Set up Plaid Dashboard MCP Server

The Plaid Dashboard MCP Server is an official remote MCP server from Plaid for production diagnostics. It is aimed at development teams running Plaid integrations in production who want to debug Items, analyze Link conversion data, or query API usage metrics — directly from an AI client or a model-powered application. It is not an end-user connector for bank data; it is a diagnostic tool at the team level.

Clarify prerequisites

Before connecting the server, make sure your Plaid account is approved for production access with at least one Plaid product. According to the official documentation, the Dashboard MCP Server works exclusively with production data; sandbox or development environments are not supported.

You need PLAID_CLIENT_ID and PLAID_PRODUCTION_SECRET for token creation. Store these credentials securely — they grant access to your team's production diagnostic data.

Option A: MCP client setup (Claude.ai, ChatGPT)

For MCP clients that support remote MCP server URLs, setup is straightforward. Open the MCP settings in your client and add the server URL:

https://api.dashboard.plaid.com/mcp/

If the client asks about the transport protocol, choose Streamable HTTP. For Claude.ai, Plaid's documentation refers to Claude's custom connector documentation for the exact steps.

After a successful connection, the five Dashboard tools are available: plaid_list_teams, plaid_debug_item, plaid_get_link_analytics, plaid_get_usages, and plaid_get_tools_introduction.

Option B: API integration with remote MCP

For applications that make a model provider API call and let the model invoke Dashboard MCP tools during that request, a programmatic token flow is required.

Step 1: Create an OAuth token

curl -X POST https://production.plaid.com/oauth/token \
  -H 'Content-Type: application/json' \
  -d '{
    "client_id": "${PLAID_CLIENT_ID}",
    "client_secret": "YOUR_PRODUCTION_SECRET",
    "grant_type": "client_credentials",
    "scope": "mcp:dashboard"
  }'

The response contains access_token (valid for 15 minutes) and refresh_token. Store the refresh token so you can request a new access token after expiry without re-entering credentials.

Step 2: Pass the token to the provider

OpenAI integrations (Responses API) pass the token in the authorization field together with server_url and type: "mcp". Anthropic integrations (Messages API, MCP Connector) use authorization_token. Consult the respective provider documentation for the exact SDK syntax.

Handle token expiry

Access tokens expire after 15 minutes. An expired token produces a 401 error from the MCP server or a provider-specific authentication error. Implement token refresh logic: on expiry, send a POST request to /oauth/token with grant_type=refresh_token and the stored refresh token.

First diagnostic session

Start with plaid_list_teams to retrieve available teams and their IDs. This is the entry point: follow-up tools such as plaid_get_link_analytics and plaid_get_usages require a team_id.

For Link metrics, call plaid_get_link_analytics with a time range. According to the sample in Plaid's documentation, the response includes metrics such as Link opens, institution selections, and handoffs.

For Item diagnostics, plaid_debug_item with an Item ID returns metadata indicating error causes — such as expired credentials or sync issues.

plaid_get_tools_introduction returns usage guidance for the tool set and is a useful first call to provide the AI client with the relevant context.

Security and data path

Plaid production data enters the context of the connected AI model. This includes Item metadata containing institution names or error codes. Prompt injection is possible if those fields contain model instructions. Control which Item IDs enter debug requests.

The server is read-only and does not write data back. End-user financial data such as transactions or account balances is not accessible through this server.

Review the privacy policies of the selected model provider (OpenAI, Anthropic, etc.) before production data enters queries. Plaid API rate limits also apply to requests through the MCP server.

The Dashboard MCP Server is under active development according to the documentation. Breaking changes are possible, and support is described as limited by the provider.

FAQ

What five tools does the server offer? plaid_debug_item, plaid_get_link_analytics, plaid_get_tools_introduction, plaid_get_usages, plaid_list_teams.

How long is the access token valid? 15 minutes according to Plaid's documentation; renew with the refresh token afterward.

Does the server work with sandbox data? No, production data only.

Can you retrieve end-user transactions? No. The server returns diagnostic metadata and usage metrics, not raw end-user data.