For Ynab MCP server

MCP for YNAB

A fast, safe MCP server for YNAB — full budget access for Claude and other AI assistants, read-only by default.

35 tools covering budgets, transactions, categories, payees, accounts, scheduled transactions, and server-computed analytics (spending summaries, budget health, recurring charges, income vs. expense, net worth, month-close coverage, category and credit-card float history). Writes are off unless you explicitly turn them on, risky writes require confirmation, and writes that edit or delete existing data can be undone.

Install

There are two ways to run Cove for YNAB: local (the npm package, one client at a time) or remote (one URL + token, works from Desktop, Code, and mobile identically).

Install (local)

Runs on your machine via npx. Works with Claude Code and Claude Desktop. Not reachable from mobile — mobile clients need the remote path below.

Claude Code:

claude mcp add ynab -e YNAB_ACCESS_TOKEN=xxx -- npx -y @walensis/cove-mcp

Claude Desktop — add this to your config (claudedesktopconfig.json):

{
  "mcpServers": {
    "cove": {
      "command": "npx",
      "args": ["-y", "@walensis/cove-mcp"],
      "env": {
        "YNAB_ACCESS_TOKEN": "xxx"
      }
    }
  }
}

Replace xxx with your personal access token (see Getting a token below). By default the server is read-only; add "YNABALLOWWRITES": "1" to env (or export it in Claude Code's -e) once you want the write tools — see Read-only by default.

Any other client that speaks MCP over stdio can run the server the same way:

npx -y @walensis/cove-mcp

with YNABACCESSTOKEN set in its environment.

Install (remote)

One URL and a token, and it works identically from Claude Desktop, Claude Code, and claude.ai on mobile — no local process to keep running. This is apps/worker, a single-tenant Cloudflare Worker you deploy on your own account, with your own YNAB token: apps/worker/README.md has the full deploy runbook and how to point each kind of client at it.

The model, in short: everything in this repo — the engine, the stdio server, and this remote endpoint — is open source and free, and it's a Tool: stateless, answers when asked, does nothing on its own. Anything autonomous or metered — always-on float monitoring, digests, LLM-written narratives, the browser extension — is a Product, and lives in a separate hosted offering we run and charge for. Self-hosting this worker gets you the Tool half for free, forever; it does not include the always-on half. If you'd rather build that yourself on top of the same open engine, see docs/build-your-own-monitoring.md.

Getting a token

1. Go to app.ynab.com and sign in.
2. Open Account Settings → Developer Settings.
3. Under Personal Access Tokens, click New Token, confirm your password, and copy the token.
4. Use that value as YNABACCESSTOKEN. Treat it like a password — anyone with it has full access to your budget for as long as it's valid.

If you'd rather not put the token directly in your MCP client config, set YNABACCESSTOKEN_FILE to the path of a file containing just the token instead.

Read-only by default, writes are opt-in

By default this server only exposes read tools — nothing it does can change your budget. To enable the write tools (creating/editing/deleting transactions, assigning money, etc.), set:

YNAB_ALLOW_WRITES=1

in the server's environment and restart it. With writes enabled:

Shortened. The full README is on GitHub.