agent-browser
/agent-browser
Drive the user's real, already-logged-in Chrome from the terminal or any MCP-capable agent — open pages, read them, click, type, screenshot, upload files — while organizing tabs into one group per task. Use when a task needs a browser: checking internal dashboards, filling web forms, scraping pages the user can already see, or reproducing a UI bug. Works without a debug port or a separate Chrome profile, so existing logins and sessions are preserved.
- Author
- Inquisitor-201 ↗
- Repository
- agent-browser-bridge ↗
- Licence
- MIT
- Last repository change
- 01.11.2026
Assessment and security
- Files in this package
- 6
- Scripts
- 0
- HTML comments
- 0
- URLs in SKILL.md
- 0
No signals detected. Repository changed on 01.11.2026, without references to retired models or APIs.
MIT, open: use and modify while retaining author attribution.
What it does and when to use it
Drive the user's real, already-logged-in Chrome from the terminal or any MCP-capable agent — open pages, read them, click, type, screenshot, upload files — while organizing tabs into one group per task. Use when a task needs a browser: checking internal dashboards, filling web forms, scraping pages the user can already see, or reproducing a UI bug. Works without a debug port or a separate Chrome profile, so existing logins and sessions are preserved.
Original author description
Drive the user's real, already-logged-in Chrome from the terminal or any MCP-capable agent — open pages, read them, click, type, screenshot, upload files — while organizing tabs into one group per task. Use when a task needs a browser: checking internal dashboards, filling web forms, scraping pages the user can already see, or reproducing a UI bug. Works without a debug port or a separate Chrome profile, so existing logins and sessions are preserved.
How it works
- Before you start
- The one rule that matters: one task, one tab group
- Typical loop
- Visual cursor
- Safety
- When something breaks
SKILL.md preview
---
name: agent-browser
description: Drive the user's real, already-logged-in Chrome from the terminal or any MCP-capable agent — open pages, read them, click, type, screenshot, upload files — while organizing tabs into one group per task. Use when a task needs a browser: checking internal dashboards, filling web forms, scraping pages the user can already see, or reproducing a UI bug. Works without a debug port or a separate Chrome profile, so existing logins and sessions are preserved.
---
# Agent Browser Bridge
Control the user's **real Chrome** (the window they actually use), not a headless
clone. The browser keeps its cookies, logins, extensions and password manager.
Two ways in, same engine:
- **MCP tools** (`browser_*`), for any agent that speaks MCP.
- **CLI** (`abb <tool> '[json]'`), for shells and scripts.
## Before you start
```bash
abb doctor
```
- `ok server ... extension connected` → good, continue.
- `FAIL no bridge server` → nothing is running; the MCP server normally owns the
port, so start your agent first, or run `node bin/mcp-server.mjs` yourself.
- `FAIL server is up ... but no extension is polling` → the Chrome extension is
not loaded, or its card needs a reload after an edit.
Never guess at a browser failure: `browser_status` / `abb status` tells you
whether the problem is the server, the extension, or the page.
## The one rule that matters: one task, one tab group
The user's Chrome is full of their own tabs. Do **not** work in whatever tab
happens to be active.
1. Start every task with `browser_group_begin { "title": "<short task name>" }`.
Use a name a human recognizes — `"pricing research"`, `"yuque weekly report"` —
not `"task 1"`.
2. Every later call targets that group automatically, even if the user is
looking at a different tab.
3. `browser_tab_new` joins the group automatically.
4. Finish with `browser_group_end`. Tabs stay open so the user can review them;
only pass `closeTabs: true` when they asked to clean up.
One session usually contains several small tasks. Open a new group per task
rather than reusing the previous group.
## Typical loop
```
browser_group_begin { title, url }
browser_snapshot → see what is on the page (accessibility tree)
browser_click { selector }→ move + click; the on-page cursor shows the target
browser_snapshot → confirm the page changed
browser_text → read the result
browser_group_end
```
- `snapshot` beats `screenshot` for finding elements, and costs far fewer tokens.
- `selector` accepts a CSS selector **or** the exact visible text of an element.
- `browser_eval` is the escape hatch when no tool fits; prefer specific tools.
- For long pages, `browser_scroll { pages: 1 }` then read again.
## Visual cursor
Clicks and typing show a large glowing cursor, a highlight on the target and a
ripple, so the user can see what is happening. It is drawn inside the page only:
it does not move the system pointer, it disappears on navigation, and it stays
silent while the tab is hidden or Chrome is in the background
(`browser_visual_cursor` to change that).
## Safety
- Never type passwords, one-time codes or card numbers. Ask the user to do that
part themselves; the browser is already logged in, so usually you do not need it.
- Ask before anything that spends money, sends a message, publishes content, or
deletes data. A click is cheap to make and hard to take back.
- Do not navigate away from a tab that has unsaved input without saying so.
- `browser://`, `chrome://` and extension-store pages cannot be scripted.
## When something breaks
- Page looks stale → `browser_open` again, or `browser_eval { expression: "location.href" }`.
- Element not found → `browser_snapshot` and match on the visible text instead.
- Everything times out → `abb doctor`; usually the extension needs a reload after
its files changed (`chrome://extensions` → reload arrow).
- Full guide: `docs/troubleshooting.md`.
Files in this package (6)
- .gitignore 1 KB
- LICENSE 1 KB
- README.md 8 KB
- README.zh-CN.md 9 KB
- SKILL.md 4 KB
- package.json 1 KB
Versions
| Repository | Score | Licence |
|---|---|---|
| This page · Inquisitor-201/agent-browser-bridge | 95 | MIT |
| vanducng/skills | 100 | MIT |
| mvschwarz/openrig | 100 | Apache-2.0 |
| mgiovani/cc-arsenal | 100 | MIT |
| desplega-ai/agent-swarm | 100 | MIT |
| nexu-io/open-design | 100 | Apache-2.0 |
| seaworld008/commonly-used-high-value-skills | 100 | MIT |
| slopus/happy | 100 | MIT |
| j62268781-alt/dsh-toolkit | 100 | MIT |
Database snapshot: 13.11.2026 · package version 3e84a8a66d4f (SHA-256 digest of the entire package)
Downloads in the last 30 days: 0 · ranking
Report a problem with this skill: harmful content, incorrect descriptions or an author’s removal request.
Similar skills
- 90agentchrome
Drive the user's real Chrome (signed-in profiles) through the agentchrome MCP server — open tabs in your own tab group, click, type, read pages, run JavaScript, inspect console and network, take screenshots and record video clips. Use when a task needs the user's logged-in… - 95browser-control
Drive the user's own Chromium browser (Chrome, Edge, Whale, Brave, Arc, Dia) with the `bctl` CLI — read pages, click, fill forms, borrow logged-in tabs, debug console/network — inside a separate Agent Window. Use when a task needs a real browser with the user's existing logins,… - 95browser-control
Drive the user's real Chrome browser from the shell - read pages, click, type, scroll, screenshot - inside a dedicated Agent tab group, using their actual logins. Use when the user asks you to look at or act on a web page, check something in a site they are signed into, fill or… - 100use-browser
主浏览器技能:通过 browser_session / browser_tab / browser_page / browser_client 四个工具操作本机 Chrome(需 Ethan Browser 扩展)。嵌入本机真实 Cookie,能操作用户已登录的浏览器、接管当前 tab、做 snapshot/click/fill/screenshot/eval。支持多个浏览器同时连接,每个有自己的名字,agent 按需选择操作哪个端。当用户要求操作浏览器、网页自动化、点击输入、或需要用到本机真实 Chrome(含登录态)时触发。兜底用 agent-browser。 - 100chrome-use
Operate web pages in the user's own Chrome on their connected computer through the chrome-use command-line tool — open pages, read them as structured text, click, fill forms, upload files, take screenshots, and test web apps, using the sites the user is already signed in to. Use… - 95agent-webbridge
Drive the user's REAL Chrome — multiple profiles with their LIVE logins, and MULTIPLE TABS PER PROFILE, all IN PARALLEL — through agent-webbridge. Clean-room, open-source (MIT), no account, no telemetry. Built for Claude Code, Google Gemini, Qwen, Cursor, Windsurf, and DeepSeek.…