MCP for AI agents

Connect Claude, ChatGPT, or any MCP client to your own transactions with one URL.

Point Claude, ChatGPT, Cursor, or any Model Context Protocol client at your own transactions and ask your own questions of them: which categories are drifting, what changed since March, where the savings are actually hiding. The agent can write rules too, so your categorization can improve through conversation instead of form-filling.

There are two servers. Use the hosted one unless you have a specific reason not to.

The hosted server

One URL, nothing to install:

https://mcp.taz.randomfact.com

Claude Code

claude mcp add --transport http --scope user transaction-analyzer https://mcp.taz.randomfact.com

--transport http is what makes this a hosted connection rather than a spawned local binary, and --scope user registers it once for every project instead of only the current directory. Then run /mcp once and sign in through the browser tab that opens.

Claude Desktop and claude.ai

Add the same URL under Settings → Connectors as a custom connector.

Anything else

Any client that speaks Streamable HTTP MCP and can do an OAuth flow will work with that URL.

Signing in

The first tool call opens a browser tab and asks you to sign in with Google — the same account you use in the app. There is no API key to generate, paste, or rotate.

The agent acts as you, and access can be revoked without changing anything on your side: every request re-checks its token, so expiry and revocation take effect immediately rather than at the end of a session.

What an agent can and cannot reach

Only your personal workspace. A request authorized over OAuth is pinned to it and cannot be pointed at a shared workspace, whatever it asks for. An agent connected to your account can't read a household's books. See Workspaces and households.

The tools are a curated list, not everything the API can do. Reading transactions, summarizing them, managing categories and rules, testing a rule, importing and exporting, overriding a category, deleting and undeleting transactions, running a duplicate scan, and the uncategorized report are all available.

Two things are deliberately left out:

  • Uploading a statement. It's a file upload; hand your statements to the app.
  • Confirming a CSV column mapping. The mapping flow is a human-in-the-loop confirmation by design — "a proposed mapping is never applied without a person agreeing to it" is the property that makes proposing one defensible in the first place, and exposing confirm-a-mapping as a tool would hand exactly that away. A client that inspects the API spec will notice the mapping endpoints have no tools; that's expected, not a packaging bug.

The local package

@randomfact/taz-mcp on npm is a stdio server for clients that spawn a local binary.

In Claude Code:

claude mcp add transaction-analyzer --scope user -- npx -y @randomfact/taz-mcp

In Claude Desktop, add it to claude_desktop_config.json with "command": "npx" and "args": ["-y", "@randomfact/taz-mcp"], then quit and reopen the app — an in-app reload won't pick up config edits.

It signs in the same way, through a browser tab on the first tool call, and caches its tokens under your home directory with owner-only permissions. Same curated tools, same personal-workspace pinning.

The hosted server is the better story for almost everyone: nothing to install, nothing to keep updated, and no local process holding tokens on disk.

Asking good questions

The tools are shaped around the same query language the REST API uses, so an agent can ask precise questions rather than pulling everything and sifting. Worth knowing that it can, because "summarize my spending on restaurants over $50 last year" is one query, not a download.