ABAB News Logo
Back to Flash News

Meta Admits Muse Workspace is Inspired by OpenClaw

Nat Friedman, head of product at Meta's Superintelligence Lab, stated that the personal agent Muse was built from scratch but "heavily borrowed" from the open-source project OpenClaw, with some workspace file names and content being similar.

Friedman mentioned that after using OpenClaw in January this year, he procured hundreds of Mac minis for the lab, as many team members fell in love with this type of personal agent. The goal is to create a version that is "like OpenClaw, but safe, easy to use, and scalable to billions of people." He did not deny the similarities when faced with questions about the similar naming of the codebase.

OpenClaw was initiated by Peter Steinberger, with default workspace files including AGENTS.md, SOUL.md, USER.md, IDENTITY.md, TOOLS.md, MEMORY.md, and daily segmented memory logs, defining personas, rules, and memories using the local file system. In comparative evaluations, Muse's side agent reported running the same set of SOUL.md, IDENTITY.md, USER.md; its environment also includes a persistent Linux virtual machine, bash, Chromium sessions, sub-agents, reusable skills, and scheduled tasks.

Muse launched on September 8 in the U.S. on iOS, Android, and web, capable of integrating with WhatsApp, claiming to be "built from scratch for everyone," with the underlying technology being Muse Spark, running on a Muse Secure VM for each user, requiring user authorization for sensitive operations. It later launched on Mac, allowing actions within local files, information, calendars, notes, and emails. The free tier covers regular usage, with subscriptions around $20 and $100 per month. Over 2.5 million downloads in the U.S. within about two weeks of launch.

Microsoft previously announced at Build that OpenClaw could run natively in Windows execution containers and launched Scout inspired by it. OpenClaw is self-hosted under the MIT open-source license and can connect to multiple channels like Telegram, WhatsApp, and Discord; Muse integrates the same form into Meta's cloud and account system, removing the need for self-built gateways and self-supplied model keys.

In market mechanisms, this represents the open-source prototype being productized by a platform: developers and geeks validate agents with "a computer, a browser, and markdown personas" on OpenClaw, while Meta uses computing power, distribution, and compliance to package the same interaction in exchange for super app entry. The buyers are general users who do not want to build their own servers; the sellers are Meta, covering inference costs with free and subscription models, while exploring e-commerce commissions. The beneficiaries are platforms that already dominate WhatsApp and app stores; the pressured parties are the differentiated narratives of open-source projects and e-commerce platforms like Amazon that refuse Muse access. The event-driven narrative is clear: first, there was a popular open-source form, then major companies acknowledged borrowing and scaled it up.

Workspace files transitioned from auditable local text to configurations in cloud VMs; control shifted from user disks to platform sandboxes and authorization gates.

Source: Public Information

ABAB AI Insight

In the past year, Meta has shifted its narrative from open-sourcing Llama to a closed-source blitz with the Superintelligence Lab: After Alexandr Wang joined, Muse Spark was developed from scratch in nine months, first integrated into its own applications, and then made into executable tasks with Muse. This follows the same path as when it absorbed social innovations through Instagram and WhatsApp—not inventing categories from scratch, but turning validated usage methods into default options for billions. Friedman purchasing Mac minis to practice with OpenClaw means they first acted as open-source users internally before transferring file agreements and agent shapes into the Secure VM.

The capital path is transforming the open-source workspace protocol into cloud-based rented computers: each user has an isolated virtual machine, independent browser, credential vault, and disposable card number for checkout, with inference costs covered by free quotas plus $20/$100 tiers, focusing on transaction commissions rather than advertising. The motivation is to lock "doing things for you" into Meta accounts and WhatsApp threads, rather than letting users keep memory files in ~/.openclaw. Resource mobilization relies on hardware procurement to establish internal habits, using Spark as the model, and exchanging app store and social distribution for installations, rather than acquiring the OpenClaw entity.

Analogies can be drawn to Microsoft integrating Linux into Azure, incorporating OpenClaw into Windows containers to create Scout, and Google making Android a distributable Linux. The industry phase has shifted from geek self-hosting expansion to platform control: whoever owns the default workspace, default memory files, and default authorization pop-ups defines the personal agent. OpenClaw uses markdown to make personas git-compatible; Muse uses the same file names to lower the mental barrier for migration, shifting storage to its own VM.

Structural judgments belong to the reconstruction of the industrial chain. The mechanism is: open-source first establishes the file protocol for the agent operating system, then major companies use sandboxes, subscriptions, and super app distribution to abstract the protocol from local directories into cloud services. Similar file names are not details; they indicate that the interface layer has been copied, with differences reduced to hosting and compliance; thereafter, competition shifts from "can we create agents" to "who owns the memory and permissions on their disk."

ABAB News · Cognitive Laws

  1. Open-source first defines files, platforms then define entry points

  2. Borrowing file names is to take over usage habits

  3. Secure sandboxes sell hosting rights, not creativity