עברית
The MCP surface of the Priority Gateway

Your ERP, inside your AI tools

Ask questions, run reports, create and update records — from Claude, ChatGPT or Gemini. Every person connects with their own token, and you can revoke it in seconds.

Ships in every gateway Three tools, no hardcoded schema Read-only SQL channel

Priority Gateway v1.55.x · endpoint POST /mcp

Claude · Priority (MCP)
You

Which customers are over 30 days late?

Claude
priority_discover(form: "FNCITEMS") → 34 fields · key ACCNAME · open-balance columns found
priority_action(query, "FNCITEMS", filter: { FNCPAY: "N", FNCDATE: "<2026-07-09" }) → 41 rows · 0.9 s

Four customers are past 30 days, $412,900 open:

ACME Industries$186,40047 days
Bar-Lev Logistics$121,05038 days
Nadav Plastics$74,30034 days
Tzur Metals$31,15031 days

Two tabs, one endpoint: a question answered from live data, and a write that stops for a human.

Three reasons AI and Priority don't meet on their own

The model can't see the ERP

An LLM reads what you paste into it. Without a tool surface, every Priority question ends in a screenshot, a CSV export, or a wrong guess.

Generic connectors don't know your Priority

Every install has its own forms, columns and rules. A connector with a schema baked in goes stale the morning somebody adds a column.

Letting AI write is the part that scares people

Reading is easy to approve. Writing is not. The question a finance lead actually asks is: whose rules apply when the model presses save?

One endpoint answers all three.

Three tools: discover, act, or sit down and edit

The gateway consolidates 23 narrow tools into three dispatchers, with no functionality lost. A model learns the surface once and uses it on any Priority.

priority_discover

Find out what exists

Forms and procedures by name, field metadata at any subform depth, dropdown values, credential checks.

  • Serves reference docs in-band: the SQLI syntax guide, the technology guide, the n8n node authoring spec.
  • Answers without a Priority connection where the doc alone is enough.

priority_action

Do it in one call

Query, create, update, delete. Run procedures and reports, fire form actions, upload files, batch several writes.

  • Queries route themselves between OData and WebSDK; you can override the choice.
  • Read-only SQL and entity dumps for the questions a single form can't answer.

priority_session

Edit like a person would

Open a form, set one field, watch Priority answer with warnings and cascading updates, then save.

  • Subform navigation and list browsing with filtered search.
  • A warning holds the save open until someone answers it.
discover → act

Progressive discovery is why this works on your Priority. The model asks what exists before it does anything, so nothing about your install is hardcoded on our side. Add a column to ORDERS this morning and the model sees it this afternoon.

Business questions, scoped before they're asked

"How much did we sell last month" is one wrong join away from a confident wrong number. So the model gets the right slice of your schema first.

STEP 1

The question

Anything with a total, a ranking, a breakdown or a date range.

STEP 2

Scoped context

Relevant forms, discriminator filters, join paths and lookup catalogs from the graph database of your own install.

STEP 3

The model writes it

It composes the query itself, against the dialect rules it was handed.

STEP 4

Validate

Dialect and column existence checked against the graph, before anything runs.

STEP 5

Run

Executed on the read-only SQL channel and answered in the conversation.

nl_context

The gateway never writes or rewrites the SQL. It scopes the schema, states the dialect rules, and checks what the model composed. The query stays the model's, and you can read it before it runs.

The map includes your customizations

That scoping comes from a graph database of your own Priority, built by BI Generator: every table, column, form and foreign-key path, with the business meaning of each field and the filters that make a form view what it is — the entities your implementer built included. The custom table behind your quality process and the columns somebody added to the order line are on the map next to the standard ones. A model working from general knowledge of Priority would not know they exist; a model handed this context can answer questions about them on the first try.

What it may do, and what it has to ask for

The controls are the ones Priority already has. The gateway's job is to expose them, not to reinvent them.

When the model…What actually happens
reads data It runs as the Priority user pinned into the token — the same permissions that person has in the Priority client, no more.
writes SQL The SQL channel is read-only by construction. INSERT, UPDATE, DELETE and the rest of the mutating verbs are rejected before the query leaves the gateway.
creates or updates a record The write goes through Priority's own form, so Priority's validations, business rules and triggers all run. Nothing writes straight to a table.
trips a Priority warning The warning text is surfaced into the conversation instead of being swallowed. Someone confirms or cancels, and the save waits.
hits a confirmation dialog You choose per call: confirm it, cancel it, fail the call, or hand the dialog text to the model and let it decide out loud.
picks the wrong protocol The reference docs are part of the tool surface. The model looks up which technology fits before it guesses, and gets a correction hint when a query hits a known wall.

Three ways in

Same endpoint, same tokens. Pick the one your people already use.

claude.ai

Paste the URL and connect. OAuth 2.1 with dynamic client registration and PKCE; the client discovers the auth server on its own from the gateway's challenge.

# Add a custom connector
https://your-gateway/mcp
OAuth 2.1 survives restarts

Claude Desktop

Download the connector bundle from the licensing portal and open it. No config file to hand-edit, no JSON to get wrong.

# One click, from the portal
priority-gateway.mcpb
one-click bundle

Any MCP-capable client

ChatGPT and Gemini reach the gateway through their MCP support, and so does anything else that speaks the protocol over HTTP. Build your own agent on the same three tools.

# Streamable HTTP
POST /mcp
Authorization: Bearer gct_…
ChatGPT Gemini custom agents

One token model for the whole surface

The token that clears the REST API clears MCP too. Nobody types a Priority password into an AI tool.

pk_ gateway key server-to-server, per customer gct_ connector token one person, one company Priority Gateway Verifies the signature offline Pins server · user · company · language Checks revocation on every request REST API apps, scripts, n8n flows MCP endpoint POST /mcp AI connectors claude.ai · Claude Desktop

What a connector token carries

A gct_ token pins the Priority server, the user, the company and the language it was minted for. Mint it at the gateway's connect page or in the licensing portal; the default life is 365 days. A leaked token cannot be pointed at a different tenant.

Revocation, in seconds

Revoke a person from the portal or the API and the next request fails. Repoint a server and every token on it is revoked with it, so a stale token can't keep dialing the address it was minted against.

For an internet-exposed gateway

Turn on consent gating and the OAuth flow stops asking politely: it requires a pasted connector token, which binds every AI session to a real Priority identity rather than to whoever found the URL.

Restarts don't disconnect anyone

Issued connector bearers live in a persistent, encrypted store. Redeploy the gateway and connected clients keep working — and the store file on its own is useless to anyone who takes it.

The short version

Four facts worth checking before you wire anything.

3 ← 23

Dispatcher tools, consolidated from 23 narrow ones with nothing lost.

Ships in every gateway

Not a separate install, a separate licence, or a separate service to run.

Survives restarts

Connected clients stay connected through a redeploy.

Read-only SQL

By construction, not by policy. Mutating verbs never reach Priority.

Where the endpoint lives

The /mcp endpoint your client connects to runs on Amazon Web Services, with Cloudflare in front of every request.

Hosted on AWS

The hosted gateway runs in Amazon's Tel Aviv region, deployed from a versioned pipeline. Secrets live in AWS Parameter Store, encrypted, never in a database or a config file.

Cloudflare in front

Every request arrives through Cloudflare: TLS, DDoS protection, a web application firewall, rate limiting and bot filtering, plus caching and routing that shorten the trip from wherever your people are.

No open front door

The servers accept no inbound connections. They dial out to Cloudflare and traffic comes back down that tunnel, so there is no public port to find and nothing to scan. Admin surfaces sit behind a second gate.

Running the gateway yourself? MCP ships inside it. The same Cloudflare model gives your on-premises Windows service an endpoint an LLM can reach — without opening a port on your network.

Part of the Priority AI-Tools suite

One registration connects your team to Priority through AI. These products are already built on the surface this page describes.

Riding the same three tools

Priority Console — Dev

Describe a change, get working Priority code. chat.flow-chain-ai.com

App Generator

One sentence to a runnable app on your own Priority forms.

BI Generator

Dashboards from plain-language questions. bi.flow-chain-ai.com

Your own agents

The same tools, the same tokens, in whatever you build next.

See the whole suite on the portfolio hub, or read what sits underneath this surface on the Priority Gateway page.