Documentation

// Protocol

What is MCP?

The Model Context Protocol (MCP) is an open standard that lets AI applications connect to external tools and data in a uniform way. Think of it as a standard port for AI: instead of every assistant building a bespoke integration, an assistant that "speaks MCP" can plug into any MCP server.

To learn more about the protocol, see modelcontextprotocol.io.

// Overview

UPX Secure Financial MCP

UPX Secure Financial MCP gives your AI assistant secure, read-only access to your financial data so it can answer questions like "what's my balance?", "how much did I spend last month?", "when is my card bill due?", or "how is my portfolio doing?".

// Use cases

What you can build

Use UPX Secure Financial MCP as the consented financial data layer for agents, copilots, and automations that need to query real accounts safely.

Developer positioning: infrastructure for financial agents built on real, consented, regulated, read-only banking data.

// Security model

Auditable by design.

UPX Secure Financial MCP is designed so developers can inspect the access surface, not just trust a marketing promise.

// Getting started

Setup in under 5 minutes

Two steps: a one-time account setup in the UPX portal, then pointing your AI host at the endpoint.

1. One-time setup (UPX portal)

2. Endpoint

Point your MCP host at the single UPX MCP endpoint:

https://mcp.upx.com

Path note: Some hosts expect the base URL https://mcp.upx.com/ (e.g. ChatGPT). Others expect the /mcp path, https://mcp.upx.com/mcp (e.g. Claude, Cursor). You always use the same host, there are no per-product subdomains.

Connecting popular hosts

Claude

Settings → Connectors → Add custom connector → paste the URL → Connect, then complete the sign-in/authorize window.

https://mcp.upx.com/mcp

ChatGPT

Settings → Connectors → enable MCP mode → Create/Add → paste the URL → Authentication: OAuth → confirm.

https://mcp.upx.com/

Cursor / other hosts

Add a remote MCP server pointing at the endpoint with your bearer token.

https://mcp.upx.com/mcp

Your first prompts

• > List my bank connections and tell me which ones need attention.
• > What are my account balances?
• > How much did I spend last month, broken down by category?
• > When is my next credit-card bill due?

// Security

Authentication & scopes

Access is governed by OAuth read scopes. A token only grants the scopes you consented to, and a tool runs only if your token carries the scope it requires.

Available scopes

Recommended default set for a general finance assistant (covers balances, spending, and card bills without triggering step-up MFA):

finance:read:accounts finance:read:transactions finance:read:credit

// Reference

Tools (10)

Start with finance_connections_list to discover connection_ids, then pass a connection_id to the other tools to scope a read to a single institution. Omit to fan out across all of your connections.

// Conventions

Response conventions

All tools share a consistent response envelope and a few common types, documented once here.

The Money object

Every monetary value is an object, never a bare number:

{
  "amount":        2901.99,
  "currency_code": "BRL",
  "formatted":     "R$ 2,901.99"
}

• • amount, positive magnitude. Direction carried by a separate direction.
• • formatted, render this; don't recompute it.
• • null means unknown. amount: 0 is a real zero. Never sum across currencies.

Standard response envelope

Pagination

Paginated tools include a pagination object:

{
  "has_more":       true,
  "next_cursor":    "eyJ…",
  "page_size":      50,
  "returned_count": 50
}

Follow next_cursor until has_more: false. page_size defaults to 50, max 500. Replay a cursor with identical filters, changing scope mid-sequence is rejected.

Billable tools & usage

On a successful billable call the response includes:

{
  "remaining":  42,
  "limit":      50,
  "period_end": "2026-06-01T00:00:00Z"
}

finance_connections_list and finance_suggestions_list are free and carry no usage charge.

// Cache & data

Freshness & caching

Because the runtime doesn't durably store your data, every response tells you how fresh it is via as_of and freshness_state.

Data is cached for at most 1 hour and refreshed in the background. cache_age_seconds tells you how old a cached value is; next_refresh_hint_seconds hints when fresher data is expected.

// Error handling

Errors & limits

When a call can't be served, the tool returns a structured error envelope designed so an assistant can explain the problem and offer a concrete next step.

Error envelope shape

{
  "code":              "entitlement.denied.quota_exceeded",
  "message":           "You have reached your plan's usage limit.",
  "category":          "ENTITLEMENT_DENIED",
  "request_id":        "req_01HV…",
  "retry_after":       3600,
  "support_reference": "REQ-01HV-7Q3M-2X8N",
  "next_steps": [
    { "action": "wait_quota_reset", "period_end": "2026-06-01T00:00:00Z" },
    { "action": "upgrade_plan", "url": "https://…" }
  ]
}

Error categories

// Catalog

Machine-readable catalog

The runtime publishes its own catalog. Treat these as the canonical source of truth, reconcile this page against them when tools change.

Good to know

// Developer resources

Start with the inspectable surface.

This page combines human documentation with machine-readable resources so product, security, and engineering teams can review the integration.