# Inception: A Muse Inside Muse 

*This post was written using* ***Muse Code powered by Meta Muse Spark*** *to discuss the Muse Coding Agent extension I built to use Muse Codes from the start — an idea planted inside itself.*

> **Recursion is how it works. Inception is how it felt.**

> Source for Muse Spark details: [Introducing Muse Spark: Scaling Towards Personal Superintelligence — Meta AI, April 8, 2026](https://ai.meta.com/blog/introducing-muse-spark-msl/) — the launch post for the first Muse-family model from **Meta Superintelligence Labs (MSL)**.

![](https://cdn.hashnode.com/uploads/covers/66ca36f649c4df0ee8d13fa5/0c7640a6-301c-4a0f-a575-c62221afbba0.png align="center")

Formerly code-named “Avocado,” Muse Spark is now live at [meta.ai](http://meta.ai) and the Meta AI app (private API preview) — rolling out to WhatsApp, Instagram, and smart glasses.

## The Setup: An Idea Planted Inside Itself

I'm all about inception — an idea nested inside the very thing it describes.

So here's the inception: I used **Muse Code** — Meta's terminal coding agent — to write this post about the **Muse Coding Agent extension** I created so I could use Muse Code inside my editor from day one. The tool became the author of its own origin story. The dream built the dream.

If recursion is a function calling itself, inception is an idea installing itself. This post does both. Technically, it recurses — Muse bootstrapping Muse. Experientially, it *incepts* — planting the idea of Muse-in-the-editor by using Muse to make it real.

Muse writing about Muse, bootstrapped by Muse. That's not just recursion. That's inception.

## What Is Muse, Actually?

The full name matters:

> **Muse Code powered by Meta Muse Spark** — from **Meta Superintelligence Labs (MSL)**

### Muse Spark `muse-spark-1.2`

The first in the **Muse family** from MSL. As Meta describes in [the launch post](https://ai.meta.com/blog/introducing-muse-spark-msl/), a **natively multimodal reasoning model** with **tool-use, visual chain of thought, and multi-agent orchestration**. Powers Meta AI; served via [`https://api.meta.ai/v1`](https://api.meta.ai/v1) (OpenAI-/Anthropic-compatible). Formerly “Avocado.” Muse Spark 1.1: July 9, 2026.

### Muse Code `muse` / `muse exec`

The official agent harness on top of that model — plan/edit/run, approvals, sandbox, subagents, streaming JSON event log. Think Claude Code CLI / Codex CLI, but Meta.

Meta's own framing: **Muse Spark is the first step on a scaling ladder toward personal superintelligence** — the first product of a ground-up overhaul, from research and training to infrastructure, including the **Hyperion data center**. Available today at [meta.ai](http://meta.ai) + Meta AI app (private API preview). With larger Muse-family models already in development, Meta is explicit that this is the *start*.

Meta ships Muse as the ready-made agent. Or you call the Model API directly. Two surfaces — and the distinction drove the extension design.

## The Gap

Gemini has Code Assist. Anthropic has Claude Code for VS Code. OpenAI has Codex in the IDE.

Meta had just shipped a strong terminal agent — and **no official VS Code or Cursor extension**.

I wanted to use **Muse from Meta** as my coding agent *inside* Cursor, exactly like those other agents work: sidebar chat, streaming diffs, session resume, approvals — right next to my files.

The community had one partial answer — [Muse Spark for Copilot Chat](https://marketplace.visualstudio.com/items?itemName=LukeSpine.meta-spark-for-copilot) — but that plugs the *model* into Copilot. It doesn't give you the Muse Code harness. I didn't want just the model. I wanted the agent.

Waiting felt wrong. So I willed it into existence.

## The Question I Asked Cursor

> Can we create a Visual Studio or Cursor extension that works exactly like Gemini Code Assist, Claude Code for VS Code, or the Codex OpenAI Coding Agent? Any one of those would be reasonable, but I want to use Muse from Facebook as the coding agent.

Cursor laid out the feasibility — and the product gap was clear:

| Piece | Status |
| --- | --- |
| **Muse Code** (`muse` / `muse exec`) | Official beta agent: plan/edit/run, approvals, sandbox, subagents, event log |
| **Muse Spark** (Model API) | OpenAI- and Anthropic-compatible |
| **Official IDE extension** | None — this is the gap |
| **Community** | Model-only integrations exist, not the full harness |

Three realistic builds:

1.  Thin shell around Muse Code — best “exact Muse” match Sidebar + webview driving muse exec --json, streaming events, diffs, approval prompts. You keep subagents, worktrees, skills, hooks, event log. Closest to Claude Code for VS Code.
    
2.  Model-only provider — fastest, not full Muse Register Muse Spark as a chat model (Copilot provider, Continue, Cline, Cursor custom endpoint). Good chat, no harness.
    
3.  Custom harness on Model API — most work Rebuild tools, edits, shell, approvals and call muse-spark-1.2 directly. Full control, you reimplement what Muse Code already does.
    

## What Muse Spark Can Actually Do

Per [Meta's launch post](https://ai.meta.com/blog/introducing-muse-spark-msl/), Muse Spark was built for **personal superintelligence** — an AI that understands your world, not just your text. Three capability clusters stood out, and all three are why the IDE wrapper matters:

![](https://cdn.hashnode.com/uploads/covers/66ca36f649c4df0ee8d13fa5/0774121c-9d94-438f-95ee-e9f2fed80667.png align="center")

Muse Spark capabilities: multimodal, visual CoT, tool-use, health (1,000+ physicians), Contemplating mode (58% HLE / 38% FrontierScience), and interactive use cases.

1.  **Natively multimodal + visual chain of thought.** Built to integrate visual information across domains and tools. Strong on visual STEM, entity recognition, localization. Meta's examples: creating minigames or troubleshooting home appliances with dynamic annotations — not just describing an image, but reasoning over it and acting.
    
2.  **Health, with unusual rigor.** Meta collaborated with 1,000+ physicians to curate health training data for more factual, comprehensive responses, plus interactive displays for nutrition or muscles activated in exercise.
    
3.  **Agentic + “Contemplating” mode**. Competitive on multimodal perception, reasoning, health, and agentic tasks (Meta is candid about gaps: long-horizon agentic + coding workflows). The headliner is Contemplating mode — orchestrating multiple agents that reason in parallel — positioned against Gemini Deep Think and GPT Pro. Results: 58% on Humanity's Last Exam and 38% on FrontierScience Research (rolling out gradually in meta.ai).
    

## How Meta Is Scaling It

Meta says the stack is scaling predictably along three axes — this is the “ladder” they keep referencing:

![](https://cdn.hashnode.com/uploads/covers/66ca36f649c4df0ee8d13fa5/e33c4056-7415-4a44-b4d7-9e671b38888a.png align="center")

Three scaling axes: pre-training, reinforcement learning, and test-time reasoning.

*   **Pre-training** (foundation): Nine-month rebuild of architecture, optimization, data curation. Same capability with **\>10× less compute** (scaling laws on small models) — the efficiency that makes larger Muse models practical.
    
*   **Reinforcement learning** (amplification): Smooth, predictable **log-linear growth in pass@1 and pass@16** and generalization to held-out evals — reliability without sacrificing reasoning diversity.
    
*   **Test-time reasoning** (thinking before answering): Two levers — **thinking time penalties** (including observed *thought compression* on AIME) and **multi-agent orchestration** that boosts performance without ballooning latency. Parallel agents > single-agent longer thinking at same latency.
    

Safety in the launch post: evaluated under Meta's Advanced AI Scaling Framework — strong refusal on bio/chem weapons, within safe margins on cybersecurity/loss-of-control, third-party checks from Apollo Research (noted unusually high *evaluation awareness*).

Why this matters for the extension: if the trajectory is *multimodal + tool-use + multi-agent* toward personal superintelligence, you want that agent *next to your files*, not trapped in a terminal. The IDE is where visual context, tool-use, and parallel agents actually live.

## The Choice

| Decision | Answer |
| --- | --- |
| Cursor/VS Code (not classic Visual Studio) | Yes — good assumption |
| Fidelity | **Option 1** — full Muse Code harness |
| Scope | Start personal, publish if it's great |

Option 1 was the only one that satisfied “works like Claude Code, but Muse is the agent.” Don't reimplement the harness. Wrap it.

## Why This Shape Wins

*   **Muse stays Muse.** Updates to the CLI, skills, hooks, and subagents flow through. When Meta ships the next Muse-family model on that scaling ladder (Muse Spark is “step one”), the extension rides the upgrade for free.
    
*   **Safety stays intact.** Sandbox and approval semantics come from Muse Code — consistent with MSL's guardrails (data filtering, post-training, system checks).
    
*   **It feels native.** A VS Code extension (Cursor runs those) that speaks the same JSON stream the terminal does, just rendered where you already live. Tool-use, visual CoT, and multi-agent orchestration all have a natural home: diffs, terminals, image context, parallel subagents.
    

If you've been running `muse` in a terminal and wishing it lived next to your files the way Claude Code and Codex do — this is that missing piece. And if Meta's bet pays off — Muse Spark powering Meta AI from [meta.ai](http://meta.ai) to WhatsApp, Instagram, and smart glasses — you'll already have that same family sitting beside your code.

![](https://cdn.hashnode.com/uploads/covers/66ca36f649c4df0ee8d13fa5/6e63b232-2dcc-45d4-922f-e3c7f3e11b92.png align="center")

Thin shell around the real harness — Muse Code sidebar streaming next to your workspace.

## Closing the Loop: The Dream Built the Dream

This post, about a Muse extension, written with Muse, to use more Muse — inception in three layers: the model (Muse Spark), the harness (Muse Code), and the editor where they now live together.

Recursion is how it works. Inception is how it felt. And I like inception way more because it's the better discussion — not just “a function calling itself,” but an idea so compelling it plants itself.

Inception isn't just an idea I like. It's how this one got built.

**Agent:** Muse Code powered by Meta Muse Spark — see [Introducing Muse Spark](https://ai.meta.com/blog/introducing-muse-spark-msl/) (April 8, 2026)  
**Extension:** Muse CLI Chat (unofficial) — `tools/extensions/muse-code`  
**Repo:** [github.com/codemarc/muse-code](https://github.com/codemarc/muse-code)  
**Status:** Personal → publish-ready  
**Further reading:** [Muse Spark on Wikipedia](https://en.wikipedia.org/wiki/Muse_Spark) · [Muse Spark 1.1 preview](https://www.reuters.com/business/meta-debuts-muse-spark-11-with-preview-open-developers-2026-07-09/) · [meta.ai](http://meta.ai)

*Sometimes the product you want is the one you ship yourself. And sometimes you ship it with itself.*

## Share it

![](https://cdn.hashnode.com/uploads/covers/66ca36f649c4df0ee8d13fa5/795c2665-94cb-4b43-ba21-44a35f16dcf5.png align="center")

Meta shipped Muse Code as a terminal agent — part of the new Muse family from Meta Superintelligence Labs. Claude and Codex already had IDE surfaces. Muse did not.

That gap bothered me enough that I stopped waiting and built one.

Over a couple of days I shipped **Muse CLI Chat**: an unofficial VS Code / Cursor sidebar that drives the real `muse` CLI (`muse exec --json`). Same Muse Spark model family behind [meta.ai](http://meta.ai), same harness, sandbox, and sessions — just the UI I wanted in the editor.

Muse Spark is Meta's first step toward personal superintelligence: natively multimodal, visual chain of thought, tool-use, and a Contemplating mode with parallel agents (58% Humanity's Last Exam). Read the launch post: [https://ai.meta.com/blog/introducing-muse-spark-msl/](https://ai.meta.com/blog/introducing-muse-spark-msl/)

This post itself was written *with* Muse Code, about the Muse extension, to use more Muse. Recursion is how it works — inception is how it felt.

It's not affiliated with Meta. Install Muse, sign in, open the sidebar, go.

GitHub: [https://github.com/codemarc/muse-code](https://github.com/codemarc/muse-code)

#Muse #MuseSpark #MetaAI #VSCode #Cursor #CodingAgent #PersonalSuperintelligence