MCP server for interacting with Monzo banking API to manage accounts, balances, transactions, and pots
The Monzo MCP server has 4 tools with generally clear names and descriptions, but exhibits significant gaps in error handling, parameter validation guidance, and output schema documentation. Tool names follow verb_noun convention (get_balance, get_pots_information, pot_deposit, pot_withdraw), which is good. Descriptions are present and reasonably detailed (average ~180 chars), but lack critical error recovery guidance and downstream chaining information. Parameter schemas are mostly complete with type definitions and enums for account_type, but descriptions could be more specific about constraints and error cases. The server demonstrates basic composition (separate tools for balance, pots, deposit, withdraw) but lacks idempotency hints and dry-run/confirmation patterns for write operations. No tool annotations (readOnlyHint, destructiveHint, idempotentHint) are present despite having both read and write operations. Output schemas are documented in docstrings but not formally exposed in a way agents can discover them. Error handling is minimal, exceptions are raised but without actionable recovery guidance.
Returns the information about an account including the balance in the lower denomination of the specified Monzo account's currency. I.e. GBP, the balance is in pence. E.g. 9155 is £91.55.
Deposit money into a pot.
Withdraw money from a pot.
Returns the pots information of the specified Monzo account including the balance in the lower denomination of the specified Monzo account's currency. I.e. GBP, the balance is in pence. E.g. 9155 is £91.55.
No error recovery guidance. Exceptions are raised (e.g., 'Error: Unknown error') but do not tell the LLM what to do next or how to retry. Missing pattern:recovery-guide.
Write operations (pot_deposit, pot_withdraw) lack confirmation/dry-run support and idempotent hints. No tool annotations (destructiveHint, idempotentHint) present. An agent could accidentally deposit twice if it retries on transient failures.
pot_deposit and pot_withdraw descriptions do not state that these are state-modifying operations or mention idempotency. Agents cannot distinguish safe retries from potentially duplicate side effects.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-22 | C | 62 | 2026-07-28+ | v2 |
| 2026-03-09 | F | 0 | - | v1 |
Output schemas are documented only in docstrings (e.g., the 'Returns' section in pot_deposit shows dedupe_id, metadata, triggered_timestamp). However, these are not formally exposed in the JSON Schema input/output definitions visible to the agent's schema introspection. Agents cannot verify what fields to expect.
pot_deposit and pot_withdraw descriptions lack detail on expected behavior, constraints on amount (min/max), and what happens if the pot is locked or insufficient funds. Descriptions are bare minimum ('Deposit money into a pot.').
No parameter validation guidance. E.g., 'amount' accepts an int but no description states 'must be > 0' or 'max 999999999'. LLMs may pass invalid amounts without constraint feedback.
balance and pots return verbose API responses (e.g., cover_image_url, round_up_multiplier, product_id). These fields dilute signal and waste tokens. Responses should be stripped to essential fields only.
Code contains a clear bug: in get_balance(), response_data is accessed before the HTTP status check. Line 'if response.status_code != 200: raise Exception(f"Error: {response_data.get(..." will fail with NameError if status is not 200.