Skip to main content
WebRun is real-time browser automation infrastructure for AI agents. Give it a task in plain English, and an AI agent executes it in a real browser—clicking, typing, navigating, and extracting data autonomously. Core advantages:
  • Sub-100ms decisions — Hybrid CNN-LLM architecture delivers production-ready speed
  • Real-time video streaming — Watch sessions live as they execute
  • Human-in-the-loop controls — Manual takeover for credentials, approvals, or edge cases
  • Multiple integration paths — REST, WebSocket, MCP, or OpenAI-compatible APIs

​
What WebRun Handles

​
Data Extraction at Scale

Extract structured data from dynamic websites without building custom scrapers. Example task: “Search LinkedIn for engineering managers in San Francisco with 5+ years experience. Return their profiles as JSON with name, title, company, and LinkedIn URL.”

​
Form Automation

Navigate multi-step forms with conditional logic and dynamic fields. Example task: “Fill out this insurance quote form using the customer data in the attached JSON. Handle the multi-step workflow and return the final quote amount.”

​
E-commerce Workflows

Search products, compare prices, add items to cart—all through natural language commands. Example task: “Find wireless keyboards under $50 on Amazon, filter by 4+ star ratings, and return the top 3 with price and review count.”

​
Cross-Site File Transfers

Download files from one platform and upload them to another—moving documents, reports, and data between systems autonomously. Example task: “Download the monthly sales report from our analytics dashboard and upload it to the shared Google Drive folder.”

​
Adaptive Testing

Test web applications without maintaining brittle test selectors. The agent adapts to UI changes automatically. Example task: “Go through the signup flow with invalid email formats and document all validation errors you encounter.”

​
Sessions and Tasks

A session is an isolated browser instance. A task is a single instruction you give that session. Sessions can run indefinitely while active and expire after 5 minutes of inactivity, letting you chain multiple tasks sequentially. 
Session (sess_abc123)
 ├── Task 1: Navigate to site → completed
 ├── Task 2: Search for product → completed
 └── Task 3: Extract data → completed
Sessions terminate when you explicitly end them or when they hit the timeout. Always terminate sessions when finished to avoid unnecessary charges. Sessions documentation | Tasks documentation

​
Response Model

Most tasks complete in 10-40 seconds. WebRun waits up to 50 seconds before returning. If your task finishes within that window, you get the result inline. If it takes longer, you poll for the result. This gives you the simplicity of synchronous APIs for typical use cases, with async polling available for complex workflows. Response model details

​
Guardrails

When the agent needs human input—credentials, clarification, purchase approval—it triggers a guardrail and pauses. Your application detects this via the API response and provides the needed information to continue. Common triggers: Login forms, payment confirmations, CAPTCHAs, ambiguous instructions For login credentials, you can bypass guardrails entirely by providing secrets upfront when creating the session. Secrets are matched to websites by domain pattern and are never stored — they exist only in the session’s memory and are discarded when the session ends. Alternatively, use browser profiles to persist login cookies across sessions — log in once and reuse the saved browser state for future tasks. Guardrails documentation | Secrets documentation | Profiles documentation

​
Integration Options

Choose the integration method that fits your stack:
MethodBest ForReal-time Updates
REST APIServerless functions, simple integrationsPoll for updates
WebSocketLive dashboards, real-time monitoringYes
OpenAI-CompatibleDrop-in replacement for LangChain, Vercel AI SDKStream or poll
MCP ServerClaude Desktop, Cline, MCP clientsSession-based
Quick decision tree:
  • Single task, auto-terminate? → POST /start/run-task
  • Multiple tasks in sequence? → POST /start/start-session + /send-message
  • Using LangChain or OpenAI SDK? → POST /v1/chat/completions
  • Claude Desktop or MCP client? → MCP Server
REST API | WebSocket | OpenAI-Compatible | MCP Server

​
Why WebRun vs Traditional Automation

Traditional browser automation (Selenium, Playwright, Puppeteer) requires writing code that targets specific page elements. When the website changes—new class names, restructured HTML, different IDs—your automation breaks. WebRun’s AI agents don’t rely on selectors. They see the page visually, understand the content semantically, and adapt to changes automatically. No maintenance when websites update their UI. Speed comparison: WebRun delivers sub-100ms response times per action. Traditional tools average ~50ms per action, but require manual coding and ongoing maintenance that compounds costs at scale.

​
Get Started

Quickstart

First automation task in 5 minutes

API Reference

Complete endpoint documentation

Integration Guides

REST, WebSocket, OpenAI, MCP

Video Streaming

Watch sessions execute in real-time