rules we wrote, and a test: a prompt that mentions none of them, and a session that followed all four. Plus where Muse Code looks for rules and memory, and why untrusted repositories need care.
Quiz 2 · Building with Muse Code

First API call with the OpenAI SDK2:14
Point the OpenAI SDK at the Meta Model API, get a key with a spending alert, and read the usage of your first call: most of the output tokens were reasoning, billed at the output price.
Reasoning effort, caching and cost2:36
The same 14 commits at minimal, low, medium and high effort (minimal wrote the same notes as high, four times cheaper), prompt caching measured with and without prompt_cache_key, and why output is most of the bill.
Wonder or Spark? Measure it2:16
A ten-line automatic quality check and three runs per model: Muse Wonder wrote complete notes 3 of 3 times and was the fastest; Muse Spark was cheaper but complete only 1 of 3 times.
Structured output: valid JSON, wrong notes2:08
Release notes as JSON with a schema. The JSON was always valid, but 0 of 9 runs were complete and correct. What a schema guarantees, what it does not, and why the bot keeps free text.
Muse Wonder reads a screenshot1:41
Send a user's error screenshot with the OpenAI image format and get a bug title, the likely cause and a release-note line in 3.7 seconds, for under one cent. It found the impossible date.
ReleaseNotes Bot end to end, and what's next1:58
The whole bot in one command, what the course cost me, and what to improve next: prompt caching, a quality check on every release, human review and re-measuring when the preview model changes.
Quiz 3 · Muse Wonder through the API

Requirements

Basic Python: you can read a short script and run it from a terminal
A Muse Code plan (the five-dollar Everyday plan is enough) and a Meta Model API key with a payment method; the API examples cost a few cents
Windows, macOS or Linux

Description

This course contains the use of artificial intelligence.

On October 12, 2026 Meta released Muse Wonder, a model built for coding, long-running agents and computer use. You can use it in two places: inside Muse Code, Meta's coding agent for the terminal, and through the Meta Model API, which works with the OpenAI SDK.

In this course you build ReleaseNotes Bot, a small Python tool that reads the git commits of a project since its last version and turns them into release notes that normal users can read. You plan it and build it with Muse Code, and then you call Muse Wonder from your own code and measure what every choice costs.

Everything was recorded with Muse Code 1.4.7 and the October 2026 release, on the five-dollar Everyday plan, with real sessions and real API calls. Some of what you will see:

What Muse Wonder, Muse Wonder Full and Muse Spark are, where each one is available, and what Meta's own benchmarks really say.
The settings to check before your first task: the default model is not Wonder, Contributor mode is on, and Muse Code imports your Claude Code rules.
How Muse Code plans before it codes, asks questions, splits work between sub-agents, and runs every command through a policy and a sandbox.
How fast the cheapest plan runs out, and how to make it last.
Project rules in an AGENTS file that Muse Code followed without being reminded.
Reasoning effort measured: minimal effort wrote the same notes as high, four times cheaper.
Prompt caching measured, and why output, not input, is most of the bill.
Muse Wonder against Muse Spark with a small automatic test: complete notes three times out of three, against one out of three.
Structured output that always returned valid JSON, and still got the content wrong.
A user's screenshot turned into a bug report and a release-note line for under one cent.

How this course is made: I designed the project, ran every session and every API call myself, and reviewed every lesson. The narration uses an AI voice and the animations are generated from my scripts. The complete project is included as a download.

Who this course is for:

Developers who saw the Muse Wonder launch and want to know if it is useful for their work
Users of other coding agents who want to try Muse Code with a clear head about cost and privacy
Python developers who already use the OpenAI SDK and want to call Muse Wonder

Meta Muse Wonder: Build Apps and Agents with Muse Code

What you'll learn

Explore related topics

Course content

Welcome1 lecture • 2min

Muse Wonder fundamentals3 lectures • 7min

Building with Muse Code5 lectures • 11min

Muse Wonder through the API6 lectures • 13min

Requirements

Description

Who this course is for: