chartremotely-mcp

ChartRemotely — operator

A Tollbooth-DPYC Operator MCP. It meters and relays chart commands to agents running on patrons' own machines, so a security can be put on a wall monitor anywhere by voice or by tool call.

The operator never touches a broker API and never drives a GUI itself. It pairs agents, bills per request, and relays a narrow vocabulary.

Why this exists

Desktop charting applications have no automation surface. thinkorswim in particular has no API, no URL scheme, and nothing addressable — so "put Palantir on the screen in the other building" is not a thing anyone can currently buy.

The control plane is the product. Chart analysis is abundant; remote input automation with a voice front end is not.

Web app

frontend/ is the site at https://chartremotely.tollbooth-dpyc.com: a welcome page, npub sign-in, a carousel of your screens with a live snapshot on request, and a profile for your balance and paired displays. It is built on @tollbooth-dpyc/web, the shared browser SDK, and deploys to Cloudflare Pages through the fleet's shared workflow on every push to main that touches frontend/.

cd frontend && npm install && npm run dev   # tests run as part of npm run build

Tools

Tool	Price	What it does
get_shortcut	free	Returns the Apple Shortcut, ready to import
pair_agent	free	Binds an agent to the caller's npub
agent_status	free	Which displays are paired, and are they connected
forget_display	free	Removes a display and the secret it signed in with
show_chart	metered	Put a security on a paired display
read_chart	metered	What a display is currently showing
snapshot_display	metered	A JPEG of the display's chart window, right now - also kept as a capture
latest_snapshot	metered	A kept capture: the newest, a symbol's newest, or one by capture id

Every tool that takes an npub needs a proof of that npub (request_npub_proof, then receive_npub_proof); get_shortcut hands out a public file and does not.

A metered call to a display that never answers - offline, or not running the agent - is refunded, and so is one naming a display you do not have. A display keeps captures by symbol: the newest 6 of each of its last 20 symbols, sealed with the vault cipher, each for four hours. One lands after every chart change the agent pushes, and one with every snapshot_display (filed under the symbol the display says it shows, or "Chart"). The AAD binds each capture to its display, its symbol and its capture id. agent_status lists them per display - symbols by newest capture, each with its captures newest first - and latest_snapshot(capture=...) shows one exactly.

A picture reply leads with one line of text naming what is known about it - display, symbol, scale, capture time, e.g. Mac mini · PLTR · half · captured 2026-09-24 22:10 UTC - then the JPEG. Anything unknown is left out. The line survives when a client drops images to compact a long conversation, and it is all a client that cannot render images gets.

display may be said loosely: "mini mac", "mini" or "mack meeny" all find "Mac mini". A name that could mean several displays is answered with them (see PROTOCOL.md for the rules).

What it deliberately cannot do

The vocabulary relayed to an agent is fixed, and the agent enforces it independently — see vocab.py in the agent repository. Nothing in this operator can reach an order ticket, an account balance, or a position. Widening that surface would require changes in both repositories, in public.

Related

• ChartRemotely-agent — runs on the patron's Mac
• chartremotely.tollbooth-dpyc.com — the web app, built from frontend/ here

License

Apache-2.0