WhereItGoes MCP server

Ask ChatGPT, Claude or Cursor about your own crypto portfolio: real holdings across every exchange and on-chain wallet you track, full movement history, FIFO profit and loss, fees, income and risk. No API key to paste, free on every account, revocable in one click.

https://api.whereitgoes.money/api/mcp

• Transport: Streamable HTTP (stateless, JSON responses)
• Auth: OAuth 2.1, authorization code + PKCE (S256). Dynamic client registration (RFC 7591) and Client ID Metadata Documents are both supported.
• Registry name: money.whereitgoes/portfolio in the official MCP Registry
• Setup guide: whereitgoes.money/connect-ai

This repository is the public listing and documentation for a hosted (remote) MCP server. There is nothing to install or run locally.

What it does

WhereItGoes is a read-only crypto portfolio tracker. You connect exchanges with read-only API keys and wallets by public address; it rebuilds your trade and transfer history and computes portfolio value, P&L and history. The MCP server exposes those same computed numbers to an AI client, so the assistant reads your real data instead of guessing from a screenshot.

Every figure is already computed and USD-valued by the app. The server instructions tell the model to report what the tools return rather than recompute it.

Tools

15 tools: 13 reads and two propose-only writes. Each read is a GET against a fixed allowlist of the app's own endpoints. Tools that cover several related answers take a required view argument, and the answer comes back nested under the key named after that view.

Every portfolio tool takes an optional profile argument, so one question can read a named portfolio ("my trading account") without changing what the app shows. Windowed views also take period (last_24h, last_7d, last_30d, last_90d, ytd, all); omitting it means all time. A view refuses an argument it does not take, rather than ignoring it.

Tool reference

Every tool declares an outputSchema and returns structuredContent alongside a JSON text block, so a client can read typed fields (units, currency and ordering are described in the schema) instead of parsing prose. A list longer than 200 rows is cut, and the answer then carries a truncated field and a <field>_note saying how many rows were kept and which.

Version 1.2.0 renamed the 1.1.x tools (26 became 15, counting the new propose_alert_change). If you saved a prompt or config that names an old tool: get_portfolio_snapshot, get_dashboard -> get_portfolio; get_pnl_overview, get_portfolio_history, get_pnl_by_symbol -> get_performance; get_risk_metrics, get_portfolio_score -> get_risk; get_capital_flow, get_deposits, get_transfers, get_fiat_stats -> get_money_flows; get_income_by_source, get_income_timeseries -> get_income; get_asset_trade_stats, get_activity_calendar -> get_trading_activity; get_venues, get_wallets, get_sync_status -> get_connections.

Every read tool is annotated readOnlyHint: true, destructiveHint: false, idempotentHint: true, openWorldHint: false. The two write tools are annotated readOnlyHint: false, destructiveHint: true, so a cautious client asks before calling them, even though neither applies anything on the call.

On a free account both get_risk views answer with an upgrade_required result rather than an error, so the assistant can say what it would take instead of retrying.

There are also two optional, propose-only tools, available only if you grant the separate portfolio:write scope when connecting: set_target_allocation drafts a target plan, and propose_alert_change drafts a price or value alert (create, edit or delete, including whether it also emails you). Neither changes anything when called: each parks a proposal and returns a link, and the change lands only after you open that link in your own browser, read a server-rendered before/after diff, and click Apply. Alerts are a Pro feature, so on a free account propose_alert_change answers upgrade_required instead of drafting anything, and a draft is checked by the same validator the app's alert form uses. See Security model.

Connect

You need a WhereItGoes account with at least one exchange or wallet connected, because the connector reads the portfolio you already have. Start here if you have not.

Claude

1. Open Settings > Connectors > Add custom connector.
2. Paste https://api.whereitgoes.money/api/mcp.
3. A WhereItGoes consent screen opens in your browser. Approve it.

ChatGPT

1. Open Settings > Apps & Connectors > Advanced and turn on developer mode.
2. Create a connector with the URL https://api.whereitgoes.money/api/mcp and OAuth as the authentication method.
3. Approve the WhereItGoes consent screen that opens in your browser.

Cursor and other MCP clients

Anything that speaks MCP over Streamable HTTP and handles OAuth works. For clients configured with JSON:

{
  "mcpServers": {
    "whereitgoes": {
      "url": "https://api.whereitgoes.money/api/mcp"
    }
  }
}

To look at the raw tool list first, run npx @modelcontextprotocol/inspector and point it at the URL.

Then ask things like "what do I hold", "how much did I deposit last month", "what am I paying in fees", or "am I actually up this year".

Security model

Read-only by construction. The read executor hard-codes GET against an allowlist of app endpoints. There is no trading, no withdrawal, no transfer, no way to connect or disconnect a source, and no way to read your exchange API keys. Those paths do not exist in the connector; they are not permissions that happen to be switched off.

A connector token can never become an app session. MCP access tokens are HMAC-signed in a separate domain from app session tokens, so each type is unverifiable as the other. When a tool runs, the server mints a fresh short-lived session for the same user internally and re-dispatches through the app's own auth and plan checks; the MCP token itself is useless anywhere except the MCP endpoint. Tests assert the separation in both directions.

Strict OAuth. PKCE S256 only; redirect_uri matched exactly with no prefix or wildcard; an unknown client or unregistered redirect gets a dead end, never a redirect; authorization codes are single use and live 60 seconds; access tokens are short-lived and bound to the MCP endpoint as their audience.

Writes need a human, every time. The optional portfolio:write scope is granted only when explicitly requested and approved on the consent screen; an existing read-only connection cannot acquire it. Even with it, a write tool only proposes. The change applies when you open the link in your own browser, under your own session, on whereitgoes.money, and approve a before/after diff rendered by the server from the stored payload (never from text the model wrote). Proposals are single use and expire after 10 minutes. This keeps a prompt-injected tool call from writing anything without you seeing it.

Tenant scoping. Every call is re-checked against the account and portfolio it belongs to. Picking a portfolio per call never changes which portfolio the app itself has active.

Revoking access

• One connector: in the app, Settings > Connections lists every connected app with when it was approved and when it last read anything. Disconnect, and the very next call fails; the grant is checked on every request, not at the end of a session.
• Everything at once: signing out of every device revokes all connectors.
• Account deletion removes every grant with the account.

Pricing

The connector is free on every account, including the free tier, and uses no AI credit of ours: your assistant does the talking, WhereItGoes only answers for the data. The Pro tools stay Pro, as they are in the app.

Support

• Setup guide: whereitgoes.money/connect-ai
• Security overview: whereitgoes.money/security
• Email: [email protected]

Nothing returned by this server is investment advice.

License

The contents of this repository (documentation and listing metadata) are released under the MIT License. The WhereItGoes service itself is a hosted product and is not open source.