I want to build the app people open when they need to get work done.
That means more than answering a question or completing a few clicks in a browser. Real work moves between websites, desktop apps, files, messages, and development tools. I want to give an agent a goal, let it work across those boundaries, and stay in control whether I am at my computer or checking in from my phone.
That is where I am taking BrowserAI DUO: an agentic browser and desktop workspace that can eventually run on your Mac or an approved remote machine.
BrowserAI DUO is the chosen name for the project previously called BrowserRAD. It started as a separate fork of BrowserOS Neo, built on Chromium. There is an experimental browser foundation and an early macOS desktop-control prototype. The unified experience, integrations, and editions described below are work I plan to build. They are not a launch announcement or a claim that every feature already works.
The problem: the task keeps moving
Consider a workflow like this:
Research three vendors, compare their plans, download the relevant documents, build a comparison spreadsheet, and leave a recommendation ready for my review.
The research happens in browser tabs. Some documents may sit behind a login. The spreadsheet might belong in a desktop app. The finished work needs somewhere to go, and the agent needs to know when it should ask before taking the next step.
I want BrowserAI DUO to carry that task through the whole process, with a visible history of what happened and evidence that the requested result exists.
The space already includes browser assistants such as Comet, Dia, and Opera Neon, alongside broader agent products such as ChatGPT agent. Browser automation platforms such as Browserbase and Browser Use offer another approach for developers.
Those are useful reference points. My direction is to bring browser work, native desktop work, remote execution, and user control into one coherent workflow. I am using the wider category to inform the roadmap; I am not claiming feature parity with every product in it.
A browser that can carry out the work
The browser should be a place where an agent can act on an instruction and check the result.
I am planning for research across multiple tabs, structured extraction, comparisons, form filling, downloads, uploads, and work inside authenticated sites you authorize. Useful routines should become reusable workflows, with scheduling and progress reporting as later capabilities.
For example, a recurring research task could gather changes from approved sources, compare them with a previous run, and prepare a report. A separate approval would govern any step that sends that report or changes an external record.
Underneath that experience, I want the agent to use page structure and direct browser evidence wherever possible. Screenshots and visual reasoning should help when needed, with their extra latency and cost visible in evaluation.
The earlier BrowserHelm design article explains related thinking about bounded execution, live page checks, and keeping the reasoning layer connected to reliable browser actions.
Native macOS work in the same task
Desktop control is what makes this more than a browser project.
The macOS version should be able to work with approved applications, windows, menus, file dialogs, and files or folders the user has granted access to. A workflow could collect information on the web, open a document in a native app, make the requested changes, and return to the browser without starting a separate conversation.
I am starting with accessibility information and native controls where available. Visual interaction is a fallback for interfaces that need it. Desktop access should be an explicit capability with understandable permissions and a visible way to stop execution.
An initial prototype is a useful starting point, but it does not establish reliable control of arbitrary applications. Packaging, permission handling, recovery, and testing across real workflows still have to be built out. The experimental work is available on the BrowserAI DUO desktop branch.
One task, with room to step in
The experience I want is straightforward: describe the outcome, review the plan when appropriate, and see what the agent is doing.
Browser and desktop actions should share a task history. That history should make it possible to inspect progress, understand a blocker, pause execution, take over manually, and resume without reconstructing the entire job.
A status message such as “done” is not enough. If the task was to create a file, the agent should confirm the file exists and check its contents. If it changed a setting, it should read back the resulting state. If a write might have succeeded but the response is unclear, it should stop and reconcile the state before attempting it again.
This is one of the central engineering challenges: making a long task understandable and recoverable when a page changes, an app loses focus, or a login needs attention.
Control it from your phone
I already work from my phone while my Mac handles tasks elsewhere. That is part of the motivation behind Vibesomnia, my macOS keep-awake utility.
Today I can reach an AI session on my Mac from my phone using my existing remote-access setup. That session can call the browser’s MCP tools on the Mac. A dedicated BrowserAI DUO phone interface could later add independent job status, approvals, and notifications, but basic remote use does not require a third app capability.
The executing computer would still do the work. Native desktop tasks would depend on that machine being awake, connected, and in a usable macOS session. A phone connection cannot remove those requirements.
Remote access, authentication, reconnection, and approval delivery need to be designed together. The goal is to make checking in from the couch or away from home feel like a normal part of the workflow.
Give longer jobs a remote worker
Some tasks belong on a dedicated machine.
I want BrowserAI DUO to support approved remote workers for browser jobs, development environments, builds, tests, and other longer tasks. That could let a worker carry on while the user keeps their own computer free.
The execution environment matters. A remote Mac would be needed for native macOS application workflows. A compatible server or development environment could handle browser automation, scripts, and command-line tools. I am not assuming a headless Linux machine can operate someone’s Mac apps.
A developer workflow might inspect an issue, work in an isolated checkout, run checks, and prepare a change for review. The user should be able to see which machine is executing it, what it can access, and what remains unfinished.
Password-manager access for approved sites
Login is a practical requirement for useful automation.
I plan to connect BrowserAI DUO to password managers so it can request autofill for approved sites and accounts. The password manager should remain the authority for unlocking and granting access. The design goal is to keep passwords out of model prompts, ordinary logs, and saved workflow definitions.
That requires more than attaching a vault integration. The requested account and destination need to match the user’s approval. Unlock prompts, multifactor authentication, unexpected redirects, or a changed destination may require a human handoff.
I have not announced supported password managers. Provider support and the exact security properties will depend on implementation and testing. This should make approved logins easier while keeping the user involved in access decisions.
A workspace for people and other agents
BrowserAI DUO should also be useful alongside the AI tools people already use.
I am exploring MCP and API interfaces so external agents can request bounded browser and desktop actions. Claude and Codex are part of my current development workflow; an embedded conversational experience inside the app is a separate milestone.
For development work, the roadmap includes access to relevant screenshots, console and network evidence, approved terminal commands, and build or test results. Model choice, local execution options, and provider costs also need evaluation. I do not want routine actions to require an extra model call without a demonstrated benefit.
Personal and Enterprise editions
I am planning two editions around the same core idea.
Personal would focus on an individual’s browser, Mac, approved accounts, files, and workflows. The priority is making everyday work easier to start, follow, and finish.
Enterprise would add the controls organizations need to manage access and accountability. Planned areas include organizational identity and SSO, roles, managed credentials, workspace boundaries, audit records, retention policies, and deployment options for private environments.
Those are design requirements for future work. Enterprise readiness will require dedicated hardening and operational validation. There are no compliance certifications, pricing, or availability dates to announce.
What exists, and what comes next
| Area | Status |
|---|---|
| Separate BrowserAI DUO browser foundation | Experimental fork of BrowserOS Neo |
| Native macOS desktop control | Early prototype; broad application reliability remains unverified |
| Unified browser and desktop task experience | Planned |
| Phone control and approved remote workers | Planned |
| Password-manager integration | Planned; providers not announced |
| Personal and Enterprise editions | Planned |
The next gate is a dependable browser-and-macOS workflow with explicit permissions, a shared task interface, and independent checks of the outcome. Remote control and credential integrations need to build on that foundation.
I will evaluate completion, accuracy, recovery, elapsed time, and actual usage and cost where those measurements are available. There is no universal speed or token-saving claim behind this roadmap.
The ambition is big: make BrowserAI DUO the app people reach for to get work done. Earning that position means finishing useful jobs consistently and giving people enough visibility to trust the result.
Questions about the roadmap
Is the combined app available now?
No. BrowserAI DUO is experimental, and the browser-plus-desktop product described here is still being developed. This article explains the direction rather than announcing a finished release.
Will it work from a phone?
Phone control is planned for starting tasks, monitoring work, approvals, and reviewing results. The browser or desktop work would execute on an authorized host machine.
Will it support Windows and Linux desktops?
Native desktop work is focused on macOS first. Remote browser and development jobs may use other compatible environments. I am not announcing native Windows or Linux desktop support here.
Why DUO?
DUO describes the two intended local surfaces: agentic browser and desktop control. The browser exists today; desktop control is still in development.
I will share concrete progress as the pieces become usable. You can inspect the BrowserAI DUO browser branch, follow the desktop experiments, or get in touch about a workflow you would want it to handle.