What is an MCP server, and what does it change for an ERP?
An AI assistant on its own knows neither your customers nor your stock. The Model Context Protocol gives it a standard way to call tools: search a record, read an invoice, prepare a document. An MCP server is the part of the software that publishes those tools and checks who is allowed to use them.
For an ERP the difference is concrete. Without a connection, you paste figures into a conversation and the assistant reasons on an excerpt. With an MCP server, it queries up-to-date records and answers with the amounts from your own management data, within the rights you have granted.
What can an assistant connected to the ERP do?
The useful cases are the ones that save a search or a re-entry. You phrase the request in plain language, the assistant fetches the data and presents a result for you to check.
Summarise a customer record and its opportunities before a meeting.
Prepare a quote or a credit note from a dictated request.
Send an invoice and prepare reminders, after validation.
Check availability and replenishment needs.
Analyse revenue, margins and unpaid invoices.
Which safeguards should you require before connecting AI to your management data?
An assistant that can write to an ERP can also pick the wrong customer, amount or recipient. Before comparing features, compare the guardrails: they decide whether the tool can be used on real invoices.
Separate read and action tools: looking up a record must never grant the right to change it.
Distinct, revocable keys: access specific to one company, cut off without touching user accounts.
Human approval: no invoice sent and no credit note validated without an authorised person's consent.
Consent tied to the exact action: approval covers one tool and precise arguments, not the whole conversation.
Audit trail: every operation requested by the assistant appears in the history.
Inherited rights: the assistant only sees the modules and records open to the person using it.
How does Vaka apply these safeguards?
Vaka's MCP server distinguishes read tools, which have no effect on data, from action tools. Read and write access rely on distinct keys, specific to one company, that can be revoked.
An action tool does not run without an approval token issued by an authorised person. The token is single-use, expires after a few minutes and is valid for one tool, one company and precise arguments only: the slightest change invalidates it. No assistant can approve itself, and operations stay logged.
Which assistants can connect?
ChatGPT, Claude, Grok and Gemini are among the connection environments Devlab supports. An assistant's presence in this list means neither an automatic connection nor a partnership with its publisher: what is possible depends on the assistant, your subscription with it, the Vaka modules enabled and the rights defined.
Setup is done with the Devlab team, which configures the connection and specifies the records and operations that are accessible. The team answers during Papeete business hours.
When does operating by AI add nothing?
An assistant does not fix poorly kept records: if customer files are incomplete or stock counts are wrong, its answers will be too. Nor does it replace an accountant's validation; it prepares, a person decides.
If your team already enters data quickly in the interface and has no repetitive searches to run, start without AI. Vaka remains usable without an assistant, and the connection can be added later on the same records. Dashboards composed by an assistant use Vaka data without being saved automatically in the ERP.
How do you try it without risk?
Start with read-only access, on questions you already know the answer to: you measure accuracy without committing anything. Then move to one prepared and approved action, on a single type of document.
Read-only week: unpaid invoices per customer, the month's margin, items to reorder.
First action: a quote prepared by the assistant, reviewed then approved by a salesperson.
Check: find the operation in the history and verify who approved it.
Review: keep the uses that save time, cut the other accesses.