MCP For Ynab

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 (claude_desktop_config.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 "YNAB_ALLOW_WRITES": "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 YNAB_ACCESS_TOKEN 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 YNAB_ACCESS_TOKEN. 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 YNAB_ACCESS_TOKEN_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:

• Confirmation gates: destructive or bulk operations (deleting a transaction, deleting a scheduled transaction, bulk-updating more than 5 transactions) require an explicit confirm: true — and bulk updates also require expected_count to match the number of rows — so the assistant has to show you what it's about to do before it does it.
• Undo: writes that change existing data — transaction edits/deletes, category edits, budget assignments/moves, scheduled transaction edits/deletes, payee renames — are journaled locally, and the undo_last tool reverses the most recent one (up to 50 writes of history, stored in ~/.cove/undo.json). Writes that *create* something (create_category, create_payee, create_account) and import_transactions are not reversible — YNAB's API has no delete for categories, payees, or accounts, and no way to undo an import. Those are still journaled (so undo history stays in order), but undo_last will tell you a given entry can't be undone and move on to the write before it. See PRIVACY.md for details on that file.

Rate limits

The underlying API allows 200 requests/hour per token, shared across every app using that same token (this server, the mobile and web apps for YNAB, other integrations, etc.). This server tracks its own usage client-side and stops itself at 190 requests/hour to leave headroom for the rest of your apps, and prefers small, server-computed responses (aggregates, summaries) over dumping raw rows where it can.

Tools

The month-close session

A guided monthly catch-up flow (Balance → Plan) built on top of the tools above: it works through blockers until the coverage gap is trusted, attributes credit-card float changes to their cause, proposes donor-first coverage moves, and — once you approve them — writes a balance-forward record via record_month_close so next month's session has a baseline to compare against (get_month_close_ledger). On the very first run, the ledger is empty — the session runs backfill_ledger for each card first and leads with what it finds (e.g. "carrying $865.75 since 2024-08") before attributing anything further.

• Claude Code: use the /month-close skill.
• Claude Desktop and other MCP clients: the server exposes the same flow as the month-close-session MCP prompt.
• Full walkthrough: docs/playbooks/month-close.md.

Remote MCP endpoint (self-host)

The remote worker (see Install (remote) above) exposes the same 35 tools over a token-authenticated remote MCP endpoint — a bearer-header route for header-capable clients, plus a token-in-path route so claude.ai's URL-only custom connectors can reach your budget directly — built on the same library entrypoint this package exports for embedding:

import { tools, buildServer } from '@walensis/cove-mcp'

It does not include always-on monitoring — that's a separate, hosted product. See docs/build-your-own-monitoring.md if you want to build that yourself on top of the same open attribution engine.

Full deploy runbook: apps/worker/README.md.

Development

This is a pnpm monorepo: packages/core (@walensis/cove-core, the API client for YNAB and domain logic) and apps/mcp (@walensis/cove-mcp, the MCP server that wraps it).

Naming: Cove is the product; @walensis/cove-core is the engine, @walensis/cove-mcp is the MCP surface, and the MCP registry entry keeps its descriptive mcp-for-ynab name for discoverability.

pnpm install

pnpm build pnpm test pnpm typecheck

To smoke-test against a real budget (read-only — it never writes):

YNAB_ACCESS_TOKEN=xxx pnpm smoke

To validate credit_card_float_history against known-good fixture values (read-only):

YNAB_ACCESS_TOKEN=xxx pnpm validate:fixtures

Disclaimer

We are not affiliated, associated, or in any way officially connected with YNAB, or any of its subsidiaries or its affiliates.

License

MIT © walensis-labs