How to get good answers from the HighCohesion MCP Server

How to phrase questions to the HighCohesion MCP Server so you get specific answers you can check.

The HighCohesion MCP Server lets an AI assistant, such as Claude, read your HighCohesion account and answer questions about your integrations. This page explains how to ask so that you get a specific, checkable answer. For ready-made prompts, see the HighCohesion MCP Server prompt library.

The short answer

Give the assistant the exact reference, say what you expected to happen, and tell it what you want back. Then check its answer against the event it points you to.

What the assistant can see

Once the connector is set up (see Connect Claude to HighCohesion MCP Server), the assistant can read your streams (what connects to what, their status and purpose), sources and destinations, transformations (mapping logic), the systems catalogue, entities and events, issues and their occurrences (each time an issue happened), and lookup tables. It never sees credentials and only reads your own organisation's data. You do not need to know any tool names; ask in plain English.

It cannot see inside your other systems. If a record never reached HighCohesion, the assistant can check the stream's rules but not the record itself, so tell it what you know about the record.

Six habits that get better answers

1. Use the exact reference

Copy the order number, PO or SKU from the source system, including any prefix such as # or a country suffix. A near match often returns nothing.

2. Name the stream or systems

If you know which stream or which two systems are involved, say so. "Our Shopify to warehouse orders stream" narrows the search and speeds up the answer.

3. Give a time window

"Since 9am", "last week" or "between 8 and 19 September" stops the assistant guessing and keeps summaries focused.

4. Say what you expected

"It should have reached the warehouse" or "the quantity should be 1, not 2" tells the assistant what to test. Without it, you get a description rather than a diagnosis.

5. Ask for the output you need

Ask for a table you can export, a two-sentence summary for a colleague, or a yes or no. The assistant will match the format.

6. Ask for the evidence

Ask for event IDs or issue IDs with every finding. You can then open them in the Control Panel and confirm what the assistant told you.

Before and after

Less useful

More useful

"Why is my order missing?"

"Order #5678 should have reached our warehouse this morning. Where has it got to, and if it stopped, why? Give me the event ID."

"What's wrong with inventory?"

"Summarise failures on our inventory stream since 9am, grouped by cause, with a count and three example SKUs for each."

"Is this right?"

"For order #1234, compare what HighCohesion received from Shopify with what it sent to our ERP. Show quantity, price and tax as a table."

"Explain the orders stream."

"Explain how our Shopify to ERP orders stream decides which orders to pick up, for someone who doesn't write code."

Follow up in the same conversation

The first answer is a starting point. Useful follow-ups include:

  • "Show me the data HighCohesion sent for that event." (Coming in the next few weeks.)

  • "Are other records affected by the same error?"

  • "Which mapping or lookup table set that value?"

  • "Put that in two sentences I can send to our warehouse."

Check the answer

AI assistants can be wrong or overconfident. Before you act on an answer, open the event or issue it names in the Control Panel. If the assistant says no event exists, check the reference first; a typo is the most common cause.

What it cannot do

The connector is read-only, so it cannot replicate an event or change a mapping. It can help you pinpoint exactly what to ask for, so your support request is already half-solved when you raise it. For the actions themselves, see How to fix and resend a failed event and Requesting a mapping change. When you contact support, paste the assistant's answer and the event IDs it gave you.