Curated with by@mohithkumar808

№ 0958GitHub

Muse for Mac 'Endo' teardown

A static teardown of Meta's Muse macOS app (codename Endo) finding a full computer-use agent in the public build, switched off by five server-delivered feature flags.

barkleesanders/muse-endo-teardownREADME ↗

Muse "Endo" teardown — Meta's desktop agent has computer‑use built in, gated off

A static teardown of Meta's Muse desktop app for macOS (codename "Endo", bundle
com.meta.endo, v1.0). The finished implementation for a computer‑using agent — seeing the
screen, clicking, typing, running programs, driving a browser — ships in the public build but is
switched off by server‑delivered feature flags.

Found in code, not shipped. Everything here was read out of the released binary and its
Meta‑served config on a single Mac. Flag names, gates, and plans change; some features never
leave the lab. This is observation and commentary, not confirmation from Meta. No app binary is
redistributed here, and no credentials are published.

Read the illustrated report: report/endo-edition-2026-09-17.pdf


Subject

App Muse (Meta) for macOS, v1.0
Codename / bundle Endo / com.meta.endo
Binary Mach‑O arm64, ~17 MB, Swift + AppKit + SwiftUI
Signature Developer ID Application: Meta Platforms, Inc. (notarized)
Updater Sparkle

The five gated capability flags

Muse fetches its feature gates from a Meta config (endo-values-public…json, in the app's cache)
and re‑downloads them on every launch — editing them locally is overwritten within seconds.
Five booleans in the hatch_desktop namespace decide whether the agent can operate the machine,
and all five arrive false:

Flag Value Governs
hatch_desktop:enable_computer_control false Clicking / typing / controlling the Mac
hatch_desktop:endo_enable_screen_recording false The agent seeing the screen
hatch_desktop:enable_browser_automation false Driving a browser
hatch_desktop:endo_system_run_sandbox_enabled false Sandboxed program execution
hatch_desktop:endo_enable_computer_use_onboarding false First‑run "let Muse control your Mac" flow

What the binary already contains behind those flags

  • Screen vision — links Apple's ScreenCaptureKit (SCShareableContent, SCStreamConfiguration),
    calls CGPreflightScreenCaptureAccess / CGRequestScreenCaptureAccess, and carries an on‑screen
    computer‑use overlay + an accessibility‑first click mode (_showComputerUseOverlay,
    _preferAXForComputerUseClicks, `[ComputerUs
Muse for Mac 'Endo' teardown on github.com

Also filed under Benchmarks & research

See all →
  1. 0599

    Spark 1.3 tops Website Arena★

    BREAKING: Muse Spark 1.3 (xhigh) takes 1st overall on Website Arena with an Elo of 1362! This is a jump of 5 positions from Muse Spark 1.2, establishing a new Pareto frontier for Speed and Price. Only a month after the release of Muse Spark 1.2, @AIatMeta has topped this

    @DesignArena

    X post

    Benchmarks & research

  2. 0598

    Spark 1.3 moves the cost Pareto frontier★

    Last week the Intelligence Index vs Cost Pareto frontier moved out substantially. Claude Fable 5.1, Muse Spark 1.3, and GPT-6 Astra each set a new point in efficient intelligence Link to analysis: artificialanalysis.ai/#intelligence-…

    @ArtificialAnlys

    X post

    Benchmarks & research

  3. 0590

    Meta's OpenCode token share hits 45%★

    $META just went from 3.5% to 45.4% token share on OpenCode in just over two weeks Muse Spark 1.3 being good + free is enough to become the default for most users Default gets you usage → usage gets you data → data makes the next model better Anthropic and OpenAI can’t afford

    @thetomcollins

    X post

    Benchmarks & research

  4. 0541

    Spark 1.3 in Code Arena: WebDev top 10★

    Muse Spark 1.3 (xHigh) just landed @AIatMeta back in the top 10 models for Code Arena: WebDev! This release is ~#10 in Code Arena: WebDev with 1623 pts (AutoEval). It’s performance is par with Claude Fable 5 at #8 (1628 pts) and GPT-5.6 Sol at #11 (1616 pts). Muse Spark 1.3

    @arena

    X post

    Benchmarks & research

More GitHub

See all →
  1. 1091

    RedMuse

    Muse-native pentesting skill pack for Muse.ai. 13 skills, 4 scan modes (quick, standard, deep, diff), gate-first, localhost vulnerable app plus passive scans only, no active probes on live sites without a separate yes. Strix-inspired with original writing. Proven on localhost (IDOR, XSS, open redirect) and on live targets with anonymized totals: 0 critical, 2 high, 4 medium, 8 low, 14 info. MIT.

    Anonymous

    GitHub

    Coding & dev tools

  2. 1089

    Gig Driver Profit Checker — real delivery profit in minutes

    A working mobile-friendly web app built by the Meta Muse personal agent from a real courier's detailed requirements. It compares DoorDash/Uber Eats delivery offers after total miles (including empty return), waiting time, all-in vehicle costs and tolls. The surprising demo: a $9 offer can yield $19.08 net per hour while a $20 offer only yields $16.09 under stated assumptions. Muse created the standalone offline HTML and logic tests (49/49 self-reported); I downloaded it, independently verified actual browser inputs/outputs, published the source and proof screenshot to GitHub, and deployed a no-signup live demo. Live app: https://boyhoangxattpt2-blip.github.io/gig-driver-profit-checker/ . GitHub README includes exact math, source code, build story, and reproducible prompt.

    Anonymous

    GitHub

    Apps & websites

  3. 1087

    Zouroboros Workshop

    Muse is your personal assistant. Zouroboros Workshop is the workshop behind it. Workshop is an MCP service that runs on a private, single-tenant VPS provisioned for each customer. From a normal Muse conversation, a user can dispatch real software work to a coordinated team of seven AI coding executors — Claude Code, Codex, Gemini, OpenCode, Kimi, Hermes, and Pi — routed by health and capability, with automatic fallback when one executor is unavailable. Three things make it different from a single-vendor hosted builder. First, it runs on the subscriptions the user already pays for: there are no Workshop model credits and no second meter. Second, its memory is shared across every executor and survives across sessions, so a decision recorded by one tool is retrieved by another in a later Muse conversation — memory that compounds instead of resetting. Third, work is verified before it is delivered: builds run through a factory lane with deterministic review, and swarm results return with per-task evidence, not just an agent's claim that it is done. The customer-facing promise is simple: turn your existing AI tools into a workshop that delivers tested, reviewed work — a finished pull request, not a wall of code in chat.

    @Xmarlandoj

    GitHub

    Connectors & MCP

  4. 1086

    Self-Healing Agent VM

    A self-healing watchdog pattern for AI agent VMs: byte-exact restore caches in persistent storage, sentinel probes, no-clobber atomic restores, and verified heals. When the VM's root filesystem is rebuilt with pieces missing (Python stdlib, timezone data, Node), the watchdog restores them in minutes and reports only repairs and failures. Includes an AGENTS.md so any agent can set it up.

    @TMTrevisan

    GitHub

    Agents & automation

Curator picks

  1. 1057

    Muse Mail: your agent gets an email address★

    MUSE FEATURE ALERT: y'all have been asking for this one. your @muse is getting its own email address. add it to a thread or forward it stuff and it’ll handle it for you. muse mail!

    @alexandr_wang

    X post

    Agents & automation

  2. 1054

    Kevin Roose gives his Muse $100 and a Kalshi account★

    gave my Muse $100 and a kalshi account, let's see how frontier this thing is

    @kevinroose

    X post

    Agents & automation

  3. 1046

    Medical bills audited line by line, $4,000 saved★

    Got Muse logged in to my medical provider’s portal, he pulled the itemized bills, and questioned every line. So far he’s found several times I’d been double billed, asked for some discounts and has saved me over $4,000. If your moat is bureaucracy, you’re cooked.

    @Ryan_Holdaway

    X post

    Errands & personal agent

  4. 1042

    GitHub connector for Muse★

    GitHub 🤝 @Muse A new way to connect your GitHub account to review PRs, catch up on issues and notifications, and leave comments without switching tabs.

    @github

    X post

    Connectors & MCP

Want more AI builds?artificialbuilds.com