BrowserOS Launches Local Filesystem Access for AI Agents

Two brothers have quietly built an open-source alternative to the AI browsers dominating the market, and their privacy-first architecture is now paying unexpected dividends. Nithin and Nikhil, twin founders of BrowserOS, have launched a platform that runs AI agents entirely on the user's machine, not in the cloud.

The timing couldn't be more relevant. As major labs release browser-based AI agents, the founders realized their decision to prioritize local processing nine months ago accidentally positioned them perfectly for this moment. Their browser agent can now read files, write documents, and run shell commands—capabilities that server-side competitors struggle to offer without compromising user privacy.

Unlike competitors like ChatGPT Atlas and Perplexity Comet where the agent loop runs server-side, BrowserOS made a bold bet early on: run everything on the client. This decision wasn't easy. The founders initially built their agent loop inside a Chrome extension, but quickly hit technical walls.

The limitations were significant:

• JavaScript's single-threaded nature prevented parallel agent execution
• Lack of NodeJS runtime blocked access to essential npm packages like Vercel AI SDK
• No clean way to expose agent tools as an API

Two months ago, they made the hard choice to scrap everything and start over. The new architecture uses a sidecar approach, placing the agent loop in a standalone Bun binary shipped alongside their Chromium binary. Rather than rewriting from scratch, they borrowed gemini-cli's loop with modifications, creating an adapter to translate between Gemini format and Vercel AI SDK format.

When Anthropic launched Claude Cowork, the BrowserOS team noticed a critical limitation: server-side agents can't access local files without uploading them first. This creates both privacy concerns and latency issues. BrowserOS's local architecture meant adding filesystem access was simply a matter of opening the door with user permission.

The agent now operates with capabilities similar to Claude Code, reading and writing files directly on the machine. No cloud storage, no synchronization, no uploads. This local-first approach enables practical workflows that would be cumbersome with server-side alternatives.

Our agent can now read and write files just like Claude Code. No uploads, no cloud storage, no sync.

Users can demonstrate this capability through concrete examples: organizing files in desktop folders or opening top Hacker News links, extracting details, and writing summaries directly into HTML files—all executed locally.

Since their last public announcement, BrowserOS has evolved significantly. The platform now supports n8n-like graph workflows for building reliable automation sequences. Users can also integrate BrowserOS as an MCP server within development environments like Cursor or Claude Code.

Perhaps most notably, the system now supports scheduling repetitive tasks, transforming the browser from a passive tool into an active assistant. The traction reflects this utility: the project has grown to 8.5K GitHub stars and surpassed 100,000 downloads.

These metrics demonstrate real-world adoption beyond theoretical appeal. Knowledge workers—who spend most of their time in browsers managing emails, documents, spreadsheets, and research—now have a tool that operates where they already work, without forcing data through external servers.

The founders are "very bullish" on the browser as the ideal platform for AI agents. Their reasoning is straightforward: browsers are the most commonly used application for knowledge workers. Even Anthropic recognizes this, though their Claude Cowork integration relies on what the founders describe as a "janky" Chrome extension.

Owning the entire stack enables smoother experiences and differentiated features. One innovation in development is Browser ACLs—access control lists that function like IAM for agents. These guardrails would prevent destructive actions, with rules like "role(agent): can never click buy" or "role(agent): read-only access on my bank's homepage."

A prototype already exists. The team is actively seeking feedback on this thesis and the broader vision of local-first AI agents that respect user privacy while delivering powerful automation capabilities.

BrowserOS represents a fundamental shift in how AI agents interact with user data. By keeping everything local, the platform sidesteps the privacy trade-offs that plague cloud-based alternatives while enabling capabilities that server-side architectures cannot easily replicate.

The project is available for Mac, Windows, and Linux, with the codebase open for public inspection. As AI agents become more integrated into daily workflows, the choice between local and cloud processing will likely define the next generation of productivity tools.

For now, BrowserOS stands as a compelling alternative—one that proves privacy and power don't have to be mutually exclusive in the age of AI.