BrowserOS

BrowserOS is a privacy-first, open-source Chromium fork that embeds AI agents natively inside the browser itself. Unlike browser extensions that bolt on AI features, BrowserOS patches Chromium directly to expose internal browser state — accessibility trees, CDP sessions, render snapshots — to a TypeScript/Bun agent platform, enabling deep automation that external tools cannot replicate. It is the open-source alternative to ChatGPT Atlas, Perplexity Comet, and Dia.

The agent platform is a Bun monorepo exposing 53+ MCP tools for browser control: navigate, click, type, scrape, screenshot, evaluate JavaScript, manage tabs and windows, upload files, export PDFs, and run shell commands in a sandboxed workspace. A Go CLI (browseros-cli) lets you drive the browser from terminal or AI coding agents like Claude Code and Gemini CLI. The built-in MCP server makes BrowserOS a first-class citizen in any MCP-compatible workflow.

Beyond raw automation, BrowserOS ships a visual workflow builder for repeatable multi-step automations, scheduled tasks for unattended runs, persistent memory across sessions, and Cowork mode that combines browser automation with local filesystem access. Fourteen LLM providers are supported — from cloud APIs (Claude, GPT-4o, Gemini) to local models via Ollama and LM Studio — and all credentials stay on your machine.

What You Get

• AI Agent - Execute browser actions — navigate, click, type, scrape, screenshot, evaluate JavaScript — using natural language, powered by 53+ built-in MCP tools with no coding required.
• MCP Server - Expose full browser control via the Model Context Protocol so Claude Code, Gemini CLI, or any MCP client can drive BrowserOS from the terminal or external AI workflows.
• Visual Workflow Builder - Construct repeatable, multi-step browser automations using a drag-and-drop graph builder with support for conditions, loops, and sequential actions.
• Cowork Mode - Combine browser automation with local filesystem operations: scrape data from the web, then write reports, CSVs, or annotated PDFs directly to your selected folder.
• Scheduled Tasks - Run agents on autopilot — hourly, daily, or at custom intervals — for recurring jobs like price monitoring, email summaries, or competitive research without manual intervention.
• Persistent Memory - Maintain context across sessions so the agent remembers past interactions, user preferences, and task history for coherent long-running automations.
• SOUL.md - Define the AI agent’s personality, tone, and standing instructions in a single Markdown file that controls behavior globally across all tasks.
• LLM Hub - Compare responses from Claude, ChatGPT, Gemini, and other providers side-by-side on any webpage to evaluate accuracy, style, and reasoning quality.
• 40+ App Integrations - Connect to Gmail, Slack, Notion, GitHub, Linear, Salesforce, Figma, and more via MCP for cross-app automations without leaving the browser.
• browseros-cli - A Go CLI tool that lets you launch and control BrowserOS from any terminal or AI coding agent, with full MCP passthrough for programmatic access.
• Ad Blocking (MV2) - Enhanced tracker and ad blocking via uBlock Origin with Manifest V2 support, providing substantially more protection than standard Chrome.
• Cloud Sync - Sync browser settings, agent history, scheduled tasks, and workflow configurations across devices using encrypted cloud storage.

Common Use Cases

• Developer browser automation from Claude Code - A developer installs BrowserOS as an MCP server, then instructs Claude Code to test a staging site, capture console errors, fill forms, and file GitHub issues — all from the terminal without touching the browser manually.
• Morning briefing pipeline for founders - A founder schedules a daily workflow that pulls calendar events, Slack messages, and Gmail threads, then asks the agent to produce a prioritized briefing saved to a Markdown file in their notes folder.
• Competitor pricing monitor for marketers - A marketer builds a scheduled scraper that visits competitor product pages every morning, extracts prices into a structured table, and saves the result to a shared Google Sheet automatically.
• Sales CRM enrichment for sales reps - A sales rep configures BrowserOS to pull open deals from Salesforce, research each prospect’s recent LinkedIn activity, and draft personalized follow-up emails drafted in Gmail — all in a single automated workflow.
• Academic literature harvesting for researchers - A researcher schedules BrowserOS to search journal databases, extract abstracts and citation metadata, and organize findings into annotated PDFs stored in a local research directory.
• Local LLM browser automation for privacy-conscious teams - A team runs BrowserOS with Ollama so that browser automation tasks — parsing internal dashboards, filling enterprise forms — are processed entirely on-premise with no data sent to cloud APIs.

Under The Hood

Architecture BrowserOS is a dual-subsystem monorepo: a Chromium fork (packages/browseros) managed with a Python build system and patch stack, and an agent platform (packages/browseros-agent) built with Bun workspaces housing a TypeScript MCP server, a WXT-packaged React browser extension, a Go CLI, an eval harness, and shared libraries. The agent layer uses a clean layered architecture — a Bun HTTP server exposes MCP tools and an AI agent loop, decoupled from browser internals via a Chrome DevTools Protocol abstraction (CdpBackend). The BrowserSession and Browser facades own page lifecycle, while a provider-factory module handles all LLM routing without leaking provider details into the agent loop. Tool sets for browser control and filesystem access are built independently and composed at runtime based on session mode (chat vs workspace vs scheduled), enforcing a capability-boundary design where agents only receive the tools appropriate to their context.

Tech Stack The agent platform runs on Bun with TypeScript throughout; the AI agent loop uses the Vercel AI SDK (ToolLoopAgent) with provider adapters for Anthropic, OpenAI, Google, Azure, Amazon Bedrock, OpenRouter, GitHub Copilot, and local Ollama/LM Studio instances. Browser automation is built on Chrome DevTools Protocol with custom type-safe bindings (@browseros/cdp-protocol). The browser extension is packaged via WXT with React and Radix UI components. The CLI is written in Go using Cobra. Persistent storage uses Drizzle ORM over SQLite (Bun’s built-in runtime). Biome handles linting and formatting; Lefthook enforces commit hooks; the eval harness runs WebVoyager and Mind2Web benchmarks.

Code Quality The codebase has comprehensive test coverage organized into distinct groups — agent, API, browser, CDP, core, filesystem, integration, and tools — all runnable independently via a test-group runner. Test fixtures include a real browser fixture server and reusable environment setup/teardown utilities ensuring deterministic behavior. TypeScript strict mode is enforced project-wide with explicit type boundaries between server, agent, and extension packages. Error handling is structured: Result-typed config loading, typed exit codes from @browseros/shared, and Sentry integration for runtime errors. The CLAUDE.md contributor guide enforces strong conventions: extensionless TS imports, kebab-case naming, minimal comments only for hidden constraints, and shared constants via @browseros/shared to prevent magic value scattering.

What Makes It Unique BrowserOS’s core differentiator is patching Chromium’s extension API surface to expose internal browser state — accessibility tree nodes, render widget snapshots, AXNodeData — directly to the agent platform via CDP, rather than operating through the constrained web-extension layer that external automation tools are confined to. This enables the agent to achieve semantic understanding of page structure that content-script based tools cannot replicate. The MCP server architecture means BrowserOS acts as a first-class MCP host, making it natively composable with the growing ecosystem of AI coding agents (Claude Code, Gemini CLI, Codex) that speak MCP — a positioning no other Chromium fork currently occupies. The context compaction system (compaction.ts) uses a sliding-window summarization strategy with token estimation to handle long-running agent sessions without losing critical context, which is an unusually sophisticated approach for an open-source browser agent.

Self-Hosting

BrowserOS is released under the GNU Affero General Public License v3.0 (AGPL-3.0). In plain terms, AGPL is a strong copyleft license: you can use, modify, and self-host BrowserOS freely, including commercially, but if you modify the source and deploy it as a network-accessible service, you must release those modifications under the same AGPL terms. For teams using BrowserOS internally for their own workflows — without offering it as a service to third parties — the AGPL imposes no practical restrictions beyond attribution.

Operationally, self-hosting BrowserOS means running the Bun server and browser process on a macOS, Windows, or Linux machine that users have direct access to. There is no container image or Helm chart provided out of the box — BrowserOS is designed as a local-first desktop application. The agent server, CLI, and extension must be installed and kept updated manually; automatic updates are handled by the built-in CLI updater, but IT teams managing fleets will need to build their own distribution and update pipelines. The Chromium fork requires rebuilding from source for browser-level changes, which demands approximately 100GB of disk space and significant build time.

There is currently no commercial cloud or enterprise tier. All features — AI agent, MCP server, workflows, scheduled tasks, cloud sync — are available in the open-source release. Cloud sync uses an externally hosted service (files.browseros.com/cdn.browseros.com), so teams with strict data residency requirements should review what sync metadata is transmitted. Support is community-driven via Discord and Slack; there are no SLAs, enterprise support contracts, or managed upgrade paths. Teams that need guaranteed uptime, centralized administration, or vendor-backed support would need to build those capabilities themselves on top of the open-source foundation.