AI agents · OpenClaw · self-hosting · automation

Quick Answer

Meta Muse vs OpenClaw: Meta Admits Muse Was Inspired

Published:

The short answer

Muse is OpenClaw for people who will never open a terminal, and Meta now says so. After users spotted identical workspace filenames and a near-verbatim SOUL.md, Nat Friedman confirmed on September 22, 2026 that Muse was “heavily inspired as a product by OpenClaw” while being built from scratch. Muse launched September 8 in the US, drew more than 500,000 users and 250,000 daily actives in its first week, hit number one on the App Store, and gained checkout rights across Shopify’s Shop Pay on September 21. It also shipped with a zero-day that let any local Mac app steal the agent’s authentication token. OpenClaw remains the open-source original: self-hosted, provider-agnostic, and yours to configure.

Side by side

Meta MuseOpenClaw
What it isHosted personal AI agent from Meta Superintelligence LabsOpen-source, self-hosted personal agent runtime
LaunchedSeptember 8, 2026 (US)Open source; OpenClaw 2.0 shipped earlier in 2026
ModelsMeta’s Muse Spark family (Muse Spark 1.3 as of September 2, 2026)Any provider: Anthropic, OpenAI, Google, xAI, open-weight via local or hosted endpoints
Where it runsMeta’s cloud; iOS/Android app, WhatsApp, Mac app, webYour machine or server; Discord, Telegram, WhatsApp, web UI, CLI
PricingFree (~100M tokens/week, card required); Power $20/mo (500M); Maximum $100/mo (3B); $16/$80 on webFree; pay your model provider per token
ConfigurationWorkspace files including SOUL.md, mirrored from OpenClaw conventionsSOUL.md, AGENTS.md, USER.md, skills, tool policies, cron-style automations
CommerceShop Pay checkout across Shopify stores (September 21, 2026)Whatever you wire up; no built-in payments authority
Security modelMeta-managed; zero-day patched September 22 after Objective-See disclosureYou own the sandbox, approvals and tool policy; you also own the risk
AvailabilityUS only at launchAnywhere you can run it
Early scale500K+ users, 250K+ DAU, 2M+ prompts in week oneDeveloper and power-user community; no public user count

What Meta actually said

The trigger was a viral post from Ansh Nanda calling Muse “LITERALLY OpenClaw for normies,” backed by side-by-side screenshots of the two projects’ workspace files. Friedman’s response on X did three things: it conceded the inspiration, denied code reuse (“built from scratch”), and reframed the copying as a compliment. He said he had used OpenClaw personally in January 2026, bought hundreds of Mac minis for the MSL team to run it, and set out to build something similar that could be “safe, secure, easy to use, and scale to billions of people.”

TechCrunch’s framing was that this fits Meta’s familiar playbook, with Snapchat Stories as the precedent. The difference here is licensing: OpenClaw is open source, so borrowing its conventions is permitted, and identical SOUL.md structure is not a legal problem. It is a branding one for a product Meta charges up to $100 a month for.

The zero-day, and why it matters more than the filenames

On September 21, 2026, Patrick Wardle of the Objective-See Foundation disclosed that any locally installed app or terminal command, regardless of what macOS permissions it held, could rewrite a long list of undocumented Muse settings. One of those redirected Muse’s transcription endpoint; a malicious process could point it at itself and capture the authentication token used to control the agent. Wardle’s proof of concept then queried device inventory, pulled a linked iPhone’s location in Barcelona and started a Bluetooth Low Energy scan, all invisibly.

Meta removed the setting from production builds within a day. David Singleton of Meta Superintelligence Labs described it as a local privilege escalation, not a remote exploit, which is accurate and also the point: Muse is designed to hold an extraordinary amount of authority on your behalf, and as of September 21 it can spend money through Shop Pay. A config bug in a chat app is a nuisance. A config bug in an agent with purchasing power and device access is a different class of problem. VentureBeat’s post-patch assessment was that enterprise security teams still cannot see what Muse can reach.

OpenClaw does not make that problem disappear; it moves it to you. A self-hosted agent with shell and file access is exactly as dangerous as the tool policy and sandbox you give it. The difference is that you can read the policy.

Who should pick which

Pick Muse if: you want an agent that books trips, tracks expenses and buys things with zero setup, you are in the US, and you are comfortable with Meta holding the session, the model and the data. The free tier is generous for casual use; most people, Meta says, will stay on it.

Pick OpenClaw if: you want to choose the model (and pay $0.10/$0.50 per million tokens on GPT-6 Luna for clerical work, or $4/$20 on Claude Opus 5.5 for hard tasks, rather than a flat $20–$100), you need it to run on your own infrastructure, you want it in Discord or Telegram or on a server, or you want to audit exactly what it is allowed to do. Expect to spend an afternoon on setup and to own the security posture yourself.

Neither, yet, if: your threshold is “no unpatched way for local software to hijack the agent.” Muse cleared that bar after September 22; OpenClaw clears it only if you configure it to. The agent ecosystem is converging on the same file layout, the same SOUL.md, and the same MCP plumbing. What is not converging is who is accountable when the agent does something you did not ask for.

For the broader personal-agent field, see Meta Muse vs ChatGPT Agent Mode vs Gemini Spark vs Claude and Why Amazon blocked Meta’s Muse agent.

Last verified: September 24, 2026.

Sources