Pennylane MCP server: connect AI agents to Pennylane from the EU

Pennylane is a Paris-based finance and accounting platform used by French SMEs and by accounting firms that keep their clients' books in it. Its official MCP server runs at https://mcp.pennylane.com/mcp over Streamable HTTP.

The server is read-only in its current version: it never creates, edits or deletes data, files returns or moves money. A company connection covers one company on a paid plan (Essentiel and above); a firm connection lets an expert-comptable reach every client file in the portfolio. Two operations tools extend reading to fixed assets, loans and cash forecasts.

For an accounting firm, an agent can list the client files with a quarterly VAT frequency, check which VAT returns are due this month, and flag files where turnover and collected VAT do not match. For a company, it can summarise receivables, flag invoices more than 60 days late, and draft reminder emails in another connector.

Finance teams also use it for analysis: a trial balance with prior-year comparison, every movement over €500 on account 606, or bank transactions still waiting for reconciliation.

On inference.eu, add the Pennylane connector to an agent. The first time it is used, a Pennylane sign-in page opens: you log in with your own account, choose the company or firm, and approve the data types the agent may read. There is no API key to copy.

Pennylane uses OAuth 2.1 with PKCE and scopes per data type, such as mcp:customer_invoice:read. The resulting token is stored encrypted in the EU and is only available to the agents you allow.

The agent reads from Pennylane's own server, so data leaves Pennylane only as the answer to each request, as it would for any AI client. Pennylane notes that once data reaches the AI provider, that provider's terms apply, including where its servers are.

With inference.eu, the model that reads your accounting data runs in EU data centres, and your connection credentials are stored encrypted in the EU. Accounting data includes personal data of customers and suppliers, so record this processing in your register.

Because the server cannot write, the main risks are reading more than needed and passing data to another tool. Grant only the data scopes the workflow uses, and remember that the agent sees exactly what your Pennylane user can see.

On inference.eu you can require a person to approve tool calls before they run, for example any call that sends Pennylane figures to email or chat connectors. For firm connections, decide which colleagues may use an agent across the whole portfolio and which only for their own client files.