You connect ERPNext to ChatGPT or Claude by running an MCP server on your ERPNext site and adding it as a custom connector in the AI app. Then you can ask questions and update records in plain English ("show me last month's overdue customers", "create a sales order for Acme"), and the model works on your live data within one user's permissions.
Short answer
Install the open-source Frappe Assistant Core app (FAC), which turns your ERPNext site into a Model Context Protocol (MCP) server. Copy the MCP endpoint URL from FAC Admin, add it as a custom connector in Claude or in ChatGPT (Developer Mode), and sign in with your ERPNext login over OAuth 2.0. The model can then read and write ERPNext data within that user's role permissions.
This is the setup companion to my guide on whether ERPNext has AI features: that post covers what is possible, this one is the runbook I follow on client sites. I am Manoj, and I lead ERPNext and Frappe programmes at MithTech, a team of Frappe specialists in Bengaluru. I set up MCP connections for businesses on ERPNext that want Claude or ChatGPT working over their own data.
MCP is a clean way to give a general-purpose model controlled access to ERPNext, and "controlled" is the word that matters. Get permissions and staging right before anyone points a write-capable model at production.
What does connecting ERPNext to ChatGPT or Claude mean?
It means running a small server on your ERPNext site that publishes the ERP's data and actions as tools the model can call, using a shared standard, the Model Context Protocol. The model does not get a copy of your database. It sends a request ("find open invoices for this customer"), your site runs it under that user's permissions, and returns the result.
That result then goes to the model's provider, because the model has to read it to answer. Ask about overdue invoices and the customer names and amounts travel to OpenAI's or Anthropic's servers under that provider's data terms for your plan. So scope the AI user to what you are willing to send.
MCP matters because it is a standard. Before it, every AI-to-ERP link was a one-off integration. Now Anthropic's Claude and OpenAI's ChatGPT both act as MCP clients, so one server on ERPNext works with either.
Skimmable summary: the model calls scoped tools on your live ERPNext under a real user's permissions; the database stays put, but every result it reads goes to the model provider.
What do you need before you start?
Three things: a Frappe/ERPNext site you can install apps on, an AI plan that allows custom MCP connectors, and an ERPNext user whose permissions define exactly what the AI may see. Setups most often stall on the plan, because free ChatGPT cannot add custom connectors.
- A self-hosted or Frappe Cloud site on Frappe v15 or v16 (v16 shipped on 12 January 2026). You need
benchaccess when self-hosted, or the Apps tab on Frappe Cloud. - Claude (desktop or web, with connectors) or ChatGPT on a plan that allows custom MCP connectors; see the plan section below.
- A dedicated ERPNext user for the AI, with a role profile limited to the DocTypes it should touch. Never an Administrator account.
- A staging site that mirrors production, so you can test write actions safely.
Skimmable summary: an install-capable site, an AI plan with custom connectors, a permission-scoped ERPNext user, and a staging copy for writes.
How do I install Frappe Assistant Core on ERPNext?
Two bench commands fetch and install it, then FAC Admin gives you the endpoint. On Frappe Cloud you add it from the marketplace instead.
Fetch the app
On a self-hosted bench, pull the app into your bench directory:
Install it on your site
Install the app on the site, then run migrate:
On Frappe Cloud, skip the bench steps and install Frappe Assistant Core from the site's Apps tab.
Copy the MCP endpoint
In ERPNext, open Desk → FAC Admin, check the server is enabled, and copy the MCP Endpoint URL. You paste it into Claude or ChatGPT next. For a production site, follow FAC's OAuth setup guide so external clients can register and sign in.
FAC ships about two dozen built-in tools: document create, read, update and delete, search, reports, workflow approvals, schema lookup, file extraction, dashboards, and two analytics tools, run_python_code and run_database_query. Every call is written to the Assistant Audit Log. The list changes between releases, so check it against the version you install.
Skimmable summary: bench get-app, then bench install-app (or the Frappe Cloud marketplace), copy the endpoint from FAC Admin, and think twice before allowing the Python and query tools.
How do I connect Claude to ERPNext?
Add the ERPNext MCP endpoint as a custom connector in Claude's settings, then sign in with your ERPNext login.
Open connector settings
In Claude, go to Settings → Connectors → Add custom connector.
Paste the MCP endpoint
Paste the MCP endpoint URL you copied from FAC Admin and add it.
Authorise with ERPNext
Click Connect. You are sent to your ERPNext site's OAuth login. Sign in as the scoped AI user and authorise. Claude never sees the password; it gets a scoped token.
Test read-only first
Ask a question you can check: "How many sales orders were submitted last week?" Compare the number with ERPNext's own report before you trust it with anything larger.
Skimmable summary: add a custom connector, paste the endpoint, sign in through ERPNext OAuth as the scoped user, then check one read-only answer.
How do I connect ChatGPT to ERPNext?
Turn on Developer Mode, add the ERPNext MCP endpoint as a custom connector, and sign in with OAuth. The catch is the plan: custom MCP connectors are a paid feature, and free ChatGPT cannot add them.
OpenAI's developer documentation lists Developer Mode for Plus, Pro, Business, Enterprise and Edu accounts on the web; on Business, Enterprise and Edu a workspace admin turns it on. What individual plans may do with write actions has changed more than once, so check OpenAI's Developer Mode help page for your plan before you rely on writes.
| ChatGPT plan | Custom MCP connector |
|---|---|
| Free | No |
| Plus / Pro | Yes, in Developer Mode; check write actions for your plan |
| Business / Enterprise / Edu | Yes, once an admin allows Developer Mode |
The steps mirror Claude: enable Developer Mode, create a custom MCP connector, paste the ERPNext endpoint and complete the ERPNext OAuth sign-in.
Skimmable summary: ChatGPT needs a paid plan and Developer Mode to add an ERPNext connector; check OpenAI's page for what your plan allows on writes.
Should you use the MCP bridge or Raven AI agents?
Use the MCP bridge when you want ChatGPT or Claude to work on ERPNext from outside the ERP. Use Raven AI agents when you want the AI inside ERPNext as a chat teammate. Many teams run both: MCP for questions and reporting in the AI app they already use, Raven for scoped, repeatable jobs such as reading receipts into Expense Claims.
| MCP bridge (FAC) | Raven AI agents | |
|---|---|---|
| Where the AI lives | In ChatGPT or Claude, outside ERPNext | Inside ERPNext, in the Raven chat app |
| Suits | Questions, reporting, exploration | Repeatable, scoped tasks (receipts, invoices) |
| Model | Any MCP client (Claude, ChatGPT and others) | OpenAI, or a local model through LM Studio, Ollama, LocalAI or an OpenAI-compatible API |
| Setup | Install app, add connector | Install Raven, configure a bot |
For the Raven side, see what Raven is and how to set it up and the Raven agent walkthrough.
Skimmable summary: MCP puts ERPNext inside ChatGPT or Claude for ad-hoc work; Raven puts agents inside ERPNext for repeatable tasks; move any legacy Raven agents off the Assistants API.
How do you keep an ERPNext AI connection secure?
Scope the AI user tightly, test on staging and start read-only. FAC signs in with OAuth 2.0 and PKCE and runs every tool call under the connecting user's roles, so security comes down to how narrowly you scope that user: the model can never do more than the user it signed in as.
The controls I apply on every setup:
- Least privilege: a dedicated AI user with a role profile limited to the DocTypes needed.
- Read before write: grant read access, prove the workflow, then add write access on specific DocTypes only.
- Mind what leaves: leave salary, bank and personal-data DocTypes out of the AI user's roles unless you are content to send them to the model provider.
- Staging first: run the whole flow on a staging copy before production.
- Switch off risky tools: keep
run_python_codeandrun_database_queryoff unless there is a bounded, reviewed reason. - Audit: read the Assistant Audit Log regularly and revoke tokens you no longer use.
Skimmable summary: a least-privilege AI user, read before write, no sensitive DocTypes in scope, staging tests, the Python and query tools off, and a regular look at the audit log.
What are the common connection errors, and the fixes?
Most failures have one of three causes: the AI plan does not allow custom connectors, the OAuth sign-in is not set up on the ERPNext side, or the AI user lacks permission for the DocType in question. Check them in that order.
| No "add custom connector" option | Free plan, or Developer Mode off | Upgrade the plan; turn on Developer Mode |
| Connector added but "unauthorised" | OAuth not configured on the site | Follow FAC's OAuth setup guide, then reconnect |
| "No records found" for data you can see | AI user lacks read permission | Give the AI user's role that DocType permission |
| Write actions rejected | Plan or workspace does not allow writes | Check your ChatGPT plan, or use Claude |
| Endpoint unreachable | Site not publicly reachable, or a firewall | Serve the endpoint over HTTPS; check the firewall |
Skimmable summary: if a connection fails, check the plan and Developer Mode, then OAuth on the site, then the AI user's DocType permissions.
Frequently asked questions
+Can I connect ERPNext to ChatGPT and Claude at the same time?
Yes. One FAC endpoint serves any MCP client, so you can add the same ERPNext endpoint to both. Each connects as its own OAuth session under whichever ERPNext user signs in, so their access is separate. I often set up Claude for one team and ChatGPT for another on the same site, each with its own AI user and permissions.
+Do I need ERPNext v16 to connect an LLM over MCP?
No. FAC supports Frappe v15 and v16 (v16 was released on 12 January 2026). The bridge depends on the app, not on one ERPNext release. Be on a supported version with current security fixes before you expose any external connection.
+Which ChatGPT plans can connect to ERPNext?
Custom MCP connectors need Developer Mode, which OpenAI lists for Plus, Pro, Business, Enterprise and Edu accounts on the web; free ChatGPT cannot add them. On workspace plans an admin turns it on. What individual plans may do with write actions has changed over time, so check OpenAI's Developer Mode help page for your plan. Claude supports custom connectors on its paid plans.
+Is it safe to give ChatGPT or Claude access to my ERPNext data?
It is as safe as the permissions you scope. FAC runs every request under the connecting user's ERPNext roles, so the model can only see and change what that user can, and OAuth means the password is never shared. Remember that whatever the model reads is sent to its provider. I use a least-privilege AI user, start read-only, test on staging and keep the Python and query tools off unless there is a reviewed need.
+Does any of my ERPNext data leave my server?
Yes, the parts the model reads. The database stays on your server, but each tool result (records, report rows, file text) is sent to the model provider so it can answer, and it is handled under that provider's terms for your plan. FAC's optional FAC Chat add-on sends the conversation to FAC Cloud instead. Limit the AI user's DocTypes to what you are content to send.
+What is the difference between the MCP bridge and Raven AI agents?
The MCP bridge lets external models like ChatGPT and Claude work on ERPNext from outside; Raven AI agents live inside ERPNext as chat assistants for scoped, repeatable tasks. Use MCP for ad-hoc reporting and analysis, and Raven for jobs like turning a receipt photo into an Expense Claim. See my Raven for ERPNext guide for that side.
+Does connecting ChatGPT or Claude to ERPNext cost anything?
FAC itself is free and open source, so the costs are your AI plan and usage: a paid ChatGPT plan for connectors, or a paid Claude plan. There is no per-user fee for the bridge, unlike suite AI that is billed into every seat.
Comments & ratings
No comments yet. Start a new discussion.
Manoj · Published on 20 July 2026
Worked with us? Leave a Google review