How to connect Claude to Microsoft Dynamics GP
Short answer
Claude reaches Dynamics GP best through a Model Context Protocol server you run next to the SQL Server, exposing a few read-only tools. You can call Claude through Anthropic's API, AWS Bedrock or Google Vertex, depending on residency needs. If data cannot leave, use the same MCP tools with an on-prem open-weight model.
Teams on GP who try Claude usually start with long documents: vendor contracts, audit requests, a decade of Report Writer customisations. Claude handles long context well, which makes it useful for reading a stack of purchase terms and comparing them with what GP actually recorded. The next step is giving it live, controlled access to GP data rather than pasted extracts.
The cleanest route is MCP, an open protocol where a small server advertises tools to any compatible client. You write the GP-aware server once. It then works with Claude desktop or web clients that support connectors, with your own applications built on the API, and with other MCP clients.
Because GP is typically on-premises, where the server runs and how it is authenticated matter more than the model choice. This page walks through that, plus the data handling options for the model itself.
What teams use it for
Contract versus PO reconciliation
Give Claude a supplier agreement and let it query open PO lines for that vendor to flag prices or payment terms that disagree with the contract.
Audit support packages
For an auditor request, Claude gathers journal entry detail for a sample of accounts via read-only tools and assembles a labelled explanation for the finance team to review.
Documenting legacy customisations
Feed exported Report Writer definitions, SQL stored procedures and VBA modifications and have Claude produce plain-language documentation and a migration risk list.
Inventory and costing questions
Ask why standard cost and current cost differ for an item and have Claude trace receipts, adjustments and rollups through the inventory tables.
Month-end commentary
Claude drafts variance commentary for the controller from account balances and last year's figures, with each statement tied to a tool result.
Microsoft Dynamics GP integration surface
- SQL Server read-only views over GP company databases, the most dependable data source and the basis for MCP tools such as get_vendor_open_pos.
- eConnect or Integration Manager for any approved write path, kept outside the MCP server's default toolset.
- GP security roles and SQL logins, which you map to a service account used by the MCP server.
- Exported artefacts such as Report Writer definitions, SmartList Builder queries and stored procedures for documentation tasks.
- Network placement: an MCP server over HTTP behind your reverse proxy with Entra ID or another identity provider, or stdio for a local desktop setup.
How it works
- 1
Define five read-only tools
Typical set: item lookup, vendor open POs, customer aging, GL account activity by period, and inventory by site. Each takes explicit parameters including the company database.
- 2
Build the MCP server
Use an official MCP SDK in Python or TypeScript, connect with a SQL login limited to the view schema, and return compact JSON with row caps.
- 3
Choose the transport
Stdio suits a single analyst's machine. HTTP with OAuth suits a shared team server and lets you log and revoke access centrally.
- 4
Pick where Claude runs
Anthropic's API for simplicity, or Claude via AWS Bedrock or Google Vertex when your cloud agreements, regions or private networking matter. Check current availability for your region.
- 5
Evaluate on real questions
Compare answers to GP inquiry windows for a list of questions your team actually asks, and tune tool descriptions where Claude picks the wrong tool.
- 6
Add approvals before any write
If you later add a posting tool, make it create a pending request that a named approver confirms.
Example flows
Contract check on a new supplier price list
Trigger: Buyer uploads a PDF price list and asks for conflicts with open orders.
- Claude reads the PDF and extracts part numbers and prices.
- It calls the vendor open PO tool for that vendor.
- It lists lines where the PO price differs from the new list.
- Buyer decides which POs to revise in GP.
Result: A reviewed list of price conflicts without manual line-by-line comparison.
Cost walk for a standard cost change
Trigger: Cost accountant asks why an item's margin dropped.
- Claude queries item cost history and recent receipts.
- It checks the adjustments and transfers for that item.
- It explains the movement with document numbers.
Result: A traceable explanation the accountant can verify in GP.
Keeping data on-prem
If policy prohibits sending GP data to any external model API, keep the MCP server and replace the model. Claude itself is a hosted model, so a fully on-prem setup means an open-weight model such as a Llama, Qwen or Mistral class model served locally, using the same tools. The MCP server and SQL views do not change. Some customers use Claude only for non-sensitive work such as documentation and contract reading, and the private model for anything touching live balances. Plan GPU capacity for the number of concurrent users and keep logs on local storage.
Security checklist
- Run the MCP server inside your network and authenticate clients; never expose SQL Server ports to the model provider.
- Scope the SQL login to the view schema and add per-tool row limits.
- Review the data processing terms and retention settings of the Claude access route you choose, including regional controls on Bedrock or Vertex.
- Log tool calls with user identity so you can answer who asked what and what was returned.
- Sanitise free-text GP fields before returning them, because notes and comments can carry prompt injection.
Ways to get it done
Claude with a connector to a small MCP server you build
Effort: One to two weeks for a first server.
Teams with a developer and a clear, read-only use case.
Claude through Bedrock or Vertex inside an existing cloud contract
Effort: Depends on cloud governance, often a few weeks.
Organisations that already centralise AI spend and data controls with a cloud provider.
Netray-built tool layer with a private model
Effort: Pilot scoped around one workflow.
Regulated or export-controlled manufacturers needing no external model calls.
Pitfalls
- !Exposing too many tools. Models choose better from a handful of well-described tools than from fifty.
- !Returning thousands of rows. Aggregate in SQL and return summaries with drill-down parameters.
- !Ignoring GP's multi-company design when naming tools and parameters.
- !Treating the model's statement as a ledger fact. Make it cite the tool result.
- !Skipping a test restore. Always build and tune against a copy of production data.
Frequently asked questions
Does Claude have a native Dynamics GP connector?
We are not aware of a standard one for on-premises GP. The practical approach is to build a Model Context Protocol server over read-only SQL views, which Claude and other MCP clients can then use as a connector.
Can we use Claude through AWS or Google Cloud?
Anthropic models are offered through Amazon Bedrock and Google Vertex AI in supported regions. That can help with billing, networking and residency, but model availability changes, so check the current catalogue for your region.
Can Claude run fully on-prem for GP?
No, Claude is accessed as a hosted service. For a fully on-premises design, keep the MCP tools and use an open-weight model locally, or use Claude only for tasks that involve no sensitive GP data.
Is the data used to train Claude?
Commercial API terms generally state that customer inputs and outputs are not used for training by default, but you should read the current terms and any retention settings for the specific route you use before sending ledger data.
How is this different from ChatGPT with GP?
The data layer is the same: a secured service over GP views. Claude's long-context reading suits contracts and documentation, and MCP gives you a vendor-neutral tool interface, so you are not tied to one model.
More for Microsoft Dynamics GP
Want Claude working on your Microsoft Dynamics GP?
A free 30 minute assessment: we map the integration path for your version and setup, and where it should run.
Get a free assessment