Most banking integrations treat AI agents as a liability. You get read-only dashboards, manual approval queues, and zero direct account access. Meow exposes a full MCP server that lets Claude Code check balances, pull transaction history, and generate reports directly against a live business account.
This guide walks through that connection from start to finish. You'll register Claude Code against https://mcp.meow.com, authenticate with a Meow account, and run a sequence of real financial operations with the expected output annotated at each step.
Before you start, confirm two things:
- An active Meow account with a registered business checking account
- Claude Code installed locally
What You're Setting Up
Connecting Claude Code to Meow's MCP server gives Claude a typed tool interface with named operations and defined inputs and outputs. A prompt like "give me all the transactions from last month" returns a structured, staged response within a known schema, rather than requiring you to construct raw HTTP requests and parse JSON yourself.
This tutorial registers Claude Code against the https://mcp.meow.com endpoint and authenticates through claude /mcp, which opens a browser-based OAuth grant flow. Once connected, your permissions are scoped to whatever access you grant during that flow.
The full flow from registration to your first financial operation:
CLI Surface vs. OAuth Surface: Picking the Right Endpoint
Meow's MCP server exposes two distinct surfaces. Confirming which one applies to your setup before running any command saves you from authentication failures down the line.
| OAuth Surface | CLI Surface | |
|---|---|---|
| URL | https://mcp.meow.com/ |
https://mcp.meow.com/cli |
| Best for | Interactive assistants (Claude Code, ChatGPT, Cursor) | Coding agents and automation; agent-led signup |
| Auth | OAuth sign-in | Bearer API key, or guided signup for new accounts |
| Entity scope | All entities your login can reach | The entity your API key is scoped to |
This tutorial uses the OAuth surface, which is the right choice for most developers interacting with their AI agent in real time. The CLI surface makes more sense when you're building scheduled automations, like a weekly transaction export or a monitoring agent that fires an alert when your balance drops below a threshold or detects an anomalous transaction pattern.
From here, you'll establish the connection and retrieve balances and transactions via Claude Code.
Step 1: Register the Meow MCP Server with Claude Code
Registering the server is a single command. Run this in your terminal:
claude mcp add --transport http meow https://mcp.meow.com
Once the server is registered, authenticate your Claude Code client with Meow:
claude /mcp
Select the Meow MCP from the list and choose authenticate. Claude Code opens an OAuth flow in your browser:
Click Continue. Meow then shows you a permission picker where you select what access to grant your Claude Code client:
For this tutorial, check both Read access and Invoicing and billing, then click Authorize. Meow authorizes your local Claude Code instance to view your account data and stage transaction requests.
You'll see an authentication successful message in your Claude Code session once the flow completes:
Verify the connection by returning to the MCPs list and confirming that Meow shows "Auth: ✔ authenticated":
At this point, your Claude Code instance is live against your Meow account.
Step 2: Load meow.com/skills.md Before Your First Financial Command
Before running any financial operation, prime Claude with Meow's structured LLM instructions file. This step narrows tool-call variance and prevents malformed requests that require retries.
Run this prompt at session start:
"Read the instructions at meow.com/skills.md and apply them to all Meow operations in this session."
meow.com/skills.md tells Claude exactly how to format requests to each MCP tool: which parameters are required, what the expected data shapes look like, and how to handle multi-step operations.
Without it, you'll occasionally hit malformed tool calls that need retries. That's a known limitation of LLM non-determinism in agentic flows, and the skills file significantly narrows that variance.
Step 3: Your First Financial Operations
With the connection established and Claude primed, you're ready to carry out financial operations. Three prompts worth trying right away:
Check Your Account Balance
Prompt: "What is the current balance on my Meow checking account?"
Expected output: Account name, available balance in USD, and account type. If you have multiple business entities linked to your account, Claude will list each one with its corresponding balance.
Pull Transaction History
Prompt: "Show me all transactions from last month."
Expected output: A list or table with operation type (Deposit or Withdrawal), amount, currency, fees, status (Pending, Completed, or Failed), and timestamps.
You can also filter inline: "Show me only completed transactions over \$500 in March 2025." Claude passes the filter parameters directly to the MCP tool rather than pulling everything and filtering in context, which matters when you're working with high transaction volume.
Get a Cash Flow Snapshot
Prompt: ”Compare my total inflows and outflows for the last two months and tell me if the trend is positive or negative”
Expected output: A summary showing total deposits and withdrawals for each of the two months, the net position for each, and Claude's assessment of whether the trend is moving in the right direction. If you have multiple accounts, Claude will break the figures down by account before rolling them up into a combined view.
Here's what it could look like:
Rather than returning data for you to interpret, Claude aggregates two months of transaction history, calculates net positions, and surfaces a directional assessment in a single response. The kind of analysis that would normally mean exporting to a spreadsheet happens inside the conversation.
Permission Scopes
Meow's MCP scope picker appears after sign-in (or when creating new API keys from the dashboard), letting you choose what Claude Code can access. meow.read is always granted and covers accounts, balances, transactions, contacts, billing, bills, and cards. The opt-in scopes are meow.transfers (draft ACH, wire, book, and crypto payments), meow.cards (issue, update, freeze, and revoke virtual cards), and meow.billing (manage invoicing products, customers, and invoices), each shown only if your role can use it.
Granting meow.transfers doesn't mean Claude Code can move money on its own. Every tool call is checked against your role and the granted scope on that specific entity, and the narrower of the two wins. Transfers created through MCP land in a pending-approval state and execute only after a human approves them on the Meow dashboard.
What to Automate Next
You have a live Claude Code and Meow MCP session. Go to meow.com/mcp, scan the full list of available MCP tools beyond what this tutorial covered (invoicing, bill pay, corporate card management), and identify one workflow you currently handle manually.
Good starting points:
- A weekly transaction export that feeds your bookkeeping tool
- Contractor invoice payments on a recurring schedule
- A cash flow monitoring agent that pings you when available balance drops below a threshold you define
Pick the one you'd actually use, write a Claude Code prompt for it, and run it while the session is still warm.
For multi-agent architectures, generate a separate scoped API key per agent with independent permission levels. An agent reviewing incoming invoices shouldn't share credentials with the agent initiating ACH transfers. Meow supports this natively. Each agent holds its own key with its own permission boundary, and you can revoke any individual key instantly from the dashboard with no effect on the others.
Bookmark meow.com/skills.md as an ongoing reference. It updates as new tools are added to the MCP server.
Frequently Asked Questions
How is connecting Claude Code to an MCP server different from calling a REST API directly?
When Claude Code connects via MCP, it gets a typed tool interface with named operations and defined input/output schemas. It calls tools rather than constructing HTTP requests, which produces structured, schema-bound responses and eliminates most of the prompt engineering typically required to handle edge cases.
Which endpoint should I use: https://mcp.meow.com or https://mcp.meow.com/cli?
For interactive use with Claude Code (which is what this tutorial covers), use https://mcp.meow.com. That's the OAuth surface, and it covers all entities your login can reach. The /cli endpoint is designed for coding agents, scheduled automations, and agent-led account signup, authenticated with a Bearer API key.
What does the OAuth permission picker control?
After you authenticate through claude /mcp, Meow presents a permission picker. Read access covers accounts, balances, and transaction history. Invoicing and billing covers managing invoices, products, and customers. You can also grant transfer access to stage ACH, wire, and book payments. Grant only what you need for the work you're doing and scope further from there.
Why load meow.com/skills.md before running financial commands?
The skills file tells Claude the exact parameter names, data shapes, and multi-step sequencing each MCP tool expects. Loading it at session start is the most reliable way to reduce malformed tool calls and minimize retries.
Can I ask Claude to filter transaction history without pulling everything first?
Yes. Claude passes filter parameters (date range, amount threshold, status) directly to the MCP tool rather than retrieving all transactions and filtering in context. That matters when you're working with high transaction volume and want to avoid unnecessary data retrieval.
Can I run multiple Claude Code agents against the same Meow account?
Yes. Generate a separate scoped API key for each agent and assign only the permissions that agent requires. An agent reviewing incoming invoices should hold different credentials from the agent initiating ACH transfers. You can revoke any individual key instantly from the Meow dashboard with no effect on other active agents.
Is Meow's MCP integration compliant with financial data security requirements?
PII never touches LLM providers. Meow's Plaid KYC layer handles identity verification, and financial data stays within Meow's infrastructure. Meow is SOC 2 compliant (report available at trust.meow.com) and accounts are held at FDIC-insured partner banks: Cross River Bank and Grasshopper Bank, N.A.
If you've connected Claude Code to Meow and run your first financial operation, drop a comment below about what you automated first, or share the workflow you're planning to build.
Top comments (1)
Solid walkthrough. The OAuth-vs-CLI surface split and the "stage, don't execute" permission tiering (Read → Invoicing → Transfer) is the right security shape. Loading skills.md before the first call to pin parameter names is a genuinely good habit — saves a lot of malformed tool-call retries.
One thing worth flagging for anyone planning to generalize this beyond Meow: the per-bank MCP pattern doesn't scale. If every bank and neobank ships its own MCP server, an agent that touches two or three accounts across institutions ends up juggling N servers, N auth flows, and N schemas. The escape hatch is a unified account-data layer where one consent covers all of a user's banks.
That's what PSD2 AIS gives you on the EU side: one read-only consent, balances and transactions across every EU account the consumer owns, surfaced through a single API. For an agent, that collapses the N-integration problem to one. And because it's read-only AIS, the data stays at the user's own bank — no SOC 2 custody gymnastics, and you can keep PII out of the model exactly the way this guide describes.
For transparency: I maintain open-banking.io, a cert-free EU implementation of that unified AIS layer — so I'm biased toward the one-consent model. The architectural point holds regardless of provider: pick a unified read surface over N per-bank MCPs for any agent that spans accounts.
The "stage, then a human approves" separation Meow uses for transfers maps cleanly onto PSD2 PIS intent-binding too — same idea, jurisdiction-flipped.