How to connect Claude to JD Edwards World
Short answer
Claude reaches World through a connector you build: usually an MCP server that wraps read-only DB2 for i views. You then choose where Claude runs: Anthropic's API, a cloud provider's hosting such as AWS Bedrock or Google Vertex, or, for data that cannot leave, a private open-weight model using the same tools.
Teams running World on IBM i often like Claude for long documents and careful reasoning, such as reading a messy PO history or a stack of supplier terms. The question they ask is how to give it live access to the ERP without opening the Power server to the internet.
The answer is a middle layer. Claude never connects to IBM i. It calls tools that your service exposes, and your service connects to DB2 for i with a limited profile. The tool contract is the same whether Claude is the model or you later swap in another one.
This page walks through that layer and the hosting choices, and is honest about where Claude is a good fit and where a private model should take over.
What teams use it for
Plain-language answers on open orders
Ask which sales orders for a customer are at risk this month and get a grounded answer built from open order lines, work orders and available stock.
Reading long supplier and contract documents
Compare terms in a supplier agreement against what World shows as lead time and price breaks, and flag mismatches.
Documenting a legacy customization
Feed a CL or RPG program and its called objects (non-sensitive) and have Claude produce a readable description for the IBM i team.
Month-end commentary
Give Claude the variance table from a finance extract and have it draft the narrative, which the controller edits.
Claude with a view-backed MCP connector
Let planners ask about item availability, BOM components and open demand directly from a Claude client, with every call logged.
JD Edwards World (IBM i) integration surface
- DB2 for i through ODBC or JDBC (jt400), the main read path for a connector. Use SQL views in a dedicated library.
- A Model Context Protocol server that you write in Python or TypeScript and run on a separate host, exposing the views as tools.
- World programs you already have for inquiry, callable through ILE procedures or IWS if your team publishes them; names depend on your release and customization.
- IFS and data queue exports for batch or event driven contexts, when you do not want live queries.
How it works
- 1
Choose the Claude surface
Claude apps can use remote or local MCP connectors; developers can use the Anthropic API with tool use. Pick based on who needs access and how you manage identity.
- 2
Build the connector as an MCP server
Define tools like get_item_availability and list_open_po_lines with clear descriptions and typed inputs. Keep results short and structured.
- 3
Connect to DB2 for i safely
Use a service profile with access to a views library only. Set the library list explicitly and decode Julian dates in the views.
- 4
Host and authenticate
Run the connector inside your network or a private VPC. Use remote transport with OAuth for shared use, or local stdio for a single developer machine.
- 5
Pick the model hosting
Anthropic's API, Claude on AWS Bedrock or on Google Vertex AI can fit different procurement and region rules. Confirm retention and logging terms for each.
- 6
Pilot with real questions
Collect twenty real questions from planners, test the answers against World screens, then widen the tool set.
Example flows
Customer call prep
Trigger: Account manager asks Claude about a customer before a meeting
- Claude calls a tool for open orders and ship dates for the customer number.
- It calls another for lines past due and the reasons in order notes.
- It summarizes risks and open questions in a short brief.
Result: The manager walks in with accurate facts, each traceable to a tool result.
Supplier terms check
Trigger: Buyer uploads a new supplier agreement PDF
- Claude reads the PDF for payment terms and lead times.
- It calls a tool to fetch the supplier's current terms in World.
- It lists differences for the buyer to resolve.
Result: Inconsistent data is flagged before the next PO is cut; the buyer decides what to update.
Keeping data on-prem
If data must stay inside the plant or a sovereign environment, the Claude hosting options may not satisfy your policy, even with a private cloud route. Keep the MCP connector and the DB2 for i views exactly as designed, and run an open-weight model on your own GPUs behind the same tool contract. That lets you use Claude where policy allows, for example on public documentation or code explanation, and a local model for part, price and customer data. Because MCP is an open protocol, the connector does not change when the model does.
Security checklist
- Give the connector the least privilege: a views-only IBM i profile, read-only first.
- Authenticate users at the connector so each call maps to a person and a role.
- Log tool name, arguments, row counts and user for every call; retain per your policy.
- Review the provider's data retention and training terms for your chosen route before sending ERP data.
- Do not give Claude tools that post or delete in World; stage proposals for human approval instead.
Ways to get it done
Claude app with a custom MCP connector
Effort: Several weeks for a secure connector and rollout.
Small groups of knowledge workers and analysts exploring World data.
Claude via API, Bedrock or Vertex inside your own application
Effort: A development project; depends on your app team.
You are building an assistant or workflow and want Claude as the model.
Private model with the same MCP tools, with a partner like Netray
Effort: Pilot in weeks; model sizing is the main unknown.
Regulated or contract-bound data that cannot go to any external provider.
Pitfalls
- !Returning wide result sets. Large tables burn context and slow answers; summarize in the view.
- !Tool descriptions that are vague. Claude chooses tools from the description, so name units, branches and date logic.
- !Library list mistakes that point the connector at test data.
- !Skipping user-level auth and using one shared login, which makes audit meaningless.
- !Ignoring that World customizations change field meaning. Document local conventions in tool descriptions.
Frequently asked questions
Can Claude connect to JD Edwards World directly?
No. Claude calls tools exposed by a connector, typically an MCP server you host. That server queries DB2 for i using a limited profile. This keeps IBM i off the internet and lets you control exactly which data and actions are available.
What is MCP and do we need it?
Model Context Protocol is an open standard for exposing tools and data to AI clients. You do not strictly need it, since you can use plain tool calling in the API, but MCP lets the same connector work with Claude and other clients, which saves rework.
Which hosting route is right for us?
It depends on your contracts, regions and cloud standards. Anthropic's API is direct, while Bedrock and Vertex let you stay inside an existing cloud agreement. If none can be approved for your data, use a private open-weight model.
Will Claude understand our World customizations?
Not unless you tell it. Put naming conventions, status codes and local logic in the tool descriptions and a short knowledge document. Validate answers against the screens your users trust before wide rollout.
Can Claude update records in World?
It can be given that ability, but we would not for a first release. Stage a proposal and have a person post it through World. This preserves edits and audit and limits risk from a wrong answer or malicious input.
Want Claude working on your JD Edwards World (IBM i)?
A free 30 minute assessment: we map the integration path for your version and setup, and where it should run.
Get a free assessment