How to connect Claude to QAD Adaptive ERP
Short answer
The cleanest way to connect Claude to QAD is a read-only Model Context Protocol connector in front of your OpenEdge reporting copy or QXtend. Claude can run through the Anthropic API or a cloud provider such as AWS Bedrock or Google Vertex; if QAD data cannot leave, use a local open-weight model instead.
Claude is good at long, messy business context: a stack of POs, an NCR history, a spec sheet and an email thread all at once. That fits QAD work, where the answer rarely sits in one screen.
To be useful Claude needs a path to your data. Today the standard path is an MCP server, a small service that exposes named tools such as get_item_availability. Claude chooses and calls tools; your code decides what data comes back.
This page walks through the connection options for QAD Adaptive ERP, how to choose where the model runs, and where the limits are.
What teams use it for
Planner copilot
A planner asks which action messages for family X are real versus noise, and Claude reads supply, demand and lead times through connector tools before answering.
Quality investigation
Give Claude an NCR, the lot's receiving record and supplier history and ask for a root-cause hypothesis list with the data behind each.
Long document comparison
Compare a customer's revised specification against your item and routing notes to list what changed and which QAD records need review.
Finance variance commentary
Claude drafts variance explanations from standard versus actual cost data and work order histories for the controller to edit.
Onboarding assistant
New buyers ask how a process works in your QAD setup, and Claude answers from your own procedures plus live lookups.
Customisation archaeology
Provide exported OpenEdge source for a legacy customisation and ask Claude to explain what it does before an upgrade.
QAD Adaptive ERP integration surface
- OpenEdge SQL access (JDBC or ODBC) to a replica for read tools; table and field names depend on your release and extensions.
- QXtend inbound and outbound services for governed write operations and event delivery.
- QAD Browse exports and report outputs as file-based sources Claude can read directly.
- Exported ABL or OpenEdge source for code explanation tasks, subject to your licence and IP policy.
- An MCP server of your own, which is the glue that turns the above into Claude-callable tools.
How it works
- 1
Choose the hosting route
Anthropic API, AWS Bedrock or Google Vertex AI for managed Claude, or an open-weight model on your own servers when data cannot leave.
- 2
Build or deploy an MCP server
Expose five to ten read tools first: item lookup, availability by site, open POs, open sales orders, BOM explosion, work order status.
- 3
Connect Claude to the server
Use stdio for local developer use and HTTP with authentication for shared team use through a supported Claude client.
- 4
Scope permissions
Bind each user to QAD domain and site restrictions inside the MCP server, not inside the prompt.
- 5
Write system instructions
Teach Claude your terminology, units, and which tool answers which question. Include examples of good answers.
- 6
Evaluate
Run your real questions and compare with QAD screens. Fix tool descriptions before blaming the model.
Example flows
Root cause draft for a supplier defect
Trigger: Quality engineer pastes an NCR number into Claude.
- Claude calls get_receipt_history for the item and lot.
- It calls get_supplier_ncr_history for the vendor over twelve months.
- It reads the attached inspection report.
- It lists three hypotheses with the supporting records for each.
Result: The engineer starts the 8D with evidence gathered instead of spending the first day on lookups.
Upgrade impact review
Trigger: A developer uploads a custom program before a version upgrade.
- Claude explains what the program reads and writes.
- It flags calls to standard programs that may change between versions.
- It proposes a test list.
- The developer verifies against a test environment.
Result: Faster understanding of legacy code, with the human still owning the upgrade decision.
Keeping data on-prem
If your QAD environment is on-prem because of ITAR, customer flow-down clauses or plain policy, hosted Claude may not be an option for QAD records. The pattern still works: keep the MCP server and database inside your network, and swap the model for an open-weight one served by vLLM or Ollama on local GPUs. You lose some capability on long-context reasoning compared with frontier hosted models, so test your top use cases before deciding. A hybrid is common: hosted Claude for non-sensitive tasks such as drafting procedures, and a local model wired to the same MCP tools for QAD data. If you use a cloud route, Bedrock or Vertex can keep processing in a region and account you control, which may satisfy some policies but should be reviewed with counsel.
Security checklist
- Read-only database credentials and a replica, not the production primary.
- Authenticate every MCP request and map it to a QAD user with domain and site restrictions.
- Redact fields such as cost, margin and customer pricing unless the role needs them.
- Review the chosen provider's data retention and training terms; do not assume.
- Audit logs for all tool calls, retained according to your records policy.
Ways to get it done
Claude with a community or generic SQL connector
Effort: Days
Quick experiments by a technical user on non-sensitive data. Not suited to shared production use.
Custom MCP server for QAD
Effort: 3 to 6 weeks
Teams that want controlled tools, auditing and per-user scoping.
Netray private deployment
Effort: A few weeks for a scoped pilot
When QAD data must stay inside with an open-weight model and the same tool layer.
Pitfalls
- !Wiring a generic SQL tool and trusting the model not to query sensitive tables.
- !Tool descriptions that are vague, so Claude calls the wrong one.
- !Returning huge result sets that waste context and slow answers.
- !Assuming cloud region selection equals compliance. Check contracts and flow-downs.
- !Skipping evaluation against actual QAD screens.
Frequently asked questions
Is there an official Claude connector for QAD?
Do not assume one. Check with QAD and in current directories. In practice most sites build a small MCP server over a read replica or QXtend, which gives you control over exactly what Claude can see.
Can Claude run inside my network?
Claude models are hosted by Anthropic or by cloud providers. For fully on-prem needs use an open-weight model with the same MCP tools. Some cloud routes keep data in your own account and region.
Does Claude learn from my QAD data?
Check the terms of the specific plan or API you use. Commercial API terms generally differ from consumer ones, but confirm retention and training settings in writing before sending production data.
What should I expose first?
Read-only lookups: item, availability, open orders, BOM, work order status. Add write capabilities only with human approval and through QXtend so QAD validation runs.
Is Claude better than ChatGPT for QAD?
Both can work through tool calling. Test on your own questions. The deciding factors are usually data policy, hosting route and tool quality, not brand.
Want Claude working on your QAD Adaptive ERP?
A free 30 minute assessment: we map the integration path for your version and setup, and where it should run.
Get a free assessment