Overview

MCP is how an agent gets tools. Connect Walletguide and Claude can read your assets, search your transactions and ask for a real net-worth figure, using the same routes the app uses.

Take the hosted endpoint if your client supports remote MCP servers. It signs in through your browser, the grant is revocable, and there is no key sitting in a config file. Take stdio when the client only speaks stdio, or when you want the traffic to stay on your machine.

Read-only by default

Both start with 46 read tools and no way to change anything. Writes are opt in: WALLETGUIDE_MCP_WRITES=1 for stdio, or granting walletguide:write at consent time for the hosted endpoint.

That is not only caution. The read-only catalogue is about a third of the size, roughly 12k tokens against 40k, and that is context the model spends before it has read a single number.

A grant that carries only walletguide:read is never shown a mutating tool. It is not offered one and refused on use; it does not appear.

Tools

One per route, named after it. GET /wallets/{walletId}/assets is assets_list, POST /wallets/{walletId}/transactions/search is transactions_search, and the seventeen metric slices are data_netWorthTimeSeries and friends. The endpoint reference describes all of them; a tool takes the same parameters as its route.

Deletes, bulk deletes and merges carry a destructive annotation, so a client that asks before running one will ask.

Getting good answers

Ask a data_* tool for a metric. It computes over the full history on the server and applies the rules raw rows do not carry: splits, transfers between your own accounts, offset categories. An agent that lists transactions and adds them up gets one page of them and no warning. See metrics.

Amounts are decimal strings, and negative means money out. Ids are uuids, so find one with transactions_search or a list tool rather than guessing.