How to use Claude with Infor M3
Short answer
To use Claude with Infor M3, put a connector in front of the M3 API that speaks the Model Context Protocol, then connect it to Claude Desktop, a Claude team workspace, or your own app using the Anthropic API, Amazon Bedrock or Google Vertex AI. If data must not leave your network, run the same tool layer with a local open-weight model.
Teams on M3 who try Claude usually hit the same wall as with any chat model: it is brilliant at explaining and drafting, but it does not know your item numbers, open orders or lead times. The fix is not to paste more data, it is to give Claude governed tools that read from M3 on demand.
The Model Context Protocol is the standard way to do that. An MCP server exposes a few functions such as item lookup or order status, and a Claude client calls them when relevant. The same server can later serve other MCP clients, so you are not tied to one vendor.
This page focuses on the Claude-specific choices: where the model runs, how enterprise data handling differs across the Anthropic API, Bedrock and Vertex, and what changes if your M3 is on-prem and cannot reach the internet.
What teams use it for
Order status for customer service
A service rep asks Claude where a customer order stands; Claude calls a status tool and summarises line statuses, promised dates and any holds.
Procurement analysis
Claude compares supplier delivery performance from PO line history and drafts a review agenda for the quarterly meeting.
Long-document reasoning with ERP context
Combine a supplier contract or a quality procedure with live PO and item data to answer questions that need both.
Planner copilot
Claude reads planned orders and shortages for a facility and explains which exceptions deserve attention first.
Developer help on M3 extensions
Developers use Claude to understand MI program fields, draft Mashup or extension code and review IPA flows, without touching production data.
Audit and traceability questions
Auditors ask how a specific lot moved through receipt, issue and shipment, and Claude assembles the trail from tool results.
Infor M3 integration surface
- M3 API (MI programs): the data source behind each MCP tool. Typical programs include MMS200MI, MMS060MI for balances and transactions, PPS200MI, OIS100MI and CRS610MI; exact names and transactions depend on version.
- ION API Gateway: for cloud tenants, OAuth 2.0 service-account authentication using an .ionapi file, with calls routed to M3 API. The MCP server holds the credential.
- On-prem grid access: the M3 API is reachable inside the network, often through the grid proxy; confirm path and authentication for your release.
- Infor Data Lake and Compass: a good source for history-heavy questions, queried through the MCP server rather than directly from Claude.
- ION events: optional triggers for proactive summaries, such as alerting on holds or large order changes.
How it works
- 1
Define the tool set
Start with four to eight read-only tools: item lookup, stock by warehouse, customer order status, PO status, manufacturing order shortages.
- 2
Build the MCP server
Implement it with an official MCP SDK in TypeScript or Python, mapping each tool to a typed M3 API call with strict parameter schemas and row limits.
- 3
Choose a transport
Use stdio for a single desktop user during a pilot, or streamable HTTP behind your gateway for shared team use with proper authentication.
- 4
Connect a Claude client
Add the server to Claude Desktop, or to a team workspace connector, or call Claude through the Anthropic API from your own interface with the tool definitions.
- 5
Pick the hosting route
Anthropic API, Amazon Bedrock or Google Vertex AI, depending on your cloud agreements, regions and procurement preferences.
- 6
Add per-user identity
Pass the user's identity so the server maps requests to the right M3 authority, instead of one all-seeing service account.
- 7
Evaluate and monitor
Run a fixed question set against known M3 answers, log tool calls, and review a sample weekly.
Example flows
Customer service order lookup
Trigger: A rep types a customer order number into Claude.
- Claude calls get_order_status with the number.
- The MCP server queries the customer order transactions through the M3 API.
- It returns lines, delivery dates, statuses and hold reasons.
- Claude answers in plain language and offers a draft customer email.
Result: Faster answers, and the rep still sends the message.
Shortage review before the planning meeting
Trigger: A planner asks for the top shortages for one facility this week.
- Claude calls the shortage tool with the facility and date range.
- The tool returns component, required, available, and linked PO or planned order.
- Claude groups by supplier and notes which have no inbound supply.
- The planner exports the summary for the meeting.
Result: A prepared agenda built from live data, with figures traceable to tool output.
Keeping data on-prem
Claude models are not distributed for you to host on your own hardware, so for a strictly air-gapped M3 the model itself must be an open-weight one. The MCP tool layer does not change: build it once, then point it at either Claude (through the API, Bedrock or Vertex, when policy allows) or a local model served by vLLM or Ollama for restricted data. A common compromise is Claude for non-sensitive tasks such as documentation and code help, and a private model for anything touching live orders, prices or controlled technical data. For on-prem M3 reachable only inside the plant network, host the MCP server in the DMZ or on an approved egress path and review data flows with security.
Security checklist
- MCP server holds M3 credentials; Claude never sees them.
- Read-only tools by default, with write actions as separate, approval-gated tools.
- Map user identity to M3 authority so Claude cannot reveal data the user could not open in M3.
- Check the data processing and retention terms for the Anthropic API, Bedrock or Vertex route you choose.
- Audit logs of every tool call, argument and result size.
Ways to get it done
Claude Desktop with a local MCP server
Effort: Days to a couple of weeks.
Pilots with a few power users and non-restricted data.
Team-wide MCP over HTTP with Bedrock or Vertex
Effort: Several weeks including security review.
Enterprises already standardised on AWS or Google Cloud with approved regions.
Private model with the same tools, via a partner such as Netray
Effort: Medium; needs GPU hosting and evaluation.
On-prem, defense, or contractually restricted M3 data.
Pitfalls
- !Giving Claude a tool that returns whole tables; use summaries and filters instead.
- !Using one shared service account so every user sees everything.
- !Ignoring MCP server authentication when exposing it over HTTP.
- !Overlooking that older M3 versions may lack some MI transactions you assume exist.
- !Skipping a question set test, then discovering errors in front of users.
Frequently asked questions
Does Claude have a native Infor M3 connector?
Do not assume one exists for your setup. The dependable route is your own MCP server over the M3 API. If a vendor or marketplace connector appears, evaluate its scope, authentication and logging against your security requirements before adopting it.
What is the difference between the Anthropic API, Bedrock and Vertex?
They offer Claude models through different providers, with differing regions, billing, identity and data-handling terms. Choose based on where your organisation already has approved cloud agreements, and have security review the specific terms for each.
Can Claude run fully on-prem for M3?
Claude models are not available to self-host, so a fully air-gapped setup should use an open-weight model with the same MCP or tool layer. You can still use Claude elsewhere for non-sensitive work.
Is MCP better than a custom API for this?
MCP gives a standard way for multiple clients, including Claude, to discover and call tools. A custom API can work but each client needs its own integration. For M3 teams planning more than one assistant, MCP reduces repeated work.
Can Claude update M3 records?
Technically yes if you expose write tools. Practically, start read-only and add narrow write tools with explicit approval, because a wrong autonomous change in purchasing or inventory is expensive to unwind.
Want Claude working on your Infor M3?
A free 30 minute assessment: we map the integration path for your version and setup, and where it should run.
Get a free assessment