Skip to content
EasySoft Tech
integracionesdolibarr

Dolibarr MCP server: connect Claude and ChatGPT to your ERP

What an MCP server really adds to Dolibarr, how permissions and read-only mode work, and what to check before an AI assistant writes to your ERP.

EasySoft Tech S.L.6 min read

The question keeps coming up in Dolibarr forums: can an AI assistant read my invoices, stock and customers without pasting spreadsheets into a chat window? Yes, with an MCP server between your ERP and the AI client. Here is what it exposes, how it differs from the REST API, and what to check first.

What a Dolibarr MCP server is

A Dolibarr MCP server is a small service that publishes Dolibarr operations as tools an AI client can discover and call.

MCP stands for Model Context Protocol: an open standard describing how an assistant finds which tools a server offers, what arguments they take and what comes back. Without it, every assistant needs custom glue code.

A tool is one named operation with typed inputs: list unpaid invoices over thirty days, get a product's stock, create a draft quote. The model does not have to guess your schema. It picks a tool, fills the arguments, gets structured data back.

A Dolibarr MCP server typically exposes third parties, products, stock, invoices, orders and projects. EasyAPI is the module we build at EasySoft Tech S.L., with a REST API and a built-in MCP server. The MCP introduction covers the protocol.

MCP vs the Dolibarr REST API

The REST API is the transport; MCP is the tool layer the AI client understands.

Dolibarr ships a REST API in the core, off by default, and each user has a key inheriting that user's permissions. But it is a catalog of endpoints: GET invoices, POST orders. A client can call them only if it knows they exist.

MCP wraps that catalog in descriptions a model can read: a tool called get_stock, a product reference in, quantity and warehouse out.

If your shop pushes orders every ten minutes, plain REST is simpler and easier to monitor. MCP pays off when the caller works from natural language and you cannot predict which operation someone needs. Our Dolibarr REST API guide covers endpoints and authentication.

What an AI assistant can and cannot do with your ERP

An AI assistant connected through an MCP server can read what its user may read and create drafts, but it cannot invent data that is not in Dolibarr.

Start with reads: unpaid invoices, stock under a threshold, a contact history, orders pending shipment. Writes should stay narrow: drafting a quotation from a described need is realistic, but validating it, sending it or recording a payment belongs behind a person.

EasyAI shows a confirmation card before it writes: you see the object and its values, then accept or reject. Demand that from any MCP setup rather than assuming it.

The assistant sees only what its user sees. Treat any figure without a source record as a warning: ask which invoice or product line the number came from, and check that record before you act on it.

Where the MCP server can live

An MCP server can be a module inside Dolibarr or a separate process outside it, and that choice decides who owns permissions, updates and business logic.

Inside, it runs in the same PHP process and database, reusing Dolibarr's permission checks, triggers and numbering rules and writing through the classes the interface uses. The cost is a module to install and update.

Outside, a container talks to Dolibarr over the REST API. It scales independently and keeps the MCP surface off your web server, at the cost of a second service to patch. It also usually holds one API key, so every conversation acts as that user unless the server maps credentials per person, which flattens your permission model.

The permission model to insist on

Before you connect any AI client, insist on a dedicated service user with least privilege, read-only access first, and an audit trail you can read.

  • A dedicated Dolibarr user, not an admin account and not your own.
  • Least privilege: a stock assistant has no business in accounting.
  • Read-only first. A week of read tools shows what people actually ask.
  • Scopes per tool or area. EasyAI limits each agent to none, read, create and edit, or everything.
  • Key storage and revocation. Keep the key out of transcripts, send it over HTTPS only, and test rotation before you need it.
  • An audit log: which user, which tool, when, what came back. Otherwise you cannot answer who created this draft.
  • A human decision before anything is validated or sent.

How the connection works, step by step

The path from a fresh Dolibarr to a working assistant has six steps.

  1. Enable the REST API and confirm a simple GET request.
  2. Create the service user, give it minimum rights, generate its key.
  3. Install the MCP server. With EasyAPI it comes in the module; otherwise it is a service pointing at your URL and key.
  4. Add it in the client: URL, authentication header, connection test. Follow the client's documentation for the exact menus.
  5. Ask for unpaid invoices over thirty days and compare with the Dolibarr screen.
  6. Enable write tools only for the objects you need.

If the assistant uses a cloud provider, the data you send goes to that provider. Choose it on purpose. EasyAI uses whichever provider and key you configure.

An assistant inside Dolibarr or an external AI client

An assistant inside Dolibarr fits daily ERP work; an external MCP client fits people who already live in Claude, ChatGPT or an editor.

Inside, it runs in the user's session: permissions are that user's and the confirmation card appears on the screen where the record will be created. That is the EasyAI pattern, with EasyAPI required underneath and each agent limited by ERP area.

External clients suit questions that arrive outside the ERP, like checking a supplier quote while reading email. They connect over the network, so the key is the security story.

FAQ

Can ChatGPT or Claude read my Dolibarr?

Yes, given a reachable MCP server and a valid API key. The model sees only what the tools return for that user's rights.

Is an MCP server safe?

It can be, if you treat it like any other API exposure: HTTPS, a least-privilege user, a revocable key, read-only scope at first, an audit log. The risk is rarely the protocol but a key with too much power.

Can it create invoices?

It can draft them. Validation, sending and payment recording should stay behind a person.

Do I need to write code?

No, with a module that already includes the MCP server. Custom servers may need development.

Where to start

Start with one module, a read-only user and a week of questions, comparing each answer against the Dolibarr screens before the assistant can change anything. EasyAPI supplies the API and the MCP server; EasyAI adds the in-ERP assistant. If you are still weighing Dolibarr, the Spanish fit guide Dolibarr te encaja sorts the questions by business type.