Releases·Tool Release
PawBrowse open-source browser automation lets AI agents read pages as tables, halving round-trips vs screenshot-based drivers
A developer released PawBrowse, an open-source Chrome extension plus a single-file MCP server that lets AI agents like Claude Code or Cursor control a real, logged-in browser. Instead of screenshots, it gives agents a compact table of clickable elements with stable refs, letting each action complete in one round-trip. On a measured Booking.com task it used 5 agent round-trips versus 10 for screenshot-based driving, roughly halving latency and token cost.
10 engagement1 sourceSun, Sep 27, 2026, 11:27 AM
AI Overview
- PawBrowse is an open-source (MIT) Chrome extension plus one-file zero-dependency MCP server letting agents like Claude Code and Cursor see and click logged-in Chrome tabs
- It hands agents a compact table of clickable elements with stable refs, so actions complete in one round-trip instead of screenshot-infer-screenshot loops
- On a filmed Booking.com Lisbon hotel-booking task, it needed 5 agent round-trips vs 10 for a perceive-then-act driver — ~2x fewer round-trips and token cost
- It drives the user's real logged-in Chrome with no remote-debug port, no relaunch, no second model, and no API key; available on Chrome Web Store and npm
Related
Products·launched
PawBrowse - faster agent browser automation via element tables instead of screenshots
PawBrowse is an open source browser automation driver for LLM agents. Instead of screenshot-based perceive-then-act loops, it gives the agent a compact table of clickable elements (accessible names, values, state) refreshed after each action, halving round-trips. It targets Claude Code users building browser-driving agents.
PawBrowse is an open source browser automation driver for LLM agents. Instead of screenshot-based perceive-then-act loops, it gives the agent a compact table of clickable elements (accessible names, values, state) refreshed after each action, halving round-trips. It targets Claude Code users building browser-driving agents.
17 engagement1 sourceredditSun, Sep 27, 2026, 11:47 AM
Products·launched
OnBridge — Chrome extension + MCP server letting coding agents drive your logged-in browser
OnBridge is a Chrome extension plus MCP server that lets Claude Code, Codex, or Cursor browse, click, and type in your real Chrome session with your existing logins and extensions, rather than a blank headless instance. An ask_user tool blocks agent actions until you approve them in a side panel, protecting against unwanted spending. It targets developers who want agents to act on authenticated sites.
OnBridge is a Chrome extension plus MCP server that lets Claude Code, Codex, or Cursor browse, click, and type in your real Chrome session with your existing logins and extensions, rather than a blank headless instance. An ask_user tool blocks agent actions until you approve them in a side panel, protecting against unwanted spending. It targets developers who want agents to act on authenticated sites.
1 engagement1 sourceredditFri, Sep 25, 2026, 07:38 PM
Products·launched
Web Draw — Chrome extension + MCP server that gives AI agents web pages as text instead of screenshots
Web Draw is a free Chrome extension paired with a local MCP server that renders the visible DOM as compact text so LLM agents can read and drive web pages. It replaces the common screenshot-plus-vision-model approach, cutting token use per step and helping models see small buttons and form validation errors.
Web Draw is a free Chrome extension paired with a local MCP server that renders the visible DOM as compact text so LLM agents can read and drive web pages. It replaces the common screenshot-plus-vision-model approach, cutting token use per step and helping models see small buttons and form validation errors.
2 engagement1 sourceredditWed, Sep 2, 2026, 10:51 PM
Products·building
BrowserClaw: dual-brain Chrome MCP for fast, low-token browser automation
BrowserClaw is a Chrome extension exposing browser automation via MCP, pairing a local micro-loop (Jev) that executes simple steps in ~250ms with a cloud reasoning model only for high-level decisions. It targets developers building agentic workflows who are frustrated by the latency, token cost, and disposable sandboxes of Playwright-style browser automation that calls remote LLMs on every click. Status per posts: building.
BrowserClaw is a Chrome extension exposing browser automation via MCP, pairing a local micro-loop (Jev) that executes simple steps in ~250ms with a cloud reasoning model only for high-level decisions. It targets developers building agentic workflows who are frustrated by the latency, token cost, and disposable sandboxes of Playwright-style browser automation that calls remote LLMs on every click. Status per posts: building.
10 engagement1 sourceredditSat, Sep 19, 2026, 01:46 PM
Social·Community
Reddit user proposes Loomwork, a portable agent format with local-first memory, to fix what MCP left unsolved
A developer posted a feedback-seeking proposal for Loomwork, a project aiming to standardize portable, memory-carrying agents the way MCP standardized tools. The post argues current agent tools (Cursor, Claude Code, Cline, Continue, Open Interpreter) lock agents to one app with no shared context, portability, or output verification. It is an early-stage idea soliciting a reality check, not a release.
A developer posted a feedback-seeking proposal for Loomwork, a project aiming to standardize portable, memory-carrying agents the way MCP standardized tools. The post argues current agent tools (Cursor, Claude Code, Cline, Continue, Open Interpreter) lock agents to one app with no shared context, portability, or output verification. It is an early-stage idea soliciting a reality check, not a release.
7 engagement1 sourceredditMon, Sep 28, 2026, 05:05 PM