Plaid Sandbox MCP Server

Plaid's MCP server is aimed squarely at the build phase of a fintech integration rather than at production banking data. It ships inside the company's AI Coding Toolkit and gives a coding agent the three things that usually slow a Plaid integration down: accurate documentation, working sandbox credentials, and realistic test data.

The four tools

• `search_documentation` — query Plaid's docs for a product or API endpoint. This matters more than it sounds: Plaid's API surface is wide and changes, and a model working from memory tends to invent endpoint shapes.
• `get_sandbox_access_token` — provision a usable sandbox access token and item ID on demand, so the agent can write code that actually runs instead of leaving a placeholder.
• `get_mock_data_prompt` — generate customised mock financial data (transactions, accounts, balances) for testing.
• `simulate_webhook` — fire sandbox webhook events at your application so you can exercise the handler path without waiting for a real one.

Scope, and what it is not

This is the sandbox server. It does not read anyone's real bank accounts; it works against Plaid's sandbox environment with sandbox credentials from the developer dashboard. The wider AI Coding Toolkit around it also includes a Plaid CLI, a separate Dashboard MCP for runtime dashboard data, an llms.txt documentation index, and a rules/ directory of product-specific guides meant to be dropped into a CLAUDE.md or equivalent as project instructions.

Setup

Needs a sandbox client ID and secret from the Plaid dashboard, and uvx (or plain Python via python -m mcp_server_plaid). Credentials can be passed as CLI arguments or as PLAID_CLIENT_ID and PLAID_SECRET environment variables — prefer the environment variables, since argument lists show up in process listings. MIT licensed, and Plaid is explicit that correctness, security and compliance of anything you ship remain yours.