| name | browser-now |
| description | Read the user's LIVE browser state — the tabs open right now, which one is focused, what's playing audio, the recent navigation flow, and DOM text from an open (possibly logged-in) tab. Use whenever the request is grounded in the present: "this page", "what I'm looking at", "the tab about X", "my current research", "do I have Y open", "save my research state", or before WebFetch when the user points at something they already have open. Powered by the tabtopia extension + its standalone MCP (tools get_context, search, capture_context, get_tab_content). Not for closed history — that's a separate browsing-history tool. |
| user-invocable | true |
browser-now — live browser state
tabtopia (a Chrome extension) pushes a live snapshot of the browser to a small
standalone MCP server. That snapshot is the ground truth for what the user is
doing right now: open tabs, which is focused, what's playing audio, and the
recent navigation flow. Every tab carries real engagement — seconds
actually spent reading — so ranking surfaces the tab they were working in over
the six they opened and abandoned.
This is a separate, standalone server that ships in this repo (mcp/). It has
no external dependencies beyond ws — no embeddings, no history corpus. For
closed history ("that article last week") use a browsing-history tool instead.
The four tools
| Tool | Use when | Notes |
|---|
get_context | Start of a conversation, or any "based on what I'm doing / this / my current research". | Returns focused tab, engagement-ranked open tabs, audio, recent flow. Always states snapshot age. |
search | "which tab is about X", "do I have Y open", "where was I reading about Z". | scope: tabs (default) · activity (recent nav) · snapshots (saved captures). Empty query = enumerate the scope. Tabs ranked by dwell × recency. |
capture_context | "save my research state", "bookmark what I'm doing" — before a context switch. | The only write. Persists a named snapshot server-side; find it later via search(scope:'snapshots'). Never drives the browser. |
get_tab_content | Read an open page a plain fetch can't see (logged-in LinkedIn/GitHub/an app), after a tab is surfaced. | Reads what's on screen — no navigate/reload. Fails gracefully if the tab isn't open. |
How to use it well
- Reach for
get_context early when the request is deictic — "this",
"here", "what I'm looking at". Don't ask for a URL they already have open.
- Prefer
get_tab_content over WebFetch for a page they currently have open,
especially anything behind auth. WebFetch gets the logged-out version.
- Trust the ranking. The top tab in
get_context / search(tabs) is the
one they actually engaged with. Lead with it.
- Every response states snapshot age. If it says "4m ago / stale", say so —
Chrome may be closed or the service worker torn down. Treat stale data as
approximate, don't pretend it's live.
- Degrade, don't flail. If a tool returns
unavailable / no_data, the
bridge daemon or extension isn't running (see Setup). Say so plainly rather
than retrying in a loop.
get_tab_content is read-only. None of these tools open, close, focus, or
navigate tabs — that's deliberately out of scope. Don't promise browser-driving.
Setup (once)
Two processes, both from this repo's mcp/ directory (npm install first):
-
Bridge daemon (always-on; the extension pushes to it, holds the snapshot):
node mcp/bridge-daemon.js
WS on 127.0.0.1:8892 (Origin-allowlisted to the tabtopia extension id),
read API on 127.0.0.1:8893. Best run under launchd (plist ships at
mcp/com.andyed.tabtopia-mcp.plist) so it survives reboots.
-
MCP server (stateless, spawned per Claude session — register once):
claude mcp add tabtopia -- node "$(pwd)/mcp/server.js"
The tabtopia extension must be loaded in Chrome — it's what pushes the snapshot.
If tools report "no live snapshot", check Chrome is open with the extension
enabled and the daemon is running (curl -s localhost:8893/health). Full
details in mcp/README.md.
