# How to use Claude with Epicor Kinetic

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

## Short answer

Connect Claude to Epicor Kinetic by running a small MCP server that wraps read-only BAQs and selected REST calls, then add it as a connector in Claude. For stricter residency, call Claude through AWS Bedrock or Google Vertex, or run an open-weight model on-prem instead.

## System

Epicor Kinetic

## Capability

claude

## Intro

Epicor users who try Claude typically start the same way: paste a BAQ export, ask for analysis, get a surprisingly good answer. The harder step is getting live, permissioned access without writing a one-off script for each question.

Claude supports the Model Context Protocol, so Kinetic can be presented as a set of named tools. Claude then decides when to call them. That keeps the ERP behind your code and your service account, and the model only sees what the tools return.

Where the model runs matters as much as how it connects. Claude can be reached through Anthropic's API and through major cloud marketplaces, and some plants will still prefer a model inside their own walls.

## Use Cases

### Ask questions across jobs and orders

**Description:** A planner asks which open jobs depend on a late part and what sales orders they feed. Claude chains an on-order BAQ, a job material BAQ and a sales demand BAQ to answer.

### Review a long engineering change

**Description:** Give Claude an ECN document plus the affected part's where-used and open job list, and ask which in-flight work needs rework.

### Customer quality response

**Description:** Draft an 8D response by combining NCR records, lot genealogy from receipts and issues, and the inspection plan, with a quality engineer reviewing every claim.

### Customisation and BPM review

**Description:** Paste a BPM directive or Epicor Function and ask Claude to explain what it does, find side effects, and suggest tests before an upgrade.

### Procurement summary memos

**Description:** Summarise supplier performance from receipts and rejects into a one-page memo for a quarterly review.

## System Surface

- BAQ execution through the BaqSvc OData endpoint is the preferred tool backend.
- Selected REST v2 business object reads for detail screens, such as a single job or PO, wrapped as individual tools.
- Service account with defined security groups, so Claude's access equals a named role and not an administrator.
- Epicor Functions for server-side shaping so the tool returns compact JSON rather than a huge dataset.
- Epicor API keys on cloud tenants, held by the MCP server and never exposed in a prompt.

## How It Works

### Choose a handful of read tools

**Detail:** For example get_job, get_open_po_releases, get_part_availability and get_sales_order. Each maps to a reviewed BAQ or a single REST call.

### Build the MCP server

**Detail:** Use an official MCP SDK in Python or TypeScript. Validate arguments, cap rows, and return concise JSON.

### Pick a transport

**Detail:** Stdio for a desktop pilot on one machine, streamable HTTP behind your gateway for shared team use.

### Authenticate the user and the ERP separately

**Detail:** The person authenticates to the MCP server (SSO), and the server authenticates to Kinetic with a scoped service account.

### Add it to Claude

**Detail:** Register the server as a connector for your organisation or as a local MCP server for pilots, then test with real questions.

### Choose the model hosting

**Detail:** Anthropic API for simplicity, AWS Bedrock or Google Vertex when your cloud agreements or regions demand it, or an on-prem model for sensitive plants.

## Example Flows

### Which jobs are at risk this week

**Trigger:** A production manager asks Claude in the team workspace.

**Steps:**

- Claude calls the late-material tool for the next five days.
- For each short part it calls the open supply tool.
- It groups jobs by customer commitment date.
- It lists the three decisions that need a human.

**Result:** A short brief with record references the manager can open in Kinetic.

### Upgrade impact review

**Trigger:** An IT lead pastes a BPM directive before a Kinetic version upgrade.

**Steps:**

- Claude explains the logic and the tables it touches.
- It flags calls that may be version sensitive.
- It drafts test cases for the sandbox.

**Result:** A checklist that shortens regression testing.

## On Prem

If the Kinetic server is on-prem, the MCP server should sit in the same network and be the only component talking to it. Whether Claude itself may see the returned data is a policy question. Many manufacturers accept Claude through a cloud provider's private endpoint in their region, with logging and no training on their data. Where that is not acceptable, as for some defense and aerospace work, run an open-weight model on your own GPUs with vLLM and reuse the same MCP tools through a compatible client. The tool layer is portable, so the work is not wasted if you change the model.

## Security

- Read-only tools first. Add write tools only as drafts that require an approving user.
- Scope the service account to least privilege and rotate its credentials.
- Log every tool call with the human requester, arguments and row count.
- Review the provider's data retention and training terms for your plan or cloud marketplace.
- Strip or tokenise fields you never want leaving the network, such as customer prices or banking details.

## Options

### Local MCP server on one machine

**Fit:** A technical champion proving value on non-sensitive data.

**Effort:** Days.

### Shared MCP server with SSO

**Fit:** A team of planners, buyers and quality staff.

**Effort:** Two to four weeks including role mapping and logging.

### Netray private deployment

**Fit:** On-prem or air-gapped plants that still want the same tool pattern.

**Effort:** Scoped pilot in weeks, with open-weight models on your hardware.

## Pitfalls

- Exposing a generic 'run any OData query' tool. The model will eventually write a bad filter.
- Returning thousands of rows, which wastes context and hides the answer.
- Using one administrator service account because it was easy.
- Assuming the cloud tenant and on-prem API paths and auth are identical across versions.

## Faq

### Does Claude have a native Epicor Kinetic connector?

**Answer:** Do not assume one exists. The reliable approach is an MCP server you control that wraps BAQs and selected API calls. Check Epicor and Anthropic directories for any official option available at the time you build.

### Can Claude write back to Kinetic?

**Answer:** It can if you build write tools, but start with drafts. Each write should name the record, show the intended change, and need a human approval step. Kinetic BPM directives will still apply to those calls.

### Should I use the Anthropic API, Bedrock or Vertex?

**Answer:** Use whichever matches your existing cloud contracts, regions and logging. The tool layer is the same. If your data cannot go to any external provider, an on-prem open-weight model is the alternative.

### Will Claude see my whole ERP?

**Answer:** No. It sees only what your tools return. That is why narrow BAQs, row limits and a scoped service account matter more than the choice of model.

### How do I stop wrong answers about quantities?

**Answer:** Compute totals and availability inside the BAQ, return record IDs with every claim, and instruct the model to cite them. Spot check against Kinetic's own screens during the pilot.

## Search terms

epicor kinetic claude, claude for epicor, claude epicor integration, claude mcp epicor, anthropic epicor kinetic, claude bedrock erp epicor, connect claude to epicor, claude ai epicor baq, epicor kinetic llm private

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