Autonomous Agent2026Shipped
Web Agent
An LLM agent that browses, fills forms and downloads over raw CDP.
An agent that navigates real websites, fills forms and downloads files. Each page becomes a semantic tree built from live Chrome DevTools Protocol DOM traversal instead of XPaths, which keeps real parent-child structure, ARIA roles, and text nested in inline elements that flat element lists lose.
livelast push
~/web-agent
Project Telemetry
Task success
64%
Benchmark
50 tasks
Websites
10
LLM providers
8
How it works
~/web-agent/architecture · step 1/3
01
live page via CDP
02
semantic DOM tree
03
WebMCP site tools
04
LLM planner · 8 providers
05
click / fill / download
↺ loops to 01
01OAuth 2.0 Authorization Code + PKCE end to end: local callback server, token exchange, bearer injection, and disk-persisted tokens with silent refresh.
028 LLM providers behind one interface; WebMCP support for discovering tools a site exposes at runtime.
03Benchmarked on 50 self-defined tasks across 10 websites: 64% success.
Project Info
DomainAI Agents
RoleAuthor
Year2026
StatusShipped
Tech Stack
PythonChrome DevTools ProtocolOAuth 2.0 PKCEWebMCPMulti-provider LLMs