Meta Says Muse Was Built From Scratch, Inspired by OpenClaw

Meta’s Muse has an acknowledged predecessor. Nat Friedman, head of product at Meta Superintelligence Labs, said the personal AI agent was “definitely heavily inspired as a product by OpenClaw,” according to TechCrunch’s September 22 report. In the same statement, he said Meta built Muse “from scratch.”

A team can adopt another product’s ideas without using its code. But Friedman’s comments followed social posts alleging similarities between files in Muse’s agent workspace and OpenClaw’s. His acknowledgment explains the product resemblance; it does not settle what happened with those files.

Meta introduced Muse on September 8 as an agent that can carry out tasks across apps, continue working after a user closes the app, and ask for approval before sensitive actions. Friedman’s statement leaves a narrower question than whether the products feel alike: where does the admitted design influence end, and what evidence would establish reuse of implementation?

Friedman Credits OpenClaw’s Product Design

Friedman said he used OpenClaw in January, bought hundreds of Mac minis for his team afterward, and found that many colleagues became enthusiastic about OpenClaw and other personal agents. He credited OpenClaw creator Peter Steinberger’s agent harness, the software framework around an AI model, as pioneering.

Meta’s aim, Friedman said, was to build something like OpenClaw that could be made safe, secure, easy to use, and suitable for a much larger audience. That is a statement of intent, not evidence that Muse has reached billions of users or that its protections have been independently validated. TechCrunch reported that a Meta spokesperson pointed back to Friedman’s statement and offered no further comment.

Similar Workspace Files Prompted the Questions

The discussion began with a post from AI app co-founder Ansh Nanda comparing Muse with OpenClaw. As TechCrunch describes the exchange, Nanda shared a conversation in which Muse characterized similarities between its system files and OpenClaw’s as a match. Other people posted comparisons of file names, while one alleged that the contents of a SOUL.md file were nearly identical.

SOUL.md is a Markdown file used to describe aspects of an agent’s identity and behavior, such as its tone and boundaries. Similar file names can indicate a shared design convention. Similar contents, if authenticated, would raise a more specific question about whether text was reused.

Asked about the reportedly matching file names and SOUL.md content, Friedman replied, “we thought that Peter got those things exactly right,” according to TechCrunch. His answer reinforces the acknowledged influence but does not specify whether Meta copied a file, adapted its wording, or independently wrote text that follows the same design.

Nanda’s exchange also needs to be read with care. An AI agent’s description of its own files is not a code audit: its answer may reflect material available in its working context, and it can be mistaken about provenance. The reported comparisons give Meta a specific question to answer. They do not prove that Muse runs OpenClaw under the hood.

Product Inspiration and Code Provenance Are Different Claims

A product can borrow an idea about how a personal agent should behave, adopt workspace conventions such as file names or instructions, reuse text, or incorporate another project’s software code. Evidence for one does not automatically establish the others.

Calling Muse “built from scratch” is not independent verification of every component’s origin. If the alleged near-identical file text is confirmed, Meta would have a reasonable question to answer about what it reused and how it understands that phrase. Matching configuration text, however, would not establish that Muse’s core agent software is an OpenClaw fork.

The public account cited here does not provide a documented comparison of the products’ source code or an authenticated, complete account of the disputed files. It also does not establish a licensing violation. Such a judgment would require knowing what material was reused, its provenance, and the terms that apply to it. Open-source software can be studied and used, but the existence of an open-source project is no substitute for checking those particulars.

Friedman has confirmed substantial product influence. Whether the reported file similarities reflect reuse of implementation remains open.

Muse’s Consumer Package Is Central to Meta’s Bet

Meta’s Muse launch announcement presents the agent as something people can message in its app or through WhatsApp. Meta says Muse can use a browser, work across connected services, remember relevant details, and keep working on longer tasks after a conversation ends. The company identifies Muse Spark as the model powering it.

Meta is trying to turn an agent entrusted with ongoing tasks into a consumer product. According to the company, Muse can help with actions such as sending email or booking travel, while returning to the user for approval when required. It can also assist with purchases through Stripe’s Link, which Meta says uses a one-time-use card so the agent does not see a user’s real card details.

That ambition helps explain Friedman’s emphasis on ease of use and scale, but it says nothing about whether a particular workspace file was copied. A new interface, model, or service architecture could coexist with reused instructions. Similar instructions would also say little on their own about the rest of the system. The product and its provenance require different evidence.

Meta’s Security Claims Need Their Own Scrutiny

Meta says Muse operates inside a dedicated cloud virtual machine called Muse Secure VM, where it keeps the agent and data associated with the user. The company also describes a separate Sentinel agent that reviews what Muse sends to the internet. According to Meta, users can choose which services to connect, set some access permissions, and inspect an audit trail of Muse’s actions.

These are concrete design claims, not a guarantee that every agent action is safe. A personal agent’s usefulness depends on access to consequential accounts and tasks; an error involving email, a form, or a purchase has different stakes from an incorrect chat answer. Friedman’s promise to make an OpenClaw-like product secure for a broad audience should be judged against how those controls work in practice.

The launch post distinguishes protections Meta says Muse has now from one it plans to add. A Muse Confidential VM, which Meta says would encrypt the virtual machine with a key held only by the user, was described as coming later in 2026. It should not be counted as a shipped feature. Meta says users can opt out of having their Muse interactions used to train its AI models and that conversations and data in the virtual machine are not shared with its ad systems; those are company statements, not independent findings in the cited reporting.

Early Popularity Does Not Resolve the Dispute

Muse was rolling out in the United States on iOS, Android, and muse.ai when Meta announced it. Meta described it as free for most needs, with subscription plans for heavier use, without specifying prices in that announcement. TechCrunch reported on September 22 that the app had recently reached No. 1 on the U.S. App Store.

That ranking suggests strong initial interest. It does not measure how many people keep using Muse, how reliably it completes tasks, or whether its safeguards work as described. It also cannot answer how Muse was built.

Final Thoughts

Friedman’s acknowledgment makes OpenClaw’s influence on Muse clear. His “built from scratch” statement draws a boundary at implementation, but the reported similarities in agent files leave that boundary insufficiently explained.

The evidence cited here establishes neither that Muse is secretly OpenClaw nor that the file questions have been answered. Meta could clarify the distinction by addressing the provenance of the specific files people identified. For a company asking users to let an agent handle messages, accounts, and purchases, precision about what it built deserves as much attention as the product’s polish.