Meta Muse Turns the AI Assistant Into a Computer, and Open Source Is Following

Meta Muse Is Really an AI Computer

Meta launched Muse on September 8 as a personal AI agent that can operate a browser, work across connected services, fill forms, send emails, plan trips and continue working when the user leaves the application. Meta says Muse can also manage longer-term goals rather than waiting for a new prompt every time something needs to happen.

That makes Muse more interesting than another improvement to conversational AI. The underlying idea is to give the model a persistent environment in which it can actually work.

Muse runs inside what Meta calls Muse Secure VM, a dedicated virtual machine with its own browser and user-specific data. It can maintain task state, remember relevant information, and operate across services over time. Sensitive actions, such as sending an email or completing a purchase, may require user approval.

The distinction matters. A chatbot can tell someone how to organize a trip. An agent with a browser, account access, and persistent memory can potentially organize parts of that trip itself.

Meta is increasingly extending that model beyond phones and computers. At Connect 2026, the company said Muse will also come to its AI glasses, letting the agent act on what the wearer is looking at. That points toward a larger strategy in which Muse becomes a persistent interface across Meta’s software and hardware ecosystem.

Muse Spark 1.3 Is Built for Agentic Work

Muse is powered by Meta’s Muse Spark model family. Muse Spark 1.3, released on September 2, was trained specifically for agentic workflows and coding rather than only conversational tasks.

Meta says the model is better at maintaining instructions during long tasks, working across conflicting information, identifying gaps in its plan, and asking the user for help when it cannot proceed safely. The company also says Spark 1.3 is more efficient than Spark 1.2 in internal coding comparisons, using roughly 20% fewer tool calls and 25% fewer tokens. These are Meta’s own measurements, not independent benchmark results.

Muse Spark 1.3 also has a one-million-token context window and is available through Meta Model API as well as Muse Code. That matters because Meta isn't keeping its agent strategy confined to the Muse consumer product. Developers can build their own agentic applications using the same model family.

Muse Spark 1.3 Pricing

Meta’s pricing is particularly interesting because agentic workloads can consume far more tokens than a normal chatbot session. A single job may involve repeated reasoning, browsing, tool calls, and revisions before it delivers anything.

The Contributor tier creates an unusually direct trade-off between price and data use. Developers willing to let Meta use their prompts and completions for future model development receive dramatically lower inference pricing.

For the consumer version of Muse, Meta says most usage is free, with paid options for heavier users. Reuters reported subscription tiers of $20 and $100 per month.

Security Becomes the Hard Part

Giving an AI agent more capability also increases the consequences of failure.

Meta says Muse uses a separate Sentinel agent that reviews activity before it reaches the internet. Credentials are stored separately so Muse does not directly see passwords or payment information, while users can control which applications and permissions the agent receives.

These protections are not a side feature. They are fundamental to whether personal agents can move beyond demonstrations.

A chatbot providing a poor answer may waste a user’s time. An agent with access to email, payments, or connected accounts can create a much larger problem if it misunderstands an instruction.

That is why the future competition in agents will not be based solely on reasoning benchmarks. Permission systems, credential isolation, auditability, and human approval mechanisms may become equally important.

OpenMuse Brings the Idea Into Open Source

Meta’s architecture did not remain only a proprietary product concept.

CopilotKit’s OpenMuse is an MIT-licensed personal-agent application explicitly inspired by Meta Muse. It combines a persistent Chromium browser, task management, files, goals, memory, and an optional Linux workspace. Users can inspect what the agent is doing and take control of its browser when required.

But OpenMuse should not be described as an open-source version of Meta’s actual Muse code.

It is an independent implementation of similar ideas, and the project itself labels the current release an alpha. Some of Meta Muse’s more ambitious capabilities, including automated checkout and broader integrations, are not yet implemented.

The technical architecture is also different. OpenMuse currently uses persistent Chromium together with an optional Docker-based Linux workspace rather than Meta’s managed Secure VM infrastructure.

Its importance, then, is not feature parity. It shows that developers can reproduce the personal-agent product model using open components and modify it for more control over the stack.

OpenClaw Changes the Story

Open source is not simply following Meta.

OpenClaw existed before Muse, and Meta has acknowledged its influence.

OpenClaw takes a more infrastructure-oriented approach. It runs on users’ own hardware and connects AI agents to services such as WhatsApp, Telegram, Slack, Discord, Signal, and other messaging platforms. Its Gateway manages sessions, tools, models, and channels, while users can choose different model providers rather than depending on a single proprietary model.

More importantly, Meta Superintelligence Labs product head Nat Friedman said Muse was “heavily inspired” by OpenClaw, while emphasizing that Meta built Muse from scratch. He said Meta wanted to create something similar that could be made safe and easy enough to scale to billions of users.

That makes the relationship between the three projects more interesting than a simple proprietary-versus-open-source comparison.

OpenClaw helped demonstrate the personal-agent model. Meta is turning that idea into a mass-market consumer product. OpenMuse is now rebuilding parts of Meta’s interpretation in an open application.

Meta Muse vs OpenMuse vs OpenClaw

The table also shows why these products should not be treated as direct substitutes.

Meta is trying to hide the complexity of running an agent. OpenMuse gives developers access to much of that complexity. OpenClaw goes further by making user-controlled infrastructure part of the product philosophy.

TDisrupt View: The Important Battle Is Moving Above the Model

Muse matters because it suggests the next stage of AI competition will not be decided only by which company has the highest-scoring model.

The execution environment surrounding the model is becoming equally important.

An agent must remember context, use software reliably, understand permissions, maintain long-running tasks, handle failures, and know when it needs human approval. The company that solves those problems can create something significantly more useful than a chatbot, even if rival models are relatively close in raw intelligence.

Meta has another advantage that most agent startups and open-source projects cannot easily reproduce: distribution. Facebook, Instagram, WhatsApp, and Meta’s growing AI glasses portfolio give the company multiple routes for putting Muse in front of users. Muse has already expanded rapidly since its September launch, and Meta is moving quickly to integrate it further across its products.

But open source changes the competitive equation.

OpenClaw shows that important agent concepts can emerge outside large AI labs. OpenMuse shows that sophisticated consumer-agent interfaces can be reconstructed quickly without controlling a frontier model.

For TDisrupt, the larger shift is clear: AI is moving from software that produces answers toward software that operates computers on our behalf.

If that transition continues, the most valuable AI platform may not be the model users talk to. It may be the agent that sits between the user and everything else they do online.

Meta wants Muse to occupy that layer.

Open source ensures Meta will not define it alone.

References

Meta: Introducing Muse

Meta AI Research: Introducing Muse Spark 1.3

Meta Model API: Muse Spark models and pricing

Meta Model API: Pricing and rate limits

CopilotKit: OpenMuse repository

OpenClaw repository

TechCrunch: Meta acknowledges OpenClaw's influence on Muse

Reuters: Muse pricing and business-model reporting