Browd: AI Browser Agent

Open-source браузерный AI-агент. Веб-задачи в side panel Chrome: извлечение данных, заполнение форм и не только.

Browd is an open-source browser agent — a self-hosted alternative to OpenAI Operator and Claude Computer-use. It runs in your own Chromium side panel, drives your real browser session, and works with whatever model you bring. Apache-2.0. No subscription, no cloud VM, no credential hand-off. WHY BROWD - Open source, end-to-end. Apache-2.0. Read the code, fork it, audit exactly how your browser is being driven. No black box. - Runs in your real session. GitHub, Gmail, dashboards — Browd sees them the way you do. Nothing to re-log-into, no cookies to paste, no remote VM. - Bring your own model. Any OpenAI-compatible provider; OpenRouter routes Anthropic, Google, Meta and others through one key. Planner, Navigator and Judge roles set independently. HOW IT WORKS You type a task in plain language — "open the LM Arena leaderboard and tell me the top three open-source models", "find the arXiv paper on X and summarize the abstract", "fill this form with my details". Browd plans the task, then executes it step by step in a dedicated browser tab, adapting as the page reacts. It runs a Plan-and-Execute loop: a planner breaks the task into subgoals, an executor works through them with browser tools (click, type, scroll, navigate, screenshot, web search), and a replanner re-checks the plan after every step — so when a page differs from what was expected, the plan is rewritten rather than blindly followed. - Tab isolation. The agent works in its own tab, marked [Browd]. Your other tabs are read-only metadata to it — it won't touch them unless you explicitly hand one over. - Full trace. Every step is logged in the side panel — the tool call, the result, a screenshot thumbnail so you see exactly what it did. - Untrusted-content wrapping. Page text, search results and fetched content are wrapped before reaching the model, to reduce prompt-injection from third-party pages. HONEST LIMITS - Token usage is high on multi-step tasks. Browd shows a live token counter; watch it. - Some sites with hard anti-bot protection block automated clicks. Browd detects the loop and asks you to click the blocked element yourself, then continues. - It's a browser agent for concrete tasks on concrete pages — not a research tool or a scraper. Forked from Nanobrowser (Apache-2.0) and rebuilt around a LangGraph.js Plan-and-Execute runtime. Source, issues and changelog: https://github.com/wyddy7/browd