How to use Claude with ECI E2 Shop System
Short answer
Use Claude with E2 Shop System by putting a read-only MCP server in front of the E2 SQL database and connecting it to Claude, or by calling Claude through the API, AWS Bedrock or Google Vertex from your own service. For export-controlled data, run an open-weight model locally instead.
Claude is strong at reading messy, long inputs: a 12-page customer PO with special terms, a vendor quote with odd line items, a chain of emails about a changed drawing revision. That fits job-shop work, where much of the real information lives in documents around E2 rather than inside it.
The cleanest way to give Claude live E2 context is the Model Context Protocol. An MCP server exposes a handful of named tools, such as get_job or list_open_po_lines, and Claude calls them as needed. Claude never gets a database connection, only the tool results you choose to return.
The other question is where Claude runs. Hosted Claude, Claude through a cloud provider you already contract with, and a private on-prem model all give different compliance outcomes, and the right one depends on the data.
What teams use it for
Read a customer PO against the E2 order
Give Claude the PO text and the E2 sales order via a tool and ask it to list mismatches in quantity, price, delivery date and special terms.
Review vendor quotes
Feed several supplier quotes for the same material and have Claude normalise units, lead times and minimums into one comparison for the buyer.
Explain a job's cost overrun
With read access to estimated versus actual hours by operation, Claude writes a short explanation the estimator can check.
Draft work instructions and traveler notes
From a customer spec and prior similar jobs, Claude drafts inspection steps and process notes for a person to review.
Triage a shared inbox
Claude classifies inbound emails as RFQ, PO, change request or shipping question and summarises each with the related E2 record.
ECI E2 Shop System integration surface
- SQL Server read-only views behind an MCP server: the primary live surface for jobs, orders, purchasing, inventory and labor. Verify table names against your version.
- MCP transports: stdio for a desktop client on the same machine, or HTTP-based transport for a shared service on your network.
- Documents and email: customer POs, drawings and vendor quotes live outside E2 and are best supplied as files or via a separate document tool.
- Reports and exports: Crystal-based reports and Excel export give quick data for one-off conversations without any server.
- Authentication boundary: E2 user rights do not carry over to a SQL login, so role rules must be re-implemented in the MCP layer.
How it works
- 1
Decide the hosting model
Options include Claude via Anthropic directly, via AWS Bedrock or Google Vertex under your existing cloud agreement, or a local open-weight model. Pick based on your contracts.
- 2
Build a small MCP server
Use an official MCP SDK in a language your team supports. Implement read-only tools backed by parameterised queries on curated views.
- 3
Scope the tools
Return only needed columns, cap row counts, and omit cost fields for roles that should not see them.
- 4
Connect to a Claude client
Register the server with the Claude client your team uses, or call it from your own app via the API tool-use interface.
- 5
Add documents
Supply POs and quotes directly in the conversation or through a document tool so Claude can compare them to E2 data.
- 6
Validate
Run a set of known questions with known answers and track accuracy before expanding access.
Example flows
PO versus sales order check
Trigger: A customer PO arrives by email.
- Buyer drops the PDF text into Claude.
- Claude calls get_sales_order for the referenced quote or order.
- It lists differences in price, quantity, ship date and terms.
- Order entry corrects E2 or contacts the customer.
Result: Mismatches are caught before the job is released to the floor.
Supplier quote comparison
Trigger: Three quotes are received for a bar stock purchase.
- Claude reads each quote and normalises length, grade and price per unit.
- It calls get_part_history for the last purchase price and vendor.
- It outputs a comparison table with lead time risk noted.
Result: The buyer decides faster with the assumptions visible.
Keeping data on-prem
Hosted Claude is not an air-gapped option. If your E2 data is covered by ITAR, CUI or customer NDAs, evaluate whether a cloud-provider deployment under your own compliance program is acceptable, or run an open-weight model such as a Llama, Qwen or Mistral class model locally with vLLM or Ollama. The MCP server design carries over unchanged because MCP is model-agnostic: the same tools can sit behind a local model. Many shops start with Claude on non-sensitive documents and move controlled workloads to a local stack. Check current vendor data-handling terms yourself, since they change.
Security checklist
- MCP server uses a read-only SQL login and never accepts raw SQL from the model.
- Per-tool allow lists and row limits; log every call with user and parameters.
- Do not send controlled technical data to any hosted model unless your compliance owner approves it.
- Authenticate the HTTP transport and keep the server on an internal network.
- Treat text inside emails and PDFs as untrusted input that could carry prompt injection; keep write tools out of the same session.
Ways to get it done
Claude chat with exports and pasted documents
Effort: Minutes to start.
Quick wins on non-sensitive document review.
In-house MCP server plus Claude
Effort: One to three weeks for a read-only first version.
Teams with a developer who can maintain a small service and whose policy allows hosted AI.
Netray-built private deployment
Effort: Scoped pilot over several weeks.
Controlled-data shops wanting local models, an E2-aware schema map and audit.
Pitfalls
- !Giving the model a general SQL tool instead of narrow named tools.
- !Assuming Claude knows your E2 schema; it must be given tool descriptions that explain each field.
- !Mixing read and write tools in a session that also processes untrusted email text.
- !Skipping the compliance review for cloud-hosted models because the pilot was small.
- !Not versioning tool descriptions, so behaviour shifts when someone edits them.
Frequently asked questions
Does Claude have a ready-made E2 connector?
Not that I can confirm. The practical route is a custom MCP server over read-only views of your E2 database. If a third-party connector appears, evaluate its permissions and hosting the same way.
What is MCP and why use it?
Model Context Protocol is an open standard for exposing tools and data to AI clients. Building one MCP server lets Claude, ChatGPT-compatible clients and other tools reuse the same E2 access with the same permissions and audit.
Can Claude run fully on-premises?
Hosted Claude models are not offered as a self-hosted air-gapped product in the way open-weight models are. For fully local operation, use an open-weight model and keep your MCP server unchanged.
Can Claude write back to E2?
It can propose changes, but writes should be gated by a human approval step and applied through a supported path. Start read-only.
Is Bedrock or Vertex better than the direct API?
If your company already has a cloud agreement, security review and billing with AWS or Google, using their hosting can simplify procurement and data governance. Capabilities and regions differ, so check current availability.
Want Claude working on your ECI E2 Shop System?
A free 30 minute assessment: we map the integration path for your version and setup, and where it should run.
Get a free assessment