Definition
MCP server for ERPA server that publishes ERP records and operations as Model Context Protocol tools, so an AI assistant, such as Claude or an in-house agent, can read data and start actions without a custom integration for each AI application. MCP is an open standard for connecting AI applications to external systems, introduced by Anthropic in November 2024.
How is an MCP server different from an ERP API for AI agents?
Both give software access to the ERP, and an MCP server usually sits on top of the API rather than replacing it. What differs is who calls it.
| Aspect | ERP REST API | MCP server for ERP |
|---|---|---|
| Caller | Code a developer writes against fixed endpoints | An AI model choosing tools as it works |
| Described by | API reference for developers | Tool descriptions and input schemas the model reads |
| Discovery | Read in the docs at build time | Listed by the server at runtime (tools/list) |
| Permissions | API keys or user permissions | The same credentials; it should never do more than the API |
What does an MCP server for ERP need to do in wholesale distribution?
Reading a record is the easy part. Customers write ‘4x16 cu 100m’, their own code or another brand’s part number. Search the catalogue with that phrase and you get something back, often the wrong thing.
An MCP server that only exposes an ERP’s records gives an agent access, not answers: the agent still has to guess which catalogue item each line means, and that is where distributor quoting goes wrong. A useful MCP server separates lookup tools (does this code exist, which account is this) from identification tools that decide which product a line means and say how sure they are, with a match score. Write actions create drafts, so a confused agent’s worst case is a draft nobody approves, not a wrong price in a customer’s inbox.
ERP Agent’s MCP tools, by job
ERP Agent’s MCP server is built around that split. Its identification tools run the same agent as AI quote automation for distributors and handle vague descriptions, competitor codes and customers’ own part numbers.
| Job | Tools | What the agent gets back |
|---|---|---|
| Check the connection | erp_whoami | Which organisation the key belongs to and what it may do |
| Look up | erp_product_search, erp_customer_search | Catalogue rows by text or product code; customer records by name or number |
| Identify products | erp_product_match, erp_product_match_get, erp_match_list | One line per requested item with the matched product code and a confidence score from 0 to 100 |
| Draft a quote or order | erp_offer_create, erp_offer_get, erp_offer_list | A draft with product code, quantity, price and match confidence on every line, plus a review link |
| Send documents | erp_file_create | An upload link for RFQ documents, bills of material and spreadsheets |
Example: an AI assistant drafts a quote through MCP
A sales rep pastes a customer’s enquiry into an AI assistant connected to ERP Agent’s MCP server.
Enquiry from an electrical contractor: ‘Please quote for delivery next week.’
- 20 x copper cable 4x16 mm², 100 m drums
- 50 x wall plugs 6x15
| Step | Tool the agent calls | What comes back |
|---|---|---|
| 1 | erp_customer_search | The customer number for the company in the signature |
| 2 | erp_offer_create (request_type: offer) | An offer id; the run identifies each product and brings in the customer’s pricing |
| 3 | erp_offer_get, polled until the run finishes | Lines with product code, quantity, price and match confidence, plus a review link |
| 4 | None: the rep opens ERP Agent’s review | The rep checks the marked lines and approves the draft |
Illustrative example with a fictional customer and request; the tool names are real.
What should you check before connecting an AI agent to your ERP?
| Check | Why it matters | In ERP Agent |
|---|---|---|
| Scope | A key should reach only the tools it needs | MCP access needs the mcp:use scope on the organisation’s API key |
| Drafts, not bookings | A wrong guess must never reach a customer | erp_offer_create leaves a draft: nothing is sent to the customer or booked in the ERP until a salesperson approves it |
| Isolation | One key must never reach another company’s data | Keys are scoped to one organisation, stored only as a hash and shown once |
| Limits | A looping agent must not flood your ERP | 60 requests a minute and 5 concurrent runs per key by default |
| Data handling | Customer requests carry business data | Models run with zero data retention, and ERP Agent is GDPR-compliant |
How does ERP Agent’s MCP server work?
The first agentic AI layer for wholesale distribution ERP systems. ERP Agent runs a remote MCP server at https://erp-agent.com/api/v1/mcp with ten tools, over Streamable HTTP without sessions and on the same organisation-scoped keys as the quote automation REST API. Setup for each client is in the MCP server documentation.
Runs are asynchronous. A typical request takes 3–8 minutes, and fast mode handles requests of up to about 50 lines in 1–3 minutes; according to our case study, a sales rep spends on average 4–7 minutes per line by hand. The agent sends nothing to customers, and a draft reaches your ERP, CRM or other system only after a salesperson approves it.
The MCP server is part of the same agentic AI layer as the agent mailbox and the web app, and it works with any ERP (all ERP connections).
ERP Agent in numbers
- 95%
- average automation across our customers
- Over 95%
- accuracy in quote-line automation at LVI-WaBeK
- 0–100
- match score on every line, with lines under 80 marked for review
