What is MCP, in two minutes
Before MCP, every connection between an AI assistant and a system was its own development project. MCP defines one language: the server announces which tools it has and what each one takes, and the assistant, whether Claude, ChatGPT, Gemini, Grok or GitHub Copilot, decides when to call them during a conversation.
In Priority it looks like this: you ask "who were our ten largest customers this quarter?", the assistant calls the server's analytics tool, and the server reads from Priority and returns an answer that says what it includes and from what time the data is.
How it works
- Your question"Who were our ten largest customers this quarter?"
- asksAI assistantClaude, ChatGPT, Gemini, Grok or GitHub Copilot
- calls a toolMCP for Priority
- Who is asking?Each user signs in for themselves, with their own permissions
- Semantic searchFinds the right tables, and an exact name comes firstPaid add-on
- Table cacheA copy kept for your account alone, refreshed every hourPaid add-on
- what the copy does not holdPriorityYour company's data
- returnsThe answerSays what it includes and from what time the data is
How do you connect?
Prepare Priority
Set up an API user in Priority with the permissions the server needs. Toward Priority the server works through it, with its permissions only.
Add the server to the assistant
In Claude, ChatGPT or Copilot, add an MCP connector with the server's address.
Each user signs in
Sign-in is personal, and each user's read, write and module permissions decide what they see and what they can write.
Ask
From then on, a question about the company's data is answered from Priority.
What should you check in an MCP server before giving it ERP access?
Before you connect an MCP server to the system that holds your finances, customers and stock, these are the questions to ask of it, and how MCP for Priority answers each one:
| What to check | Why it matters | In MCP for Priority |
|---|---|---|
| An identity for each user | So permissions and the log belong to a person, not to "the system" | Each user signs in for themselves, with read, write and module permissions |
| Approval before writes | An AI assistant can make mistakes | Writing is off by default, and in a conversation every operation is shown for approval first |
| No deletion | A mistake you can fix beats data that is gone | No tool in the catalog removes Priority data |
| An audit log | To know who asked and who changed what | Every tool call is logged, chained with SHA-256 hashes |
| Isolation between organizations | On a shared server, your data is only yours | Encrypted credentials per organization, and every query filtered by organization |
| Where it runs | Where the data is and who processes it | Google Cloud, me-west1 in Tel Aviv. The AI features, when turned on, process content outside Israel |
What else does MCP for Priority do?
Beyond the safety checks, these are what make MCP for Priority work with your Priority, not just connect to it:
- It knows your Priority: it reads your company's own data dictionary, custom fields and screens included, and searches it in Hebrew or English.
- Semantic search, an optional paid add-on: ask for "customer invoices" and it finds the right tables, and an exact name comes first. It is an AI feature, so it works only once your organization turns the AI features on.
- Table cache, an optional paid add-on: the tables you ask about most answer from a copy kept for your account alone. The copy refreshes every hour and reads only what changed, and every answer from it says what time its data is from.
- Analytics that say what they counted: profit and loss, sales trends, customer churn and stock movement. Every summary states what it includes, and a partial answer is marked partial.
- Every AI assistant: Claude, ChatGPT, Gemini, Grok and GitHub Copilot connect to the same server, with the same permissions and the same log.
Beyond questions
The same server is the base of AI agents. An agent preparing a quote or a collection reminder calls the same tools, with the same permissions and the same log. n8n has a dedicated endpoint in it, so automation scenarios use the same tools too. An automation that runs with no person in front of it, such as n8n, skips the approval step only with an API key that was given write access and explicit permission to skip it, and that is off by default.
Common questions
- What is an MCP server?
- A server that implements the Model Context Protocol: it gives an AI assistant a list of tools, such as search, analytics and creating records, and the assistant calls them during a conversation.
- Which AI assistants work with MCP for Priority?
- Claude, ChatGPT, Gemini, Grok, GitHub Copilot and any other client that supports MCP. n8n has a dedicated endpoint, so the same tools are available to an automation that runs with no person in front of it.
- Do you need to install anything?
- Not on your side, and not inside Priority either. MCP for Priority is a managed service: you add its address to the assistant, and each user signs in for themselves. On the Priority side, cloud or an on-premises server reachable over HTTPS, you need an API user with an API license and a personal access token (PAT).
- Does it work with on-premises Priority, or only the cloud?
- Both. MCP for Priority connects to Priority's REST API in the cloud or on a local server reachable over HTTPS, and nothing is installed inside Priority. An on-premises server behind a firewall allows the service's fixed egress addresses.
- Can MCP for Priority write to Priority?
- Yes, when writing is turned on: orders, invoices, customers and remarks. Writing is a paid add-on and is off by default. In a conversation every operation is shown for approval first, and before an update the server checks nobody changed the record since the preview.
- Can MCP for Priority delete records?
- No. It has no ERP delete tool at all. It reads, and when writing is turned on it creates and updates, always within the API user's permissions.