How to use Claude with Katana
Short answer
Claude connects to Katana most cleanly through an MCP server that wraps the Katana REST API, which lets Claude read stock and orders on demand. You can reach Claude through Anthropic directly, AWS Bedrock or Google Vertex AI depending on where your data governance sits.
Katana teams who try Claude usually start by dropping a spreadsheet into a chat. It is a fine start, but the value shows up when Claude can ask Katana questions itself and carry the result into a longer piece of work, like a supplier review or a build plan for next week.
That is what the Model Context Protocol is for. It gives Claude a standard way to call tools, and Katana's REST API is easy to wrap. This page is about the Claude side of the decision: which route to reach Claude, what it should be allowed to do with Katana, and how that differs from using another assistant.
What teams use it for
Weekly planning memo
Claude reads open sales orders, manufacturing orders and stock, then writes a plan memo that states assumptions and lists decisions needed from the owner.
Recipe and cost review
Given recipe rows and current material costs, it checks which products have drifted in margin and why.
Supplier performance summary
Compares expected and received dates across purchase orders per supplier and writes a short scorecard.
Long document grounding
Combine Katana order data with a customer's specification PDF to confirm that ordered variants match required options.
Data cleanup assistant
Scans variants and materials for duplicate SKUs, missing units or inconsistent naming, and proposes a cleanup sheet for a human to apply.
Katana integration surface
- Katana REST API with bearer key authentication, wrapped by a local or hosted MCP server so Claude Desktop or other MCP clients can call it.
- Claude can also call tools defined directly through the Anthropic Messages API for your own applications, using the same wrapper functions.
- Katana exports (CSV and similar) for a no-engineering route, uploaded to a Claude conversation or project.
- Webhooks, if you want a Claude-powered backend job triggered by Katana events rather than a person asking.
- Rate limits on the API key apply, so MCP tools should cache catalogue data and return trimmed results.
How it works
- 1
Choose how Claude is reached
Direct Anthropic access is simplest. If your company already governs AI through AWS or Google Cloud, Claude on Bedrock or Vertex AI keeps usage inside that contract and region setup.
- 2
Pick the interface
People using chat can connect an MCP server. Backend jobs use the API with tool definitions. Start with chat to learn which questions matter.
- 3
Build the Katana tool wrapper
Implement read-only tools for stock, orders, manufacturing orders, purchase orders and recipes. Keep descriptions precise.
- 4
Write a project or system prompt
Define your terms: which location is the main warehouse, how you treat safety stock, and when Claude must ask instead of assuming.
- 5
Review outputs with a checklist
For the first month, a planner checks the numbers against Katana before acting, and misses become prompt or tool fixes.
- 6
Add controlled writes later
If needed, add a draft-only write tool and a human confirmation step before the real API call.
Example flows
Friday build plan
Trigger: Owner asks Claude to plan next week's builds
- Claude pulls open sales orders due within two weeks.
- It reads recipes and stock for each required variant.
- It identifies materials that must be purchased and lead times from supplier notes.
- It writes a plan with a build sequence, risks and questions.
Result: A plan memo in minutes, with every figure traceable to a tool call.
Margin drift check
Trigger: Month end close
- Claude reads sales order rows and current material costs.
- It recomputes expected margin per variant.
- It lists products whose margin fell and the material that moved.
Result: A short list of pricing or sourcing decisions for the owner.
Keeping data on-prem
Katana is cloud only, so there is no on-prem Katana to connect to privately. If you need Claude-grade reasoning without leaving a cloud boundary, use Claude through a cloud provider region you control, such as Bedrock or Vertex AI. If you need the model itself on your own hardware, Claude is not a self-hosted option, so use an open-weight model behind the same MCP or tool layer. Netray builds that layer so you can swap the model without rewriting the Katana tools.
Security checklist
- Review Anthropic's current commercial terms for retention and training; commercial and API usage has different defaults than consumer chat, but confirm in writing for your plan.
- Use a dedicated read-only Katana key for the MCP server.
- Run the MCP server locally for personal use, or behind authentication if shared.
- Remove or mask customer personal data from tool outputs unless required.
- Log tool calls so you can reconstruct what Claude saw for any answer.
Ways to get it done
Upload exports to Claude
Effort: Minutes
One-off analysis with no setup.
MCP server over the Katana API
Effort: Days to a couple of weeks
Live questions from planners using Claude Desktop or another MCP client.
Claude on Bedrock or Vertex with Netray
Effort: A scoped build
Company governance requires a specific cloud account, region, logging and approvals.
Pitfalls
- !Giving Claude too many tools so it picks the wrong one; keep the set small and descriptive.
- !Skipping the system prompt, which leaves Claude guessing at your location and safety stock conventions.
- !Returning full API payloads and wasting context on fields nobody needs.
- !Assuming a working chat demo equals a safe write path.
- !Ignoring API rate limits during long multi-step analysis.
Frequently asked questions
Does Katana have an official Claude connector?
Check the Claude connector directory and Katana's integrations list for the current status. If none exists, an MCP server you run over the REST API is a straightforward alternative, and the same tools work with other MCP clients too.
Should I use Anthropic directly, Bedrock or Vertex?
Use Anthropic directly if you want the newest features and simple billing. Use Bedrock or Vertex AI if your organisation already centralises security, billing and regions on those clouds. The Katana tool layer stays the same either way.
Can Claude run on-prem for Katana data?
Claude is a hosted model, so not on your own servers. Katana is also cloud only. If you need a private model, use an open-weight one behind the same tools. Netray can set that up and keep prompts and logs in your environment.
Can Claude change inventory in Katana?
Only if you build and enable a write tool. We suggest starting read-only, then adding draft proposals such as a suggested purchase order, with a human approving before the API write is executed.
Is this better than Katana's own reports?
Katana's reports are better for standard views and are always consistent. Claude is useful for open questions, narrative explanations and combining Katana with other documents. Use both and let Katana remain the source of truth.
Want Claude working on your Katana?
A free 30 minute assessment: we map the integration path for your version and setup, and where it should run.
Get a free assessment