How to connect Claude to Sage X3
Short answer
Connect Claude to Sage X3 by running an MCP server that wraps SData or web services, then attaching it to Claude. Choose the Anthropic API, AWS Bedrock or Google Vertex for hosting, or a private open-weight model when data must stay on-prem.
Teams on Sage X3 who like Claude usually want two things: to ask questions against live ERP data, and to keep a record of what it saw. The cleanest pattern is Model Context Protocol, where a small server presents X3 as a set of tools Claude can call.
There is no need to move your ERP to make this work. The MCP server sits next to Syracuse, speaks SData or your published web services, and enforces what Claude may read.
The harder question is where the model runs. That is a data-governance decision, and Sage X3 customers in manufacturing and defence often have strong views on it.
What teams use it for
Ask questions over live open orders
Through MCP tools, Claude finds which customer orders are at risk from a late component and cites the lines.
Review a supplier contract against PO history
Upload a contract and have Claude compare terms to purchase order delivery performance pulled via a tool.
Explain a failed X3 import
Paste an import error log and a template to get likely causes, such as missing reference values or wrong site.
Draft month-end commentary
Give Claude trial balance movements and have it draft variance commentary for the controller to edit.
Document customisations
Feed exported X3 scripts and screen definitions to produce readable documentation for a handover.
Sage X3 integration surface
- SData 2.0 REST on Syracuse for typed reads of orders, items, suppliers and stock, subject to what your folder publishes.
- SOAP web services for estates where SData is not exposed.
- Database replica views (Oracle or SQL Server) for aggregate questions, accessed through parameterised queries only.
- X3 script and class export for code-oriented tasks such as explaining a custom subprogram.
- Reporting outputs and attachments stored with documents, usable as context when permitted.
How it works
- 1
Decide the hosting route
Anthropic API for simplicity, AWS Bedrock or Google Vertex if your cloud contracts and region controls already sit there, or a private model for on-prem.
- 2
Build the MCP server
Implement tools such as find_sales_order, get_item_stock and list_open_po_lines. Return compact JSON and clear errors.
- 3
Choose a transport
Local stdio for a single power user testing, or remote HTTP behind your gateway for a team.
- 4
Authenticate to X3
Map each Claude user to an X3 user or role, or use a service user with narrow rights. Avoid one all-seeing account.
- 5
Attach to Claude
Register the server as a connector in your Claude plan or application. Features differ by plan, so check current documentation.
- 6
Log and review
Keep a record of tool calls and review them weekly during the pilot.
Example flows
At-risk orders question
Trigger: Planner asks which customer orders depend on a delayed component
- Claude calls get_item_stock and list_open_po_lines for the item.
- It calls find_sales_orders_using_item.
- It lists orders by due date with the supply gap.
Result: The planner gets an evidence-backed list and decides what to expedite.
Import error diagnosis
Trigger: A consultant pastes a rejected import log
- Claude reads the log and the template columns.
- It identifies missing site or tax code references.
- It proposes corrected rows for review.
Result: Fewer rounds of trial and error before the import passes.
Keeping data on-prem
Claude models are delivered through Anthropic and its cloud partners, not installed on your servers. If policy says nothing leaves your network, run the MCP server and tool layer on-prem and swap the model for an open-weight one served with vLLM or Ollama, since MCP works with other clients too. Some customers use Claude for non-sensitive work such as documentation and drafting, and a private model for anything touching item costs, customer prices or controlled data. Check current Anthropic and cloud provider terms on retention and regional processing when making the call.
Security checklist
- Expose read-only tools first and require approval for any write.
- Run the MCP server inside your network and publish it only through an authenticated gateway.
- Scope tools by X3 role so Claude cannot see data the user could not.
- Log every tool call with user, arguments and result size.
- Review vendor data handling and regional options for your plan before connecting real records.
Ways to get it done
Claude with exports and uploaded files
Effort: Hours.
Quick start for drafting and analysis with low-sensitivity data.
In-house MCP server over SData
Effort: Two to four weeks for a first read-only set.
Fits teams with an X3 developer and API experience.
Netray with your X3 partner
Effort: A scoped pilot over several weeks.
When you want a tested tool layer, approvals and a private model option.
Pitfalls
- !Exposing too many tools. Models pick better from a short, well-named set.
- !Returning entire X3 records, wasting context and leaking fields.
- !Using one service account so audit shows nothing about the real requester.
- !Assuming connector availability is the same on every plan.
- !Skipping site and folder parameters, producing answers for the wrong entity.
Frequently asked questions
Does Claude have a native Sage X3 connector?
We know of none maintained by Sage, and offerings change, so check current directories. The practical route is an MCP server you or a partner build over SData or web services. It is a modest amount of code if the API is already published.
Can Claude run on-premises?
Not as a self-hosted model that we can point you to. You can use Anthropic's API or a cloud provider's hosting, or run an open-weight model on your own hardware and reuse the same MCP tools.
Which hosting route is best?
It depends on where your contracts, region rules and security reviews already sit. Bedrock or Vertex suit teams already on those clouds. The direct API is simplest. For regulated records, on-prem with an open model avoids third-party processing.
Can Claude update Sage X3 records?
It can be given write tools, but start read-only. Where writes are needed, produce a draft that a user approves, and execute it under X3 rights with your normal audit trail.
Do I need to expose X3 to the internet?
No. Keep Syracuse internal. Put the MCP server on the same network and expose only that, if at all, through an authenticated gateway or VPN.
Want Claude working on your Sage X3?
A free 30 minute assessment: we map the integration path for your version and setup, and where it should run.
Get a free assessment