TL;DR
- Manus launched My Computer, a native desktop AI agent that runs locally on macOS and Windows — not in the cloud.
- The agent operates through terminal commands with explicit user permissions and trains on local GPU hardware.
- It handles file management, coding tasks, and app control directly on your machine, competing with Anthropic Dispatch and Perplexity’s agents.
- Privacy and security concerns linger despite built-in approval safeguards — you’re handing terminal access to an AI.
Manus Brings the AI Agent to Your Actual Computer
Manus dropped My Computer this week, and it’s not another chatbot wrapper. It’s a native desktop agent that runs locally on macOS and Windows, executing tasks through terminal commands with explicit user permissions at every step. The company — which operates under the Meta Manus brand — designed the system to handle file operations, coding workflows, and app control without routing everything through cloud infrastructure.
The agent trains on local GPU hardware, which means your machine does the heavy lifting. No data leaves your device unless you explicitly tell it to. That’s a sharp departure from the cloud-first design philosophy most AI companies have embraced since ChatGPT kicked off this arms race.
My Computer slots into a growing category of autonomous desktop agents. Anthropic shipped Dispatch. Perplexity rolled out its own agent features. But most of those tools still lean heavily on cloud compute and API calls. Manus is betting that local execution — with all its friction and hardware demands — will win over developers and privacy-conscious users who don’t want their workflows uploaded to someone else’s servers.
Why Local Execution Changes the Agent Game
Here’s the thing: cloud agents are fast, scalable, and easy to deploy. But they also require you to trust that your data, your code, your file structure won’t get swept into a training corpus somewhere. My Computer flips that trade-off. You get full control and zero data exfiltration risk, but you also need a machine with enough GPU horsepower to run the model locally.
That’s not a small ask. Most consumer laptops can’t train or run inference on meaningful models without choking. Manus is clearly targeting developers, data scientists, and power users who already have the hardware — or are willing to buy it. For everyone else, this is a preview of where the category might go, not a tool they’ll adopt tomorrow.
And yet, I think this is the right bet for a certain slice of the market. Developers don’t want their proprietary code touching third-party servers. Researchers working with sensitive datasets can’t risk cloud leakage. My Computer gives them an option that didn’t really exist before: an agent that’s genuinely autonomous but also genuinely local.
The explicit permission model is critical here. Every terminal command the agent wants to execute requires user approval. That’s friction, sure. But it’s also the only way to make this remotely safe. An AI with unchecked terminal access is a disaster waiting to happen — one bad hallucination and it’s recursively deleting your home directory.
Think of it like handing your car keys to a valet. You’re trusting them to park it, not joyride across state lines. My Computer’s permission gates are the equivalent of making the valet ask before leaving the parking lot. Annoying? Sometimes. Necessary? Absolutely.
Security Concerns Won’t Disappear With a Checkbox
But let’s not pretend the permission model solves everything. Users are notoriously bad at evaluating risk in the moment. If the agent asks to run a script and you’re deep in a workflow, you’ll probably just click yes. That’s human nature. The approval gate only works if people actually read what they’re approving — and most won’t.
The bigger issue is what happens when something goes wrong. Who’s liable if the agent misinterprets a command and wipes a critical file? What happens if a bug in the local model causes data corruption? Manus hasn’t publicly detailed its error-handling or rollback mechanisms, and that’s a gap that matters.
Privacy advocates will love the local-first architecture. Security researchers will have questions. And both are right. My Computer eliminates one entire class of risk — cloud data leakage — but introduces new ones around local execution, permission fatigue, and user error. That’s not a knock on Manus. It’s just the reality of building autonomous agents that actually do things.
The competitive context sharpens the stakes. Anthropic Dispatch and Perplexity’s agents are already in market, and both have momentum. Dispatch integrates tightly with Claude‘s reasoning capabilities. Perplexity’s agent leans on search and real-time data. My Computer’s edge is its local execution and GPU training, but that’s also its constraint. You can’t just spin up My Computer on any laptop and expect it to work. It demands hardware most people don’t have.
The Industry Is Sprinting Toward High-Autonomy Local Agents
Manus isn’t alone in this push. The entire industry is racing toward locally deployed, high-autonomy agents. Apple reportedly has similar projects in the works. Microsoft is embedding agentic features into Windows. Google’s Gemini Nano runs on-device. The trend is clear: the next wave of AI won’t just live in the cloud. It’ll run on your phone, your laptop, your desktop — wherever the compute exists.
That shift creates new infrastructure demands. Local models need optimization. Inference needs to be fast enough that users don’t notice latency. Training needs to happen without melting your GPU. Manus is solving some of those problems with My Computer, but the broader ecosystem still has work to do.
The other challenge is user experience. Cloud agents are simple: you type, they respond, everything happens invisibly. Local agents require setup, hardware checks, permission flows, and troubleshooting when something breaks. That complexity is a feature for power users and a dealbreaker for everyone else. Manus is clearly targeting the former, which is smart. But it also means My Computer won’t be a mass-market product anytime soon.
Still, this is where the category needs to go. If AI agents are going to handle sensitive tasks — managing finances, writing code, controlling infrastructure — they can’t all route through centralized cloud providers. The attack surface is too large. The privacy risks are too high. Local execution isn’t just a nice-to-have. It’s a necessity for certain use cases.
What Happens When Local Agents Go Mainstream
The immediate question is adoption. Who actually uses My Computer in its current form? Developers who need to automate repetitive terminal tasks. Data scientists who want an AI assistant that doesn’t phone home. Security-conscious users who refuse to send their workflows to OpenAI or Anthropic. That’s a niche, but it’s a valuable one.
The longer-term question is whether Manus can scale this beyond power users. That probably requires hardware partnerships — preloading My Computer on gaming laptops with beefy GPUs, or bundling it with workstations aimed at creative professionals. It also requires better onboarding and smarter defaults. If every single action requires manual approval, most users will abandon it after a day.
Watch how Manus iterates on the permission model. If they can build trust layers that let users whitelist certain actions or set approval thresholds, the friction drops dramatically. Watch how competitors respond. If Anthropic or Perplexity ship local execution modes, Manus loses its primary differentiator. And watch the hardware side. If Nvidia or AMD start optimizing chips specifically for local agent inference, the entire category accelerates.
FAQ
What is My Computer by Manus?
My Computer is a native desktop AI agent from Manus that runs locally on macOS and Windows. It operates through terminal commands with explicit user permissions, handles file management and coding tasks, and trains on local GPU hardware without sending data to the cloud.
How does My Computer differ from cloud-based AI agents?
Unlike cloud agents from Anthropic or Perplexity, My Computer executes all tasks locally on your machine using your GPU. No data leaves your device unless you explicitly direct it to, eliminating cloud data leakage risks but requiring more powerful local hardware.
Is My Computer safe to use with terminal access?
My Computer requires explicit user approval for every terminal command it wants to execute, which adds a safety layer. However, security concerns remain around permission fatigue and user error — if you approve commands without reading them carefully, risks increase.
What hardware do I need to run My Computer?
My Computer trains on local GPU hardware, so you’ll need a machine with sufficient graphics processing power. Most consumer laptops won’t handle it well — the agent targets developers, data scientists, and power users with dedicated GPU setups.
Source: MarketingProfs
