@tekmidian/browsr

Browsr — browser control for agents

Browsr lets any MCP-capable client (Claude Code or anything else that speaks MCP) drive the user's real running Chrome — the one with their logged-in sessions, cookies and extensions. It lists tabs, snapshots a page's DOM as an accessibility tree, clicks, types, reads text, evaluates JavaScript, screenshots and reads console output.

• Real Chrome, no --remote-debugging-port, nothing headless, no Playwright
• No debugger attachment — no "started debugging this browser" banner
• Provider independent — plain stdio MCP, no coupling to any one vendor
• The browser-domain sibling of Clickr (screen/native): anything in a web page → Browsr; screen and native apps → Clickr

MCP client ─stdio─ browsr MCP server ─ws://127.0.0.1:8757─ native host ─native messaging─ extension ─tabs+scripting─ Chrome

Quick install

Requires node ≥ 22 (the MCP server uses the global WebSocket client).

git clone <this repo> && cd browsr
npm install
npm run build

# 1. Load the extension: chrome://extensions → Developer mode → Load unpacked
#    → select the extension/ folder → copy its 32-char ID
# 2. Register host manifest + MCP server:
node dist/cli.js install --extension-id <the-32-char-id>
# 3. Restart Chrome (native host manifests are read only at startup —
#    reloading the extension is NOT enough), then restart Claude Code

Full walkthrough and troubleshooting: docs/browsr.md.

Tools

Tool	What it does
tabs_list	List tabs: id, title, url, active
tab_open {url, active?}	Open a tab
tab_select {tab}	Bring a tab to the front
tab_close {tab}	Close a tab
dom_snapshot {tab}	A11y-tree YAML of the page, with [ref=sN] ids
dom_click {tab, ref}	Click an element (scrolls into view first)
dom_type {tab, ref, text}	Focus an element and type into it
page_text {tab}	Read document.body.innerText
eval_js {tab, code}	Evaluate JavaScript (awaits promises)
tab_screenshot {tab}	PNG screenshot of the visible tab, base64
console_logs {tab}	Console output captured since the last snapshot

Take a dom_snapshot before clicking or typing: refs (s1, s2, …) are assigned per snapshot in DOM order and stay valid until the next snapshot or navigation. Browsing actuation is Browsr's job — there is no control gate in v1; the operator asking you to browse is the authorization.

CLI

browsr install --extension-id <id>   write the Chrome native-host manifest,
                                     register the MCP server, print next steps
browsr uninstall                     remove both registrations
browsr status                        show what is registered and where

Configuration: BROWSR_PORT overrides the bridge port (default 8757), BROWSR_LOG overrides the host log (default /tmp/browsr.log).