How to connect Claude to abas ERP
Short answer
Connect Claude to abas ERP by putting a Model Context Protocol server or a small API in front of EDP or the REST interface, then letting Claude call it. Hosting can be the Anthropic API, AWS Bedrock or Google Vertex, and an on-prem open-weight model is the alternative when data cannot leave.
Teams running abas often want an assistant that can answer a question like which orders are blocked by a missing purchase part, without opening five screens. Claude is a strong fit for reading structured results and long documents such as drawings notes and supplier contracts.
The sticking point is the same for any large language model: abas data is not exposed as a convenient SQL endpoint. You need a controlled bridge, and the Model Context Protocol is the cleanest way to give Claude that bridge with explicit tools rather than free access.
This page walks through the connection options, the hosting choices that matter for a European manufacturer, and what to do when your policy says no cloud.
What teams use it for
Order status through MCP tools
Claude calls an MCP tool to fetch sales and production order status, then explains the delay in plain language for the sales team.
Contract and certificate review
Claude reads supplier contracts or material certificates attached to abas records and extracts terms or test values for a buyer to confirm.
Explaining customisations
Developers use Claude with source access to understand legacy FOPs and infosystems before an upgrade.
Cost variance commentary
Given actual versus planned cost output from an infosystem, Claude drafts a variance explanation for the controller to edit.
Multilingual supplier communication
Claude drafts and translates correspondence between German, English and other languages while quoting exact order references.
abas ERP integration surface
- EDP selections wrapped by an MCP server: the usual path, with each MCP tool mapping to a bounded abas query.
- Java object layer: lets the MCP server read typed objects instead of parsing raw field text.
- REST interfaces on newer abas versions can back the MCP tools when available, which simplifies deployment. Verify what your release exposes.
- File exports and attachments: documents linked to records can be retrieved by the MCP server and passed to Claude for reading.
- abas user and role model: the MCP server authenticates as a service user, so Claude inherits only those permissions.
How it works
- 1
Define the tool list
Write down five or six read-only tools such as get_order_status, search_parts and get_open_po_lines, each with a strict input schema.
- 2
Build the MCP server
Implement it with an official MCP SDK in a language your team supports. Use stdio for a local desktop pilot and streamable HTTP for a shared deployment.
- 3
Authenticate to abas
Use a dedicated abas service user. Keep credentials in a secret store and restrict the server host's network access to the abas application server.
- 4
Choose where Claude runs
Pick Claude through the Anthropic API, AWS Bedrock or Google Vertex AI based on your region, contracts and data handling requirements.
- 5
Connect and test
Register the MCP server with your Claude client, test with real questions, and compare answers to abas screens.
- 6
Review and expand
Read the tool-call logs, remove fields not needed, then consider approved write tools.
Example flows
Why is this order late
Trigger: Sales rep asks Claude about a customer order
- Claude calls get_order_status with the order number
- It calls get_production_order for the linked job
- It calls get_open_po_lines for short materials
- It summarises the cause and the earliest realistic date
Result: The rep answers the customer with a grounded explanation and the references used.
Certificate extraction
Trigger: A goods receipt with a material certificate is booked
- MCP tool fetches the attached PDF
- Claude extracts heat number, grade and test values
- Values are shown next to the purchase order requirements
- Quality staff confirm or reject
Result: Faster incoming inspection with human sign-off retained.
Keeping data on-prem
Claude is delivered through the Anthropic API and cloud providers, not as software you can run inside a closed plant network. If your abas runs on-prem and policy forbids sending records out, keep the MCP server and gateway local and replace the model with an open-weight one served by vLLM or Ollama. Because MCP is an open protocol, the same server often works with local clients too. A hybrid is common: Claude for non-sensitive drafting and code explanation, a private model for anything pulling live abas records.
Security checklist
- MCP server uses a least-privilege abas service role and read-only tools first
- Choose a hosting route with the data processing terms and region your legal team accepts
- Never pass raw credentials or unfiltered tables through tool results
- Log every tool call with user, parameters and returned record identifiers
- Review tool descriptions for prompt injection risk when documents are read
Ways to get it done
Build your own MCP server
Effort: A few weeks for a useful read-only set.
You have developers and want a reusable bridge for several clients.
Use a community or partner connector if one exists
Effort: Short to install, longer to review.
You prefer not to build, but verify security and maintenance before relying on it.
Netray-built bridge with cloud or private model
Effort: Scoped pilot on one workflow.
You want the bridge, permissions and hosting decision handled together.
Pitfalls
- !Tools that are too broad let Claude pull far more data than the question needs. Keep tools narrow.
- !Documents can contain instructions that try to steer the model. Treat retrieved text as untrusted.
- !Cloud hosting terms differ between routes. Do not assume the Anthropic API, Bedrock and Vertex have identical contracts or regions.
- !Customised abas fields need documentation in the tool descriptions, or Claude will guess.
- !A shared service account hides who asked what. Pass the user identity into your audit log.
Frequently asked questions
Is there an official Claude connector for abas?
We are not aware of a vendor-supported one, and availability can change, so check with abas and the connector directory. In practice most teams build a small MCP server over EDP, the object layer or REST, which gives better control over permissions.
Can Claude run fully inside our network?
Not as a self-hosted model. Claude is accessed through Anthropic's API or a cloud provider. For a fully private deployment use an open-weight model locally with the same MCP or gateway design.
Should we use Bedrock or Vertex instead of the Anthropic API?
Choose based on where your company already has cloud contracts, regional requirements and procurement. Features and model availability can differ between routes, so test the exact model you need before committing.
Does Claude need a copy of our abas data?
No. With MCP it fetches only what a tool returns for each question. That avoids building a separate copy, though returned content is still sent to the model, so tool design matters.
Can Claude change abas data through MCP?
Technically yes if you build write tools. We recommend read-only first, then proposals that a person approves, with the server performing the write under an audited identity.
Want Claude working on your abas 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