How to use Claude with Sage 100
Short answer
To use Claude with Sage 100, expose a small read-only connector over the Sage 100 ODBC driver or SQL Server, usually as an MCP server, and let Claude call it. You can reach Claude through the Anthropic API or through AWS Bedrock or Google Vertex AI; if data cannot leave your network, you swap in a locally hosted open-weight model behind the same connector.
Sage 100 teams come to Claude for the same reasons as everyone else: long documents, careful reasoning on messy data, and readable drafts. A controller wants help with a variance explanation. A buyer wants a vendor email. An IT person wants help understanding a Visual Integrator job someone wrote years ago.
There is no assumption to make about a ready-made Claude connector for Sage 100. What exists is a standard way of connecting tools to Claude, the Model Context Protocol, and a Sage 100 data layer that is accessible enough, through ODBC or SQL Server, to wrap in a small server.
The decision that shapes the project is where inference happens. Claude is a hosted model family. Reaching it means sending prompts to a cloud endpoint, through Anthropic directly or through a cloud provider you already use. For many Sage 100 shops that is fine for some data and not for others, and it is worth deciding by field, not by system.
What teams use it for
Explain a gross margin variance
Claude reads a period comparison pulled by a tool and writes a plain-language explanation tied to specific items and customers.
Read vendor contracts against PO terms
Give it a contract PDF and the PO header terms and ask where payment terms or lead times disagree.
Document old customizations
Feed it exported Customizer scripts and Visual Integrator job definitions and have it write a readable description for your wiki.
Draft collection and expedite messages
With aging and open PO data from tools, it drafts messages that match your tone and cite the right documents.
Reconcile an intercompany or cut-off issue
Provide extracts from AR and GL detail tables and ask for likely causes of a mismatch, then verify in Sage.
Month-end commentary for the board pack
It takes already-approved numbers and drafts commentary, leaving the numbers themselves untouched.
Sage 100 integration surface
- Sage 100 ODBC driver: the simplest read path from a Windows host to the data files; wrap it in tools with parameterized queries.
- SQL Server in Premium: allows a dedicated read-only login and direct queries against tables such as AR_Customer, CI_Item, SO_SalesOrderHeader and GL transaction tables. Names depend on version and modules.
- Business Object Interface via COM or ProvideX: the supported write path, which for Claude should sit behind a draft-and-approve step.
- Visual Integrator and report exports to CSV or Excel for situations where live access is not worth the risk.
- MCP over stdio for a developer desktop or over streamable HTTP behind a gateway for a shared deployment; the Sage 100 data stays on your server either way, only the tool results go to the model.
How it works
- 1
List the questions Claude should answer
Draft 15 real questions staff ask: open PO by vendor, items below reorder, customers over credit limit. Tools should map to these.
- 2
Build the MCP server
Implement tools such as search_items, get_open_orders and get_ar_aging against ODBC or SQL, each with typed parameters and capped row counts.
- 3
Choose the Claude access path
Use the Anthropic API directly, or Claude on AWS Bedrock or Google Vertex AI if procurement and data processing terms are easier there.
- 4
Connect the client
Register the MCP server in the Claude client your team uses, or call it from your own application with the API tool-use loop.
- 5
Scope data in the tools
Return only fields needed. Drop bank details, tax IDs and anything under confidentiality clauses before it reaches the model.
- 6
Test with a restored copy of the company
Run the 15 questions, compare answers to Sage reports, and fix tool descriptions that cause wrong tool choices.
- 7
Roll out to a pilot group
Give three or four users access, collect failures, and expand only when answers are consistently right.
Example flows
Vendor delay explanation
Trigger: A buyer asks why a production item is short.
- Claude calls get_open_po_lines for the item.
- It sees two lines past due and calls get_vendor_history for lead time trends.
- It writes a short summary with PO numbers and promise dates.
- It drafts a chase email for the buyer to send.
Result: The buyer has facts and a draft in under a minute.
Close commentary draft
Trigger: Controller finalizes period numbers.
- A tool returns period versus prior period by account group.
- Claude drafts commentary and flags items over a threshold.
- The controller edits and approves.
- Final text goes into the board pack.
Result: Faster close writing with numbers sourced from the ledger, not the model.
Keeping data on-prem
Claude itself is not an open-weight model you can install on your server, so a strict on-prem requirement means using a different model for the sensitive parts. A common pattern is an MCP server that talks to Sage 100 and is model-agnostic: point it at Claude for non-sensitive work, and at a locally hosted open-weight model served by vLLM or Ollama for restricted data. Because MCP is a standard, you do not rebuild the connector. If nothing may leave your network, run only the local model and keep the server host with no outbound route.
Security checklist
- Run the MCP server with a read-only database identity and no access to administrator Sage 100 accounts.
- Apply field-level filtering in the tool, not in the prompt; the model cannot leak what it never receives.
- Review the data retention and training terms of the access path you choose and record the decision.
- Log each tool call with user, parameters and row counts for audit.
- Require authentication on HTTP transports and bind them to the internal network or a VPN.
Ways to get it done
Claude plus a lightweight MCP server you build
Effort: Two to four weeks for a first version.
Good if you have a developer and mainly read-only questions.
Claude through Bedrock or Vertex
Effort: Similar build plus cloud setup.
Good if your IT policy already approves one of those clouds.
Hybrid or fully local with a partner like Netray
Effort: A scoped pilot with a model-agnostic connector.
Good when part of the data cannot leave the network.
Pitfalls
- !Giving Claude a tool that accepts free-form SQL; use named, parameterized tools.
- !Returning thousands of rows, which wastes context and lowers answer quality.
- !Forgetting that the ODBC driver can lock files under load; keep queries short.
- !Letting a desktop stdio setup become the production architecture without authentication.
- !Assuming a numbers answer is verified; add a check against a standard Sage report.
Frequently asked questions
Does Claude have a native Sage 100 integration?
Do not count on one. Claude connects to systems through tools, commonly MCP servers. For Sage 100 you would build or obtain a connector that wraps the ODBC driver or SQL Server. Check with your reseller whether any third-party connector exists for your version.
Can I run Claude on my own server?
No, Claude is accessed as a hosted service, directly from Anthropic or through cloud providers such as AWS Bedrock or Google Vertex AI. If data must remain on your hardware, you would use an open-weight model locally, which can sit behind the same connector.
Is my Sage 100 data used to train Claude?
Terms differ by plan and access path and can change, so read the current commercial terms for your route and record them. Regardless, send only the fields a tool needs, and keep sensitive records out of tool results.
Can Claude post transactions to Sage 100?
It can if you build a write tool over the Business Object Interface, but start read-only. For writes, use a draft-and-approve step so a person confirms each posting. This protects against a wrongly interpreted instruction.
What do I need on the Sage 100 side?
A read-only identity, either an ODBC data source with a restricted role or a SQL Server login in Premium, and a host that can run the connector. The Sage 100 client is also needed on the host if you use the Business Object Interface.
Want Claude working on your Sage 100?
A free 30 minute assessment: we map the integration path for your version and setup, and where it should run.
Get a free assessment