The terminal, inside your agents, notebooks and spreadsheets.
Three ways in: an MCP server for Claude and your own AI agents, a Python SDK and API for your own systems, and Excel and Google Sheets. Each carries the same rights, point-in-time and audit checks as the desktop, and every call is a read.
MCP server
Inside Claude and your own agents.
Teams use it inside Claude or their own AI agents: ten read-only verbs, called as tools.
How it works ->API and Python SDK
Into your systems and notebooks.
Fund engineers pull thesis checks and exposure into their own systems and notebooks.
How it works ->Excel and Sheets
In the spreadsheet you already use.
One formula, =QT(), in Microsoft Excel, Google Sheets and the terminal’s own Workspace.
Inside Claude and your own agents.
The MCP server exposes the terminal as ten tools an assistant can call. Claude, or an agent your team builds, can resolve a company, read its fields with their provenance, open the filings a figure came from, score a thesis, trace exposure through your book, and draft a plan for a person to approve.
{
"mcpServers": {
"quantify": {
"command": "<terminal>/.venv/bin/python",
"args": ["<terminal>/tools/quantify_mcp.py"]
}
}
}
The server runs beside the terminal on your own machine, and we set it up with you. Next, the same server over HTTPS, so a team’s agents can reach it without a local install.
- resolve_company a ticker, ISIN or name to one canonical id, with the identifiers and fields known for it.
- get_facts canonical fields with the source, the period, and when each first became knowable.
- get_evidence the filings, most material first, each with the stored sentence and its offsets.
- thesis_health a thesis scored, with every assumption’s status.
- look_through_exposure the routes from a company to what you hold, each step named.
- model_diff what a recompile would propose against your committed model, cell by cell.
- red_team the seven seats, each arguing from one evidence path.
- hedge_architect what protecting a position against a fall you name would cost.
- propose_plan and review_plan a plan made only of declared steps, for a person to approve. Neither executes one.
Into your systems and notebooks.
Fund engineers pull thesis checks and exposure into their own systems and notebooks. The Python SDK calls the same server, so a notebook and an agent get the same answers: a typed method for every verb, exceptions instead of error envelopes, and a NotYet answer that cannot be mistaken for data.
from quantify_sdk import QuantifyClient with QuantifyClient.local() as qt: nvda = qt.resolve_company("NVDA")["entity_id"] health = qt.thesis_health(nvda) routes = qt.look_through_exposure(nvda)
The same verbs over HTTPS, with a key per desk, for systems that do not run Python beside the terminal. The rights, point-in-time and audit checks belong to the server, so they come with it unchanged.
If your team needs it, tell us which calls come first.
In the spreadsheet you already use.
One formula, the same everywhere. The desktop terminal generates the connector for each spreadsheet, with the formulas and the connector address already in it.
- Connect Excel generates a VBA module. Import it into Microsoft Excel once, and formulas in your workbook call Quantify Terminal.
- Google Sheets copies an Apps Script connector. Paste it under Extensions › Apps Script, and the sheet does the same.
- The Workspace spreadsheet runs the same formulas inside the terminal, with CSV, JSON and XLSX import and export.
=QT("AAPL","price")
The connector is yours: you deploy it once, from the folder the terminal hands you, and the desk that deployed it holds the key. It returns market data and portfolio figures; for fields with their provenance, the door is the MCP server.
What an agent cannot do.
The same rules hold for the MCP server, the SDK and the spreadsheet connectors, because they are the server’s rules, not the door’s.
- Write anything.Every verb is a read; nothing is saved, refreshed or spent.
- Place an order.Nothing here prepares one, let alone sends it.
- See data it is not licensed to see.A value that may be shown on your screen but not placed in someone else’s prompt is withheld, named and counted. See how we handle data rights.
- Read the past without point-in-time.An as of read on a plan without it is refused, with the same sentence the desktop shows.
- Act without an audit row.Every rights decision is recorded with the account that asked and the call it belonged to, with the reason.