How to use Claude with Infor VISUAL
Short answer
The cleanest way to use Claude with Infor VISUAL is a small Model Context Protocol server that exposes read-only tools over the VISUAL SQL Server database. Claude then calls those tools for live answers. For data that cannot leave the plant, the same tool layer can be pointed at an open-weight model instead.
Teams on VISUAL tend to ask Claude two different things. One is analysis and writing: summarize a list of late jobs, explain a routing, draft a note to a customer. The other is live questions about the system: what is the status of work order W1234, which POs are overdue.
The first works with exported data. The second needs a connector, and Claude's connector model is MCP, a standard way to expose tools and data to an AI client. Because VISUAL has no widely shipped MCP server of its own that I would assume, you or a partner write one, and it is typically small.
The deployment question follows: Claude is reached through Anthropic's service or through cloud marketplaces, and many VISUAL plants are on-prem and cautious. Both paths are described below.
What teams use it for
Work order status in plain language
A supervisor asks where a job is. Claude calls a tool that reads WORK_ORDER and OPERATION and answers with the current operation, last labor posted and what is next.
Shortage explanations
Claude reads requirements and supply for a job and explains which part is short, which PO covers it and whether the date works.
Routing and BOM review
Paste or fetch a structure and ask Claude to look for missing operations, odd quantities or inconsistent units before release.
Customer communication drafts
Given an order line and its status, Claude drafts a clear update. A person reviews and sends.
Long document questions
Upload quality manuals, procedures or customer requirements and ask Claude how they apply to a given job, citing sections.
Infor VISUAL integration surface
- The VISUAL SQL Server database, which an MCP server can query with read-only views. Typical objects: PART, WORK_ORDER, OPERATION, REQUIREMENT, CUSTOMER_ORDER, CUST_ORDER_LINE, PURC_ORDER and LABOR_TICKET. Verify names in your version.
- The Visual API toolkit for any supported creation or update, which you would wrap behind an approval step and not expose as a free tool.
- Crystal Reports output as an upload source for ad hoc work without any connector.
- File shares holding drawings, certificates and procedures that Claude can read alongside ERP lookups when you add a document tool.
- Network placement: the MCP server needs to reach SQL Server on the plant network and be reachable by the client, either locally or over an authenticated HTTPS endpoint.
How it works
- 1
Define the first five tools
For example get_part, get_work_order_status, get_open_po_lines, get_customer_order_status and get_shortages. Keep names and descriptions clear because Claude chooses tools from them.
- 2
Build the MCP server
Use an official MCP SDK in a language your team supports. Each tool runs parameterized SQL against read-only views and returns compact JSON.
- 3
Choose the transport
Stdio suits a single user on a workstation. Streamable HTTP suits a shared service, and then you need authentication, TLS and per-user identity.
- 4
Connect a client
Add the server to Claude Desktop, a team workspace that supports custom connectors, or your own application built on the Anthropic API. Capabilities depend on your plan.
- 5
Pick where Claude runs
Anthropic's API, or Claude through AWS Bedrock or Google Vertex AI if your organization already has a cloud agreement and prefers to keep traffic there.
- 6
Test with real questions
Collect twenty questions planners actually ask, compare Claude's answers to the VISUAL screens, and fix tool descriptions where it chose wrongly.
Example flows
Where is my job?
Trigger: Supervisor asks about a work order in Claude
- Claude calls get_work_order_status with the base id.
- The server returns operations, status and labor posted.
- Claude calls get_shortages for the same job.
- It answers with the current step and any blocking material.
Result: One question replaces three windows, and the underlying lookups are in the log.
Review a procedure against a job
Trigger: Quality engineer uploads a customer requirement document
- Claude reads the document.
- It fetches the job's routing through a tool.
- It lists steps in the routing that the requirement affects.
Result: A checklist for the engineer to confirm, with the source paragraphs quoted.
Keeping data on-prem
Claude is a hosted model, so a plant that must keep all data inside its network cannot run Claude itself there. What you can keep on-prem is the MCP server and the data. Where policy forbids sending business data out, point the same tool layer at an open-weight model served locally with vLLM or Ollama, since MCP is not tied to one model. Some organizations use Claude for non-sensitive work and the local model for controlled data.
Security checklist
- Give the MCP server a SQL login limited to SELECT on selected views.
- Authenticate HTTP transports and map calls to named users, not a shared token.
- Return only the fields the tool needs; leave out cost and pay rate by default.
- Review your Anthropic plan or cloud provider terms on retention and training before sending business data.
- Do not expose write tools until an approval step exists outside the model.
Ways to get it done
Build a small MCP server yourselves
Effort: Days for a first read-only set, more for authentication and hosting.
You have a developer comfortable with SQL Server and want direct control.
Uploads and exports only
Effort: None beyond a routine for masking data.
You want analysis and drafting with no connector.
Partner build such as Netray
Effort: Scoped pilot, then expansion by tool.
You want an audited tool layer, an approval path and the option to swap in a local model.
Pitfalls
- !Writing tools that return huge result sets, which waste context and slow answers.
- !Vague tool descriptions, so Claude picks the wrong one.
- !Assuming a connector is available on every plan; check what your plan supports.
- !Letting the server share one database login across all users with no audit trail.
- !Treating a successful demo on clean data as proof for the full production schema.
Frequently asked questions
Does Claude have a ready-made connector for Infor VISUAL?
I would not assume so. The practical route is a custom MCP server that you or a partner write over the VISUAL database. It is usually a small project because the tools are simple lookups.
Can Claude run fully on-prem for VISUAL?
Claude itself is a hosted model, so not in the strict sense. You can keep the MCP server and data on-prem and send only the needed results, or swap in a locally served open-weight model for data that must not leave.
Should I use the Anthropic API, Bedrock or Vertex?
Pick the one that matches your existing cloud agreements and compliance posture. The model capability is similar, but billing, regions, logging and data terms differ, so involve your security team.
Can Claude write back to VISUAL?
It can prepare changes, but you should apply them through the supported toolkit after a human approves. Direct writes to the database by an AI tool are best avoided.
How is this different from the ChatGPT route?
The data path is similar: a narrow read-only layer over SQL Server. The difference is the standard. MCP lets one tool server work with Claude and other compatible clients, so you are less tied to one vendor.
Want Claude working on your Infor VISUAL?
A free 30 minute assessment: we map the integration path for your version and setup, and where it should run.
Get a free assessment