How to connect Claude to Baan IV and Baan V
Short answer
Claude reaches Baan through a connector you build, typically an MCP server over a read-only copy of Baan data. You can run Claude through Anthropic directly or via a cloud provider such as AWS Bedrock or Google Vertex, or use an open-weight model on-prem if data cannot leave.
Baan sites tend to have two kinds of knowledge problems: the data is hard to ask questions of, and the logic is buried in customized sessions and scripts. Claude is a strong fit for long documents and code reading, which makes it useful on both fronts, once it can see the right material.
The route in is the same as for any system without a modern API: you create a controlled interface. For Claude the standard interface is the Model Context Protocol, so a Baan connector can be reused by Claude Desktop, Claude Code and other MCP-capable clients.
How you host Claude matters as much as how you connect it. This page lays out the options and where each fits for a manufacturer on Baan.
What teams use it for
Read customized session documentation
Give Claude exported session help, DDC dumps and your own change notes so it can answer questions about how a customization behaves.
Review a 3GL script before a port
Use Claude to summarize what a custom script does, which tables it touches and what would need to change on a migration to Infor LN.
Ask questions of exported BOMs and routings
Upload structured exports and ask for loops, orphan components or routing gaps.
Live lookups through MCP
With an MCP connector, ask about stock, open orders or item attributes, with answers fetched on demand.
Draft migration mapping tables
Map Baan tables and fields to a target ERP's data model with Claude proposing and a consultant verifying.
Baan IV / Baan V integration surface
- Database access: read-only ODBC or a replicated database is the base for any connector. Table names follow the package, module, number pattern, which vary by version.
- Source and dictionary: Baan tools can export domain, table and session definitions and 3GL source, which is useful context for code reading.
- AFS and sessions: the supported path for writes, to be kept behind approvals rather than given to the model.
- Scheduled ASCII exports: simplest source for documents you want Claude to read.
- Infor LN web services: relevant only if you are on a later LN generation; Baan IV and V lack them.
How it works
- 1
Choose how Claude is hosted
Options include the Anthropic API, Claude via AWS Bedrock or Google Vertex AI in your cloud tenancy, or a Claude-compatible workflow that you decide to keep off ERP data entirely.
- 2
Stage the data
Create a read-only replica or export set that excludes what must not leave.
- 3
Build an MCP server
Expose tools like get_item, open_purchase_lines and stock_by_warehouse. See the MCP page for design details.
- 4
Connect a client
Point Claude Desktop or your internal client at the server using stdio on a workstation or HTTP for shared use.
- 5
Test with real questions
Use ten questions your planners actually ask, and compare answers to Baan session output.
Example flows
Customization audit before migration
Trigger: Project team starts a move from Baan to Infor LN.
- Export custom session and script sources.
- Claude summarizes each customization and the tables it touches.
- Consultant marks each as keep, replace or retire.
Result: A scoped list of customizations built in days rather than weeks of reading.
Order status question in chat
Trigger: Customer service asks about a sales order.
- Client calls an MCP tool with the order number.
- Tool returns lines, status and promised dates from the replica.
- Claude writes a short reply for the agent to send.
Result: Answer in seconds without navigating sessions.
Keeping data on-prem
Claude models are not available as downloadable weights, so a strict air-gapped Baan environment cannot run Claude itself. Options are a cloud route you accept under policy, such as Bedrock or Vertex inside a region and tenancy you control, or an open-weight model on local GPUs behind the same MCP connector. Because MCP is a protocol, you can design the Baan connector once and swap the model later without rebuilding the data layer.
Security checklist
- Review the data processing terms of whichever route you choose before connecting live data.
- Run the MCP server with a read-only database account on a replica.
- Filter by user role so a request cannot return data the user could not see in Baan.
- Never expose AFS or write tools without an approval step.
- Keep logs of tool calls for audit.
Ways to get it done
Claude on documents and exports only
Effort: Days.
Code reading, documentation and migration planning.
In-house MCP connector
Effort: Weeks.
A developer comfortable with your schema and a policy that allows the chosen Claude route.
Partner build with Netray
Effort: A scoped pilot.
You want a connector that can switch between Claude and an on-prem model.
Pitfalls
- !Sending whole database dumps rather than the minimum needed.
- !Assuming Claude knows your customizations; it only knows what you provide.
- !No version control on exported source, so answers cannot be traced.
- !Letting the connector return raw table names that confuse users.
Frequently asked questions
Does Claude have a Baan connector?
Not that we know of as a ready-made product. You build an MCP server over your own Baan data. Check current marketplaces before you commit, as the landscape changes quickly.
Can Claude read Baan 3GL code?
It can read and summarize it reasonably when you provide the source, but Baan 3GL is uncommon, so verify its claims against how the session behaves. It works better as an assistant than as an authority.
Bedrock, Vertex or direct API?
Choose by where your contracts, billing and data residency requirements already sit. If you are an AWS shop, Bedrock keeps traffic inside your account boundary. Direct API is simplest to start.
Can we run Claude fully on-prem?
No, the model weights are not distributed for self-hosting. If you need on-prem, use an open-weight model. Keep the connector model-neutral so you can change later.
Is Claude safe for write-back to Baan?
Treat it like any model: propose, review, then apply through AFS or a session. Direct writes without approval are not recommended on a system with years of customization.
Want Claude working on your Baan IV / Baan V?
A free 30 minute assessment: we map the integration path for your version and setup, and where it should run.
Get a free assessment