What it is
Plaid offers two MCP servers for developers building on its platform. The hosted Dashboard MCP server at https://api.dashboard.plaid.com/mcp/ uses Streamable HTTP and works with Production data only, so your team needs approved Production access for at least one Plaid product. You mint a token from Plaid's /oauth/token endpoint with your client ID, Production secret and the mcp:dashboard scope; access tokens expire after 15 minutes and are renewed with a refresh token. Plaid documents it for Claude.ai and ChatGPT custom connectors and for the Anthropic Messages and OpenAI Responses APIs.
Its five tools are plaid_debug_item (diagnose a connected account), plaid_get_link_analytics (Link funnel, conversion and errors), plaid_get_usages (product usage and API volumes), plaid_list_teams and plaid_get_tools_introduction. Plaid says it is under active development, with breaking changes possible and limited support.
For development, the local server in plaid/ai-coding-toolkit (uvx mcp-server-plaid, version 0.1.1 from August 2026) runs with sandbox keys. Its four tools are search_documentation, get_mock_data_prompt, get_sandbox_access_token and simulate_webhook, and its Plaid client is hard-coded to the Sandbox environment.
Install
https://api.dashboard.plaid.com/mcp/Our review
These servers are for engineers who build on Plaid, not a way to point an assistant at your own bank account, and within that scope the hosted server's Item debugging and Link funnel tools save real Dashboard time. Asking why a specific Item keeps failing, or where users abandon Link this week, becomes a question rather than a hunt.
The Dashboard server is built for backends rather than chat apps. Tokens come from a client-credentials grant using your Production secret and last 15 minutes, so a Claude.ai connector needs re-authorising or a backend that refreshes for you; Plaid's own examples mint the token in Python and pass it to the Messages or Responses API. All five tools are read-only, and Plaid publishes no rate limits for them.
On data, going by the tool descriptions, none of the five returns account balances or transactions: results are Item metadata, Link metrics and usage volumes. Those still travel to whichever model provider your client uses, and the token is derived from a Production secret, which belongs in a secret store rather than a desktop config.
The sandbox server is small, at 31 GitHub stars, but well judged. Hard-coding the Sandbox host means it cannot reach live data even with the wrong keys, and webhook simulation covers the dullest part of integration testing. Documentation search is not local: questions go over a websocket to Plaid's AskBill service. Compared with Stripe's server, which can create refunds and payments, Plaid has kept both servers away from end-user financial data, which narrows their reach and their blast radius alike.
Where it earns its place
A backend engineer on a lending app's integrations team
Work out why a customer's bank connection keeps failing
Ask Claude to run plaid_debug_item on the Item ID from the support ticket and explain the result, instead of digging through the Dashboard. There is no extra fee, but it needs Production access and a token refreshed every 15 minutes.
A growth product manager at a fintech watching onboarding conversion
See where users abandon Plaid Link this week
plaid_get_link_analytics returns opens, institution selections and handoffs for a date range; Plaid's example shows 58 opens falling to 18 handoffs over a week. A scheduled agent report replaces a manual Dashboard export.
A developer building a first Plaid integration in Cursor
Write and test webhook handling before going live
The local server mints a sandbox access token and fires simulated webhooks at the app, while search_documentation answers API questions. Sandbox is free; Limited Production allows 200 live calls per product before a paid plan.
Before you commit
- Setup
- Hosted: get Production approval, call /oauth/token with grant_type client_credentials and scope mcp:dashboard, then add the URL as a custom connector in Claude.ai or ChatGPT, or pass it to the Messages or Responses API. Local: uvx mcp-server-plaid with sandbox keys; README configs cover Claude Code, Cursor, Codex and VS Code.
- Integrations
- Hosted: five read-only tools (plaid_debug_item, plaid_get_link_analytics, plaid_get_usages, plaid_list_teams, plaid_get_tools_introduction). Local: four tools, of which get_sandbox_access_token and simulate_webhook act on the Sandbox only.
- Data and privacy
- The Dashboard token is minted with your Production secret and scoped to mcp:dashboard; results (Item metadata, Link metrics, usage volumes) go to your model provider. The local docs search sends your questions to Plaid's AskBill service.
- Cost at scale
- No fee for either server. Plaid publishes no dollar prices: Pay as You Go bills per connected account or per successful call, Growth needs a 12-month commitment, and MCP rate limits are not published.
- Limits
- Hosted server works on Production data only, may change without notice and has limited support; access tokens last 15 minutes; there are no write tools and no transaction data.
What's good
- Item debugging and Link analytics from an assistant
- All hosted tools are read-only
- Sandbox server cannot touch live data
- Webhook simulation for integration tests
What's not
- Production access required for the hosted server
- 15-minute tokens are awkward in chat clients
- Breaking changes possible; limited support
- No published rate limits
Best for
Engineering teams integrating Plaid who want faster debugging and sandbox testing.
Alternatives
Plaid MCP servers: common questions
- Is Plaid MCP servers worth it?
- Useful for Plaid teams chasing broken connections and Link drop-off, and deliberately kept away from anything that reads customers' money or moves it. We rate it 3.6 out of 5.
- How much does Plaid MCP servers cost?
- Free to use for Plaid customers; Dashboard server needs approved production access.
- What are the downsides of Plaid MCP servers?
- Production access required for the hosted server; 15-minute tokens are awkward in chat clients; breaking changes possible; limited support; no published rate limits.
- Who is Plaid MCP servers best for?
- Engineering teams integrating Plaid who want faster debugging and sandbox testing.
- How much does Plaid MCP servers cost for a team or in production?
- No fee for either server. Plaid publishes no dollar prices: Pay as You Go bills per connected account or per successful call, Growth needs a 12-month commitment, and MCP rate limits are not published.
- Is Plaid MCP servers safe for company data?
- The Dashboard token is minted with your Production secret and scoped to mcp:dashboard; results (Item metadata, Link metrics, usage volumes) go to your model provider. The local docs search sends your questions to Plaid's AskBill service.
- What are the best Plaid MCP servers alternatives?
- The closest alternatives we have reviewed are Stripe MCP and Square MCP Server.
Answers checked Oct 23, 2026.