Use Quantra from your AI agent

Quantra has an MCP server at mcp.quantra.io. Any agent that speaks MCP can use it: Claude, Claude Desktop, Claude Code, Cursor, and others. Once connected, the agent can build curves, price swaps, bonds, swaptions, CDS, options and inflation products on the Quantra engine, and show you how each number was computed. Every number comes from the engine. The server computes nothing itself.

The server

One URL, streamable HTTP, no account, no key:

https://mcp.quantra.io/mcp

Add it wherever your client configures MCP servers or connectors. The agent will see 47 tools: calendars, curve construction from quotes or from pasted discount factors and zero rates, one pricing tool per product, the engine's own example requests, DV01 and key rate ladders, a tool that changes one field of a request and reprices, and the methodology pages.

Instructions for your agent

Every client has a place for standing instructions: a project in claude.ai, custom instructions in Claude Desktop, a CLAUDE.md for Claude Code, rules in Cursor. Put this there. It matters more than anything the server says, because it decides how the agent treats market data.

You are a pricing assistant for business users, backed by the Quantra pricing engine
through the Quantra MCP server.

Rules that override everything else:
1. You never invent, estimate, recall or type market data of any kind: no curve quotes,
   discount factors, zero rates, volatilities or fixings. Not as placeholders, not as a
   test, not labelled approximate.
2. When the user shows a trade (screenshot, ticket, description) and has not given market
   data, you answer only with: whether it can be priced, exactly which market data is
   missing and in what simple form to paste it (for example "the discount factors from the
   curve screen as date, value rows"), and the conventions you will assume. Then you stop.
   You do not call any pricing or curve tool.
3. Market data exists only once the user has pasted it or attached it in this conversation.
   Then you build the curve from it, price, and put the result next to the number on the
   user's screen.
4. Every number you report comes from the engine. When asked how something is computed or
   why it differs from another system, use the engine's methodology pages and test each
   hypothesis by repricing with one change, rather than asserting a cause.
5. Speak in trade and market terms. Do not mention tools, presets, sessions or request
   bodies unless asked how it works.

What a session looks like

Paste a screenshot of a trade, for example a swap pricing screen from another system, and ask whether Quantra can price it and what it needs. The agent reads the trade off the screen, says which market data is missing and how to paste it, lists the conventions it will assume, and stops. Paste the curve screen. The agent builds the curve from it, checks its discount factors against yours, prices the trade, and puts its number next to the one on your screen.

Then ask the questions that matter:

• Why does the number differ? The agent changes one convention at a time and reprices, and reports what moved the number and by how much.
• How is DV01 computed? What is bumped, by how much, and where is that in the code?
• Show me the engine code that computes the NPV, the fair rate and the leg BPS, with links.
• Show me every cash flow on both legs with its discount factor.
• Plot the discount factors, zero rates and forward rates of the curve you built.
• Give me the exact request you sent, so I can run it myself.

The answers cite the engine source at the pinned version, file and line, with a link to the line on GitHub. The engine is open source, so you can check every claim.

Rules the server enforces

• Market data for your trade comes from you. Every curve and pricing tool requires a declaration of where the data came from: pasted by you, read from a file or screenshot you attached, an engine example you asked to run, or a market built earlier in the session. There is no value for estimated or recalled data.
• Every result carries the exact request sent to the engine and the engine's response verbatim. Engine errors are returned as they are.
• Every convention a tool applies is listed in the result with its source.
• The server has no access to licensed market data. Curves from Bloomberg, Refinitiv or any vendor must be pasted.

Client setup

claude.ai and Claude Desktop

Claude Code

claude mcp add --transport http quantra https://mcp.quantra.io/mcp

Put the instructions in the project's CLAUDE.md. Start a new session and check with /mcp.

Cursor and other clients

Add a server of type streamable HTTP with the URL above in the client's MCP settings, and put the instructions in its rules or custom instructions. Any client that supports remote MCP servers works the same way.

Running it yourself

The server is a small Python package that talks to any Quantra engine over its JSON API. Run the engine and point the server at it with QUANTRA_ENGINE_URL. The repository has the install steps, a Dockerfile and a compose file with both services: github.com/joseprupi/quantra-mcp. It needs engine 0.7.0 or later.