Your AI reads the signed-in tabs you have open and everything behind them — your history, downloads, the pages you closed — searched by meaning, not keywords, in plain language. Ask it "what did I work on today?" or "pull up that article I read last week" and it actually can. And when you ask, it can click, type, and watch a tab too — even behind a strict CSP.
Your browser is the highest-fidelity record of your work and attention — every doc, dashboard, article and design you've touched. Yet your AI can't see any of it, and Chrome itself can barely search it.
You talk to your AI in plain language; it runs the right Handle tools for you. A few things people ask:
Most "AI that browses" tools open a fresh, robot-controlled Chrome signed into nothing — a stranger to your work. Handle reads the browser you're already logged into. Programs are walled off from the browser by design, so it keeps several ways in — a Chrome add-on, the built-in macOS bridge, or an opt-in developer channel — and tries the smoothest first. Because there's always a backup, a read almost never just fails.
Handle hands your AI only the part you'd actually read — clean Markdown, not the page's markup and scripts. Roughly 95%+ less to chew on: cheaper, faster, and less for the model to get lost in.
Behind one plain "read the tab," Handle tries the cleanest route first and drops to the next only if it must — so you get usable content, or a clear reason why not.
Handle exposes 31 tools over MCP, so any agent can use them as first-class functions. You rarely call these by name — your AI picks the right one — but here's the full surface.
read_tabpage text — HTML, Google/Office export, or screenshot, by type
active_tabthe tab you're looking at right now
find_tabresolve "the figma one" to a tab
grep_tabssearch the on-page content of every open tab
list_tabsevery open tab, with groups + stale flags
crawlfollow a collection's links → each page's full text, combined
screenshot_tabcapture a tab as image(s) to read with vision
grab_clipboardread what you copied — the reliable floor
ask_tabsanswer a question across your open tabs, cited
search_webfull-text search every page you've read through Handle
save_tabscapture tabs into a dated research bundle
recall_bundleload a saved bundle back as context
list_bundleslist your saved research
historyquery your history by meaning (+ search terms)
closedrecently-closed tabs, real titles → reopen
bookmarkssearch your bookmarks
downloadsrecent downloads → local file paths
journeysChrome's own topical clusters of your browsing
most_visitedthe sites you live in
click_tabclick an element — by selector, text, or intent (CSP-safe)
type_tabtype into a field, optionally submit (CSP-safe)
meetingscrape a Zoom recording's transcript → JSON + Markdown
open_tabfocus a tab, or open any URL
close_tabclose tabs for real
group_tabmove a tab into a group
note_tabattach a note to a tab
pin_tabpin / unpin a tab
refresh_tabsre-scan Chrome into Handle
consoleconsole logs + errors for a tab (via the extension)
ext_statusis the Chrome extension connected
auditthe append-only trail of what Handle did in your session
read_tab --cdpread via DevTools Protocol when you've opted in
Connect it in one line: claude mcp add handle -- node ~/claude-apps/handle/mcp/server.js
You'll need macOS, Google Chrome, an AI assistant you already use (like Claude Code), and Node + Python 3 for the local bridge (most dev machines already have both). It's free and open source, runs entirely on your own computer — no account, no API key. It reads your tabs; it never sends them anywhere, and it can't touch anyone else's.
git clone https://github.com/abhitsian/handle.git ~/claude-apps/handle
One line, and your assistant can read your tabs:
claude mcp add handle -- node ~/claude-apps/handle/mcp/server.js
a. The first read pops a macOS prompt to let it control Chrome — click OK.
b. In Chrome, turn on View → Developer → Allow JavaScript from Apple Events. This one's a menu toggle, not a popup, and it's off by default — it's what lets Handle read the actual text on a page. Skip it and tabs list fine but reads come back empty.
c. Only for screenshots of designs and PDFs (like "what does this Figma show?"): grant Screen Recording to your terminal in System Settings → Privacy & Security. Text reads don't need it.
"Summarize the doc I'm looking at." "What does the figma show?" "Answer this from my open tabs." That's the whole thing.
Optional: there's a small Chrome companion that makes it sturdier — your real tab groups, cleaner screenshots, and page errors — installed once, no browser restart, and it talks only to your own computer. Most people never need it; the steps above work on their own.