# How to connect Claude to Macola

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

## Short answer

Claude reaches Macola through a connector you control, typically an MCP server that exposes a few read-only tools over the Macola database or web services. Models can be used via Anthropic's API, through AWS Bedrock or Google Vertex, or replaced by an on-prem open-weight model when the data cannot leave.

## System

Macola

## Capability

claude

## Intro

Macola teams who try Claude usually start by pasting a report and noticing how well it handles long, messy text: supplier terms, engineering change notes, a ten-page customer specification. The next question is how to let it see live order and inventory data without handing over the keys to the ERP.

Claude supports tool use and the Model Context Protocol, which means the integration is a small server you write once and any compatible client can call. For Macola, that server sits on your network and talks to the database or the vendor's services. The model never touches Macola itself.

The remaining decision is where the model runs. This page lays out the choices and the trade-offs for a typical on-prem Macola site.

## Use Cases

### Reading long supplier and customer documents beside live data

**Description:** Claude reads a customer contract or quality clause and checks it against actual lead times and on-hand from Macola, flagging conflicts for the account manager.

### Explaining customizations and stored logic

**Description:** Developers feed Claude a custom report, trigger or stored procedure and ask what it does before they touch it during an upgrade.

### Order desk research

**Description:** A CSR asks about an order, and Claude pulls status, shipments and open backorders through tools, then writes the customer reply.

### Month-end commentary

**Description:** Claude reads variance data from curated views and drafts the first pass of management commentary for the controller to edit.

### Upgrade and migration analysis

**Description:** Compare field usage, custom tables and report dependencies to plan a move from Macola 10 to a newer release or to another ERP.

## System Surface

- Database layer: SQL Server for Macola 100 and, on older lines, a Progress database reached through ODBC. A read-only login on curated views is the safest tool backend.
- Vendor web services or SDK: if your version includes them, they are the right path for writes. Confirm availability and names in your release documentation.
- MCP transport: a local stdio server suits a single analyst desktop, while an HTTP server on the LAN suits a team. Choose by who needs it.
- Documents: shop packets, certificates and PDFs often live on file shares beside Macola, and Macola's own document attachments may point to them. A document tool can expose those paths.
- Reports: existing Crystal Reports are the best statement of business definitions to encode in tool descriptions.

## How It Works

### Choose the Claude access route

**Detail:** Options include the Anthropic API directly, Claude through AWS Bedrock, or Claude through Google Vertex AI. Pick the one that matches your existing cloud contract and data residency needs.

### Build a read-only MCP server

**Detail:** Expose tools such as lookup_item, order_status and open_po_lines. Write precise tool descriptions, because Claude uses them to decide what to call.

### Authenticate users, not just the service

**Detail:** Map the person using Claude to a Macola role equivalent so cost and margin fields are removed for people who should not see them.

### Add a documents tool

**Detail:** Index approved folders such as specifications and certificates, and return citations with file names so answers can be checked.

### Pilot with five users for a month

**Detail:** Give a planner, a CSR, a buyer, an accountant and a developer access. Collect the questions they ask and tune tools to match.

### Decide on any write tools last

**Detail:** If there is demand, add one narrow write behind a confirmation step in your own UI, not through the chat alone.

## Example Flows

### Contract clause against lead time

**Trigger:** Sales uploads a customer purchase agreement requiring 4 week delivery.

**Steps:**

- Claude extracts delivery and penalty clauses.
- Calls item lookup tools for the top parts quoted.
- Compares quoted lead time with purchasing lead times and current inbound.
- Returns a list of lines that cannot meet the clause.

**Result:** A risk summary before the quote goes out, with source rows cited.

### Customization explainer for an upgrade

**Trigger:** Developer opens an upgrade planning task.

**Steps:**

- Pastes the custom stored procedure and report SQL.
- Claude explains what each does and which tables it touches.
- Developer asks which parts depend on version specific fields.

**Result:** An upgrade checklist of customizations needing rework.

## On Prem

If data cannot go to a cloud model, you have two sound patterns. One is to keep the MCP server on-prem and run a private model behind the same tool interface, so the experience stays similar while inference stays in your network. Open-weight models such as Qwen, Llama or gpt-oss class served with vLLM can handle the tool calling and summarizing these workflows need. The other is to use Claude through a cloud account for non-sensitive work only, and route anything touching pricing, customers or controlled data to the private model. In both cases the MCP server and its audit logs stay inside your firewall. For defense and aerospace customers, the second pattern with strict routing rules is commonly easier to defend to an auditor.

## Security

- Run the MCP server with a read-only database login and curated views only.
- Keep tool outputs small and field-filtered; do not return whole tables.
- Log every tool call with the human user who triggered it.
- Review the provider's current data handling and retention terms for the route you pick, and record them.
- Do not include secrets or connection strings in prompts or tool descriptions.

## Options

### Claude desktop or workspace with a self-built MCP server

**Fit:** A small team with a developer and non-sensitive data.

**Effort:** Roughly 2 to 4 weeks for a first set of read-only tools.

### Claude via Bedrock or Vertex with your cloud controls

**Fit:** You already have a cloud tenant and security review for it.

**Effort:** 3 to 6 weeks including networking from the on-prem gateway.

### Private on-prem model behind the same MCP interface, with Netray

**Fit:** Restricted data, air-gapped networks, or contract requirements.

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

## Pitfalls

- Giving Claude a single shared database login so every user sees every field.
- Vague tool descriptions that make the model guess which tool to call.
- Assuming a cloud route removes the need for a data classification decision.
- Letting the model generate free-form SQL against production when curated tools would do.

## Faq

### Does Claude have a Macola connector?

**Answer:** I know of no vendor-published one, so plan on building an MCP server or using a partner's. The work is mostly in the data views and role handling, not in the MCP code itself.

### What is MCP in this context?

**Answer:** The Model Context Protocol is an open way for AI clients to call tools. You write one server that wraps Macola lookups, and Claude and other compatible clients can use it without bespoke integration for each.

### Can Claude run entirely on-prem?

**Answer:** Claude itself is accessed through Anthropic or cloud partners, so a fully air-gapped setup uses an open-weight model instead behind the same MCP tools. Check current offerings if you need a specific vendor model.

### Which Macola versions work?

**Answer:** The pattern works across versions because it relies on the database or web services you already have. Table names, drivers and write options differ, so scope per version.

### Is it safe to let Claude write back?

**Answer:** Only through narrow actions behind human confirmation. Start read-only, and measure accuracy for a month before enabling even one write.

## Search terms

macola claude, claude for macola, macola claude mcp, connect claude to macola, macola anthropic, macola erp claude bedrock, claude erp on-prem macola, macola 100 claude, macola ai assistant claude

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