# How to use Claude with SYSPRO

> Source: https://netray.co/integrations/syspro-claude
> Publisher: Netray (netray.co), AI and integration engineering for ERP and manufacturing systems.
> Section: Integrations

## Short answer

To use Claude with SYSPRO, put a Model Context Protocol server in front of e.NET or a read-only SQL Server login, then connect it to Claude Desktop, a Claude connector or your own app built on the Anthropic API, Bedrock or Vertex. For data that must not leave the site, use an open-weight model with the same tool layer.

## System

SYSPRO

## Capability

claude

## Intro

SYSPRO teams interested in Claude are usually after two things: strong reading and reasoning over long documents, and reliable tool use. Both matter on an ERP. A planner's question rarely maps to one table, and a good answer needs several lookups followed by a careful explanation.

Claude does not talk to SYSPRO natively. The usual bridge is an MCP server, a small program that exposes named tools such as get_item_stock or get_open_po_lines. Claude clients discover those tools and call them with arguments. The MCP server is where you enforce read-only access, row limits and logging, so it is also where your security review should focus.

Where Claude runs is a separate decision. You can use Anthropic's own apps and API, reach Claude models through AWS Bedrock or Google Vertex AI under your cloud agreements, or avoid hosted models entirely for sensitive workloads.

## Use Cases

### Planner Q&A over live data

**Description:** A planner asks which components will short next week's jobs. Claude calls stock, requirement and receipt tools and explains the shortages in plain language.

### Read and summarise supplier documents

**Description:** Claude reads certificates, quotes and long contracts that sit alongside SYSPRO purchase orders and highlights terms that conflict with the order.

### Explain BOM and routing changes

**Description:** Given two revisions of a bill of materials, Claude produces a plain-language change summary for engineering and production.

### Draft NCR and corrective action text

**Description:** Using receipt, lot and job context pulled via tools, Claude drafts the narrative and leaves the engineer to confirm root cause.

### Audit prep question answering

**Description:** Quality managers ask which lots of a part went into which jobs and customers, and Claude walks the traceability tools to answer.

### Code help for SYSPRO customisation

**Description:** Developers use Claude Code or the API to draft e.NET XML, SQL queries and add-in code, reviewed and tested before deployment.

## System Surface

e.NET business objects: query objects for inventory, sales orders and purchase orders return XML, and a wrapper can convert it to compact JSON for tools. Exact object names and XML schemas are version dependent, so verify against your installed documentation.

SQL Server read-only login: parameterised SELECT queries over InvMaster, InvWarehouse, SorMaster, PorMasterDetail, WipMaster and lot tables give fast access for analytics, with schema checked per version.

SYSPRO operators: an MCP server should authenticate to e.NET as a dedicated operator with restricted program access, or map each Claude user to their own operator where policy requires it.

Document stores: attachments, drawings and certificates often live on file shares or in a document management product, and can be exposed as a separate read-only MCP tool or a retrieval index.

Network placement: the MCP server must reach both SYSPRO and the Claude client, which determines whether you use stdio on a user's machine or an HTTP transport inside the network.

## How It Works

### Pick the first workflow

**Detail:** Choose a question set such as shortages or order status and list the 5 to 8 tools it needs.

### Write the MCP server

**Detail:** Use the official MCP SDK in TypeScript or Python. Each tool gets a clear name, a precise description and a typed input schema, and returns compact JSON.

### Connect to SYSPRO safely

**Detail:** Use a read-only SQL login for bulk reads and e.NET queries where you need SYSPRO's business logic. Keep write capability out of the first release.

### Choose the transport

**Detail:** Stdio works for a single desktop user. A remote HTTP transport behind your identity provider suits a team and lets you centralise logging.

### Connect a Claude client

**Detail:** Register the server in Claude Desktop or a supported connector, or call the Anthropic API, Bedrock or Vertex from your own application with the tool definitions.

### Evaluate

**Detail:** Run a fixed set of real questions, compare answers to SYSPRO screens, and tune tool descriptions where Claude chooses the wrong tool.

### Decide on hosting

**Detail:** Review data handling terms for the route you chose, and keep a private-model path ready for data that policy keeps on site.

## Example Flows

### Shortage explanation

**Trigger:** A production manager asks, 'Why is job 000231 at risk?'

**Steps:**

- Claude calls get_job_requirements for the job.
- It calls get_stock_by_warehouse for each short component.
- It calls get_open_po_lines for those components.
- It explains which component is late, by whom and by how many days.

**Result:** A sourced explanation with the specific PO and supplier to chase.

### Lot traceability question

**Trigger:** A quality manager asks which customers received lot L2291 of a resin.

**Steps:**

- Claude calls the lot lookup tool for the inventory transactions.
- It follows the job issues to finished goods lots.
- It maps shipments to customers and dates.

**Result:** A customer list with dates, ready for the manager to verify in SYSPRO and use in a recall assessment.

## On Prem

If SYSPRO and SQL Server sit on site and policy keeps prompts and results there, do not route them to a hosted model. Keep the MCP server and tool layer exactly as designed, and swap the model for an open-weight one served on your own GPU hardware with vLLM or Ollama, using any MCP-compatible client or a small orchestrator. Tool-calling quality is lower than frontier models on hard multi-step tasks, so test with your real questions. Some sites run a hybrid: Claude through Bedrock or Vertex in their own cloud tenancy for general work, and a local model for costing and export-controlled data.

## Security

- Review the current data retention and training terms for the Anthropic route you choose, and prefer a business or API arrangement over consumer accounts for company data.
- Run the MCP server with read-only credentials and enforce row limits, field allow-lists and timeouts in code, not in prompts.
- Put remote MCP servers behind your identity provider with per-user tokens so every call is attributable.
- Treat tool results and documents as untrusted: they can contain text that tries to steer the model, so never give a read tool the power to trigger writes.
- Keep an audit log of tool name, arguments, user and result size, retained under your records policy.

## Options

### Claude Desktop or a connector plus your own MCP server

**Fit:** Quick way for a small planning or quality team to try grounded Q&A on live SYSPRO data.

**Effort:** 1 to 3 weeks for a read-only server.

### Custom app on the Anthropic API, Bedrock or Vertex

**Fit:** Good for embedding Claude in a portal or workflow with your own approvals and logging.

**Effort:** 4 to 10 weeks depending on scope.

### Partner-built private deployment

**Fit:** Best where data cannot leave the premises, using open-weight models with the same tool design.

**Effort:** A scoped pilot in weeks plus hardware lead time.

## Pitfalls

- Writing vague tool descriptions makes Claude pick the wrong tool, so describe inputs, units and exactly what each tool returns.
- Returning entire XML responses wastes context and hides the answer, so trim to the fields needed.
- Mapping all users to one operator hides who asked what, which hurts audit and security reviews.
- Using SQL for data that depends on SYSPRO business rules, such as available quantity, gives numbers that disagree with screens.
- Forgetting company scoping lets a question about one company pull data from another.

## Faq

### Does SYSPRO have a native Claude connector?

**Answer:** Do not assume one exists. Check SYSPRO's current release notes and partner listings. In practice most sites build a small MCP server over e.NET or SQL Server, which gives you control over tools, scoping and logging and works with any MCP-capable client.

### What is the difference between Claude on Anthropic, Bedrock and Vertex?

**Answer:** The models are the same family, but contracts, regions, billing, identity and data-handling terms differ. Bedrock and Vertex fit teams already invested in AWS or Google Cloud. Review each provider's current terms with your security team before sending ERP data.

### Can Claude run entirely on premises?

**Answer:** Claude models are provided through Anthropic and cloud partners, so a fully on-prem Claude is not the normal route. For on-site processing, use an open-weight model with the same MCP tool layer, and keep Claude for workloads where cloud processing is permitted.

### Can Claude write back to SYSPRO?

**Answer:** It can if you build write tools, but start read-only. When you add writes, make them narrow, require human approval, post through e.NET so SYSPRO validation applies, and log each action with the approving user.

### How accurate are answers about my data?

**Answer:** Accuracy depends mostly on your tools. If tools return correct, well-labelled data and Claude is told to cite it, answers are usually reliable. Test with real questions and compare to SYSPRO screens, especially for calculated fields like available stock.

## Search terms

syspro claude, claude for syspro, syspro claude integration, connect claude to syspro, syspro anthropic, syspro claude mcp, syspro bedrock claude, use claude with syspro erp, claude syspro on-premise

---
Talk to an engineer: https://netray.co/contact?intent=erp_ai | Machine catalog: https://netray.co/agent/catalog.json