BrowserOS – "Claude Cowork" in the browser

Establishes direct WebSocket or HTTP connections to Anthropic's Claude API from the browser runtime, implementing streaming message handlers that progressively render token-by-token responses without server-side proxying. Uses fetch-based request construction with automatic retry logic and connection pooling to maintain low-latency interactions with Claude models (3.5 Sonnet, Opus, Haiku).

Provides a shared browser-based workspace where multiple users can simultaneously interact with Claude, with real-time state synchronization across clients using operational transformation or CRDT-based conflict resolution. Implements presence awareness (cursor positions, active selections) and message history merging to maintain consistency across concurrent edits and Claude interactions.

Implements an agentic loop where Claude can call tools (functions) defined in the browser, with automatic tool result handling and replanning. Supports function schema definition (JSON Schema) and automatic tool invocation based on Claude's function_calls output. Implements multi-turn agent loops where Claude reasons about tool results and decides next steps, with execution logging and error recovery for failed tool calls.

Indexes the user's local codebase (JavaScript, TypeScript, Python, etc.) into an in-browser vector store or AST-based symbol table, then uses Claude to generate code completions that reference existing functions, classes, and patterns from the indexed codebase. Implements semantic search across indexed symbols to retrieve relevant context before sending completion requests to Claude, enabling structurally-aware suggestions that respect project conventions.

Enables users to construct Claude prompts by attaching files (images, PDFs, code snippets, documents) directly in the browser, with automatic format detection and encoding. Implements file preview rendering (image thumbnails, code syntax highlighting, PDF text extraction) and drag-and-drop file upload with progress tracking. Converts files to base64 or Claude-compatible formats (vision tokens for images, text extraction for documents) before sending to Claude API.

Persists Claude conversation history to browser local storage (IndexedDB or localStorage) with automatic serialization of messages, metadata, and context. Implements conversation search across stored history using full-text indexing, and provides export functionality (JSON, Markdown, PDF) to save conversations locally. Handles storage quota management (typically 5-50MB per domain) with automatic cleanup of old conversations or user-initiated archival.

Provides a templating engine for constructing reusable Claude prompts with variable placeholders (e.g., {{variable_name}}, {{file_content}}, {{timestamp}}). Implements a prompt library UI where users can save, organize, and version-control prompt templates. Supports conditional blocks (if/else) and loops for dynamic prompt generation based on user inputs or file content, with real-time preview of rendered prompts before sending to Claude.

Enables users to define multi-step workflows where Claude processes a task sequentially, with each step's output feeding into the next step's input. Implements a visual workflow builder (DAG-based) where users can chain Claude calls, conditional branches, and file operations. Handles state management across steps (storing intermediate results, passing context between steps) and provides execution logging with step-by-step progress tracking and error recovery.

Provides a sandboxed JavaScript/Python execution environment within the browser where Claude-generated code can be executed safely, with output capture (console logs, return values, errors) and visualization. Implements security boundaries using Web Workers or iframe sandboxing to prevent malicious code from accessing the DOM or local storage. Supports interactive code execution with user input handling and real-time result rendering.

Integrates Claude into a code review workflow by analyzing code diffs and generating review comments with suggestions for improvement. Implements diff parsing (unified diff format) and line-by-line annotation where Claude comments are attached to specific lines of changed code. Provides structured review output (bugs, style issues, performance suggestions, security concerns) with severity levels and actionable recommendations.

Enables users to upload documents (PDFs, Markdown, text files) which are indexed into a browser-based vector database using embeddings. Implements semantic search across the knowledge base where user queries are embedded and matched against document chunks using cosine similarity. Retrieved relevant documents are automatically included as context in Claude prompts, enabling RAG (Retrieval-Augmented Generation) workflows entirely in the browser.