Claude Duet, Matrix OS, and the new wave of personal AI agents

Why your own computer changes the agent's work

Imagine: in the evening you left a task to gather materials for an article, and in the morning you opened a source table, a draft, and a list of questionable statements. The most important thing here is the ability to continue working without a constant human presence. This is exactly the experience around which personal AI agents are being built.

The agent combines a model, tools, context, and an action loop: it receives a task, does something, checks the result, and chooses the next step. A persistent cloud computer provides a place where files and running processes can remain. This still doesn't guarantee a correct result: finishing a process, saving a file, and fulfilling the user's intent are different things.

For Vibe Coding, the change is practical. Instead of moving snippets of code between chats, you can work on a single project: set a task, receive changes, verify them, and return to the same files later.

Claude Duet: a proactive agent, but not yet for Ukraine

On October 21, Anthropic opened the research preview of Claude Duet. It is a separate application that transfers memory, skills, and connected services from Claude, prepares a morning brief, takes tasks into work, and returns for confirmation before acting on behalf of the user. The company's own examples involve searching for unpaid invoices, drafting reminders, preparing meetings, dashboards, and web applications.

The main limitation for the Ukrainian reader: as of October 22, it is a waitlist for Pro and Max subscribers in the USA; Team and Enterprise are not yet included. Anthropic also warns that the preview may misidentify urgency, misread a message chain, or miss an event between checks. Background checks and large tasks consume the Claude plan's shared limit.

The claim of "over 6,000 connectors and plugins" describes potential integration coverage but does not guarantee equal quality for every scenario. It is more useful to verify one working route: what data the agent reads, what it creates, where it asks for confirmation, and how to revoke access. Anthropic describes separate protections for credentials, actions, prompt injection, and sensitive data leaks, but you still shouldn't start a research preview with payroll, cards, or mass mailing.

What has already happened: Duet, Muse, Dots, and Grok Bot

The phrase "all major AI companies have moved to agents" is too broad. However, several major players have indeed taken concrete steps:

• October 21 — Anthropic introduced Claude Duet. Official product page describes a separate mobile and web application on Opus 5.5, morning briefs, background work, and confirmation of important actions. This is a limited research preview, not general availability.
• September 8 — Meta introduced Muse. via the official announcement, the personal agent works in a dedicated virtual machine with a browser; you can communicate with it via the app and WhatsApp. On September 29, Meta added scenarios for small businesses and new connections to work services. In this announcement, availability is specified for the USA and Canada — it should not be automatically assumed for Ukraine.
• September 29 — OpenAI introduced Dots. In company publications describe its own cloud computer, working with connected applications, and managing multiple projects. Deployment depends on the plan, market, and organizational settings. A separate nuance: background proactive research is limited to reading tools; it is not a permission to independently send messages or change data.
• October 12 — Meta announced Muse Wonder and Wonder Full. The company emphasizes memory and the ability to work longer on a task. In the announcement shared access to models via API/Muse Code and subsequent implementation into Muse. A new model and its appearance for every app user is not a single event.
• October 20 — a practical case of Grok Bot in support. The team shared about IncidentWatch, which searches for spikes in similar requests, and IncidentSweep, which prepares a response package. According to its data, during an incident on September 18, 204 requests were processed in approximately 10 minutes, including engineer verification and approval. This is the company's own case, not an independent speed test.

Our conclusion from these announcements: competition is shifting toward long-running workflows — with memory, access, and checkpoints. Comparing such products solely by the beauty of a chat response is no longer enough.

Matrix OS: a shared workspace for different agents

Matrix OS offers its own cloud workspace with files, a terminal, applications, and agents. Its idea is particularly close to those who want to create tools for their specific tasks and use different AI assistants.

The name OS does not mean you have to replace Windows or macOS. It describes the architecture shows an environment on top of server infrastructure that the user connects to. There is a managed cloud and a path for independent installation on a Linux server. The runtime code is open; self-deployment simultaneously leaves server maintenance, updates, and backups to the owner.

Instructions for coding agents directly describes Claude Code, Codex, OpenCode, and Pi. The main page additionally names Cursor and Gemini CLI. The list in the marketing block is broader than the step-by-step terminal menu, so we do not assume the same level of readiness for all integrations.

Hermes and OpenClaw have a different role: they are configurable environments for connected messaging channels. They require a model, access to a provider, and permissions. Choosing such an agent does not mean automatic access to all files and conversations.

A shared folder also does not create a cohesive team from several agents by itself. If two assistants change the same file, a workflow is needed: separate tasks, branches, or sequential changes and verification. The number of active agents is a weak indicator of utility; a completed and verified result is a stronger one.

How the approaches differ

This is a comparison of intended purpose, not a quality rating. Grok Bot's capabilities are described on the product page; the rest are in the documents mentioned above. We did not conduct the same practical test on all five systems.

How much Matrix costs and what "starting offers" means

via the payment documentation, monthly prices are $20 for Starter, $100 for Builder, and $200 for Max. The first main computer can have a three-day trial period with a linked card. After that, billing begins unless the subscription is canceled in time. Additional computers are paid for separately.

The Matrix plan pays for the computer. Subscriptions and the use of third-party AI providers are paid for separately. Therefore, $20 is not the price for unlimited work with Claude, Codex, and other agents combined. Final costs depend on the chosen model access.

The documentation also describes a scheduled one-time credit $5 for Matrix AI for eligible users. It explicitly states that the update has not yet been rolled out to everyone. Presenting this as a guaranteed bonus for every new user would be incorrect.

Even a running computer does not bypass AI provider limits, rate limits, or waiting for permission. "Can work after closing your laptop" does not mean "guaranteed to complete any task without you."

What about iOS and Android

The Desktop client is described for macOS, Windows, and Linux. Regarding phones, the current documentation talks about early access and active development. It does not explicitly announce availability in the Apple App Store; the main page is still awaiting public store links.

There is also a significant limitation: opening a created app in a mobile preview does not prove that its data will be correctly saved via the Matrix database bridge. For such apps, the documentation recommends verified web or desktop surfaces. Therefore, we do not repeat the phrase "full mobile version coming next week" without a dated confirmation.

What you can do for yourself

Below are our task examples, not reported Matrix testing results. They help determine if you need a separate agentic workspace at all.

A research folder for an author or marketer

The agent collects specified sources, records dates and links, and separates new facts from repetitions. The result is a table for an editor, not an automatic publication of everything found. Verification is simple: do the sources open, and do they support what is written?

A small CRM for a freelancer

Contact, project, next step, and deadline — this is often enough for a first version. Start with fictional entries, test editing, searching, and exporting. Connecting real email is a separate step with specific permissions.

An edit tracker for a designer or 3D animator

You can try converting client comments into a list of frames, edits, and questions. Let the agent flag ambiguities and save links to the original comment. You should not task it with independently promising new deadlines to clients or rewriting work scenes without approval.

Here is a starter task that is easy to evaluate:

Create a simple edit tracker for my project. Fields: frame, comment, priority, status, deadline. For the first run, use 5 fictional entries. Add editing, a status filter, and CSV export. Verify that entries remain after a restart. Do not connect email or send messages. Show the result, a list of checks, and known shortcomings. Ask me before performing paid actions or accessing other folders.

Help verify one repetitive task without risk. Task: [describe one process in one sentence]. Test data: use only fictional examples. Result: [name one file, table, or list]. Success criterion: the result is complete, opens, and contains a checklist. Forbidden: sending messages, publishing, paying, changing accounts, or opening other folders. If data is missing — stop and ask one specific question. At the end, show: what was done, what was not done, which permissions were needed, and how many steps required my confirmation.

Supplement text instructions with technical constraints: a separate folder, limited permissions, and a spending limit where available. To help with this, we have prepared an AI agent access map. If you need connection basics — there are the MCP explanation in Ukrainian.

When it makes sense

Matrix looks interesting for a person who regularly creates their own tools and wants to keep code, files, and agents close by. A ready-made personal assistant might be more convenient if delegating work is more important to you than setting up the environment. This is our conclusion from the product builds, not the result of a competition between them.

Practical selection criterion: which repetitive task does the agent bring to a verified result, how much does it cost, and how much of your attention does it require? Also calculate the time for fixes, model costs, and maintaining your own applications. Sometimes a regular table still wins — and that is also a useful test result.

For our community, this is a natural extension of Vibe Coding: creating small programs for your own needs and gradually testing what can be delegated. You can discuss your examples and failed experiments in the Ukrainian Vibe Coding community.