How to use Claude with Oracle Fusion Cloud ERP
Short answer
Connect Claude to Fusion through an MCP server that wraps REST and reports, and pick where Claude runs: Anthropic's apps and API, or the same models through AWS Bedrock or Google Vertex AI. For data that cannot leave your boundary, use an open-weight model on-prem instead.
Claude is useful for the reading-heavy parts of Fusion work: long supplier agreements, messy change descriptions, and multi-table questions that need an explanation. The question is how it reaches Fusion data safely.
The cleanest pattern today is the Model Context Protocol. You run a small MCP server that exposes a few Fusion tools, and a Claude client calls them. Fusion credentials stay in the server, and Claude sees only tool results.
Where Claude itself runs matters as much as the connector. Some manufacturers already have an AWS or Google Cloud agreement and prefer to consume Claude through Bedrock or Vertex AI under that contract. Others cannot send the data off their network at all.
What teams use it for
Read and compare supplier agreements
Load a contract from Fusion Procurement Contracts attachments and have Claude extract price breaks, lead times and termination terms next to current purchase order lines.
Explain a cost rollup
Using a read tool for item cost and structure, ask Claude to walk a cost rollup and point to the components or overheads driving a change.
Triage planning exceptions
Feed planning recommendations and have Claude group them by cause, such as late supply, forecast spike or lead time change, for the planner to act on.
Summarise a quality escalation
Combine receipts, inspection results and the supplier history into a short brief for a supplier quality meeting.
Answer procurement policy questions
Ground answers in your own procurement policy documents plus live purchase order data, with citations.
Oracle Fusion Cloud ERP integration surface
- Fusion REST resources for purchase orders, receipts, items, on-hand and work orders, called by the MCP server with a scoped service user. Versions and attributes vary by release.
- BI Publisher and OTBI for reports that are easier to run than to rebuild through REST, including analyses your finance team already trusts.
- Attachments on documents such as purchase orders and contracts, retrievable through the attachment child resources where your pod exposes them.
- BICC extracts to Object Storage for a searchable copy that supports retrieval over history.
- Identity domain OAuth 2.0 for service authentication, and your corporate identity provider for the human users of the Claude client.
How it works
- 1
Choose the deployment
Decide between Claude apps for staff, the Anthropic API for your own tools, Bedrock or Vertex AI under an existing cloud contract, or an on-prem open model for restricted data.
- 2
Stand up an MCP server
Implement a few read tools: purchase order lookup, item on-hand, work order status, supplier performance. Use stdio for a single user test and HTTP for shared use.
- 3
Authenticate to Fusion
The server uses a service user or delegated OAuth. Map each Claude user to a role in the server so queries can be limited by business unit.
- 4
Add document retrieval
Index contracts and policies in a store under your control and expose a search tool so Claude cites passages rather than recalling them.
- 5
Review data handling terms
Check the retention and training terms of your chosen plan or cloud provider with your legal team, and record the decision.
- 6
Pilot with one team
Start with procurement or planning, measure how often answers need correction and expand tools only after that.
Example flows
Contract versus order check
Trigger: A buyer asks whether a purchase order matches the supplier agreement
- Claude calls get_purchase_order for the lines and prices
- It calls the contract search tool for price break clauses
- It compares them and lists any lines outside the agreed breaks
Result: A short list of variances with clause citations for the buyer to confirm.
Weekly planner brief
Trigger: Monday morning request from a supply planner
- Claude pulls the latest planning exceptions through a report tool
- It groups them by item family and cause
- It highlights the five with the highest customer order impact
Result: A prioritised starting point for the planning meeting.
Keeping data on-prem
Anthropic's models are served through Anthropic's own platform and through partner clouds such as AWS Bedrock and Google Vertex AI, not as software you install in your data center. So if the requirement is that no prompt or result leaves your network, Claude is not the answer for that workload. Keep the same MCP tool layer and point it at an open-weight model served locally with vLLM or Ollama. Many companies run both: Claude for low-sensitivity analysis and a private model for controlled program data, behind one set of Fusion tools.
Security checklist
- Keep Fusion credentials in the MCP server and never in prompts or client configuration files shared with users.
- Expose read-only tools first, and make any write tool a proposal that a person approves elsewhere.
- Scope tools by business unit and role so a Claude user cannot see ledgers outside their remit.
- Redact personal and banking data at the MCP boundary before results reach any model.
- Log tool calls with user, parameters and row counts for audit and review.
Ways to get it done
Claude apps with a hosted MCP connector
Effort: Days to build and secure the MCP server.
Small teams who want quick answers over a few Fusion tools.
Claude through Bedrock or Vertex AI
Effort: Weeks including cloud account, networking and approvals.
Companies with an existing cloud agreement and governance around it.
Private model and shared MCP tools with Netray
Effort: Scoped pilot on one workflow, then extension.
Restricted data, or a need to compare hosted and private models.
Pitfalls
- !Giving the MCP server a broad super-user account means every Claude user effectively has it.
- !Contract PDFs with scanned pages need OCR before Claude can read them, and quality varies.
- !Large report outputs overrun the context. Return summaries and offer drill-down tools.
- !Assuming a cloud region choice settles residency. Confirm where inference runs for your chosen provider.
- !Skipping a smoke test after quarterly Fusion updates, when REST attributes can change.
Frequently asked questions
Does Claude have a built-in Oracle Fusion connector?
Do not assume so. The dependable approach is to build or deploy an MCP server for Fusion that you control, then connect a Claude client to it. Check current connector availability, as the list changes.
Can I use Claude through AWS or Google Cloud?
Yes, Claude models are offered through AWS Bedrock and Google Vertex AI in supported regions. This can fit existing procurement and governance. Confirm region, retention and logging settings for your account.
Can Claude run fully inside my network?
Not as an installable on-prem product. If data must stay inside, use an open-weight model on your own GPUs and reuse the same Fusion tool layer, accepting differences in capability.
Is MCP necessary?
No, but it keeps tools reusable across Claude, ChatGPT and other clients, and keeps credentials out of the model. A plain API plus function calling also works if you only need one client.
Can Claude post transactions to Fusion?
It can propose them. Keep the actual submission behind a human approval and a separate executor using REST or FBDI, so a model error never becomes a posted entry.
More for Oracle Fusion Cloud ERP
Want Claude working on your Oracle Fusion Cloud 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