orke support

Your service sends questions; AI on your PC or server answers.
Start with a simple assistant, then connect your actual service.

Hands-on guide · Create an AI → Test → Connect your serviceEnglish

What does this feature do?

It lets you create an AI that your service can call. Set its name, AI, and instructions so your service’s server can send questions and use the answers.

When an online store’s server sends “How long does shipping take?”, it answers according to your instructions. If it needs actual order status, connect a separate tool that retrieves that information.

Example: store serverYour service
→
The AI you configuredorke support
→
Returned to your serviceAnswer
This is different from “Support” on the website

Use “Support” in the top navigation for help using orke. Use “orke support” inside the app to connect AI features to your own service.

Three things to remember first

TermPlain meaningOnline store example
AIAn assistant with a defined role and response rulesAn AI named “Store assistant”
RoomA container for an ongoing conversationOne support conversation with a customer
Access keyA secret your service uses to call this AIThe key the store server sends with a question

Sending the next question to the same room continues the previous conversation. Start a new room for a new support conversation.

Example: create a store assistant

1

Check the AI’s sign-in

This uses the currently signed-in Claude or ChatGPT CLI. Check the account you want to use in Settings › Accounts. support uses the same current account as the terminal.

Processing questions consumes that account’s usage allowance. orke must remain running while receiving questions.

✓The AI’s CLI is installed and signed in.

2

Name the new AI and add instructions

Create a new AI in the orke support window. Name it “Store assistant” and choose the account and model you prepared.

Example for “Instructions”You are our online store’s assistant.
Shipping takes 2–3 business days after payment and costs KRW 3,000.
Give customers short, friendly answers in English.
You cannot look up individual order status yet. Do not guess; ask the customer to check with a support representative.

For your first exercise, start without an MCP server and with built-in tools turned off. Review the default response format and limits shown before creating the AI.

✓“Store assistant” appears in the AI list.

3

Save the access key shown once

The full access key appears once, immediately after creation. Copy it to a safe location for use by your service’s server, then press “I saved the key.”

If you lose the key, you can generate a new one, but the old key will stop working. Update your connected service to use the new key too.

✓You have saved the key for the actual service connection.

4

Ask a question in the Test tab first

In the “Test” tab, ask “How much does shipping cost?” You can use demo_01 as practice data in the “User information” field.

In the same test conversation, follow up with “How long does shipping take?” Tests run even when “Accept questions” is off. Test conversations do not appear in the regular conversation list and are deleted when you close this window. Turn “Accept questions” on before receiving questions from your actual service.

✓The answers match the KRW 3,000 shipping fee and the stated delivery time.

Send questions from your service

After testing in the app, open the “Connect” tab. You can copy the actual address and request examples. Call it from your service’s server rather than directly from JavaScript on a web page.

Practicing on your PC and connecting externally

On the same PC, use the default listener at http://127.0.0.1:48700/v1. An external service needs an HTTPS proxy address that connects to this listener. The HTTPS IP-and-port address used to open the orke workspace is separate from the support API address.

Create a room for the conversation first

The examples below use a macOS or Linux shell. Replace YOUR_ACCESS_KEY with the actual key. On Windows, adapt the requests from the Connect tab to the shell you use.

Create a room · Practice on your PC
curl --request POST "http://127.0.0.1:48700/v1/rooms" \
  --header "Authorization: Bearer YOUR_ACCESS_KEY" \
  --header "Content-Type: application/json" \
  --data '{"user":"demo_01"}'

The response’s room_id identifies the room. Replace YOUR_ROOM_ID below with this value.

Send a question
curl --request POST "http://127.0.0.1:48700/v1/rooms/YOUR_ROOM_ID/messages" \
  --header "Authorization: Bearer YOUR_ACCESS_KEY" \
  --header "Content-Type: application/json" \
  --header "Idempotency-Key: shipping-question-001" \
  --data '{"text":"How long does shipping take?"}'

The request waits until the answer is complete. A successful response has status: "done". With the default response format, read the answer from output.answer.

Fields to check in the response · Illustrative example
{
  "status": "done",
  "output": {
    "answer": "Shipping takes 2–3 business days after payment."
  }
}

Use a new Idempotency-Key for each new question. If you reconnect because you did not receive the answer to the same question, use the same key and the same text to prevent duplicate execution.

Connect MCP to look up real orders

The example above only knows the information in its instructions. To report actual order status, it needs a tool that retrieves order data.

An MCP server provides external tools for the AI to use. Once your store’s developer prepares an order-lookup MCP server, connect it through “Attach MCP server” in the AI settings. Test both the connection and questions.

The user value when creating a room identifies the person asking. Your service’s server must send a value for the signed-in customer, and the lookup service must return only orders that customer is allowed to access.

Keep built-in tools off for your first exercise

Enabling built-in tools increases what the AI can do. Claude’s shell tools run with the OS user’s permissions. ChatGPT’s shell uses the default read-only sandbox. Connect any needed external lookups through MCP tools suited to that purpose.

What to check during operation

SituationExplanation and solution
Sign-in requiredComplete CLI sign-in for the current account, then ask again.
Waiting for a usage limitCheck the account’s limit status. Distinguish queued questions from rejected questions.
MCP connection failedCheck that the MCP server is running and its address and authentication are correct, then test the connection.
Instruction changes are not appliedChanges apply to new rooms from that point onward. Test in a new room.
Room memory lostThe AI conversation for that room can no longer continue. Start in a new room.
Generate a new keyThe previous key becomes invalid. Replace the key in your service too.
Delete a roomCancels any question in progress and deletes the room’s working folder and linked AI conversation history.

API requests

RequestWhat it does
GET /v1/healthCheck the API listener’s status. This request does not require an access key.
POST /v1/roomsCreate a conversation room. user is optional and is set by your service’s server.
POST /v1/rooms/{room_id}/messagesSend a question and wait until the answer is complete.
GET /v1/rooms/{room_id}/messagesRead conversation history. Use after and limit to read subsequent pages.
GET /v1/rooms/{room_id}/messages/{message_id}Check the current status and result of one question.
DELETE /v1/rooms/{room_id}Delete the conversation and linked history. Any question in progress is canceled.

When creating a room or sending a question, send that AI’s access key in the Bearer header. Check success using both the HTTP code and the response’s status and error. Do not use an incomplete response’s output as a completed answer. The API does not provide CORS access for direct calls from web pages, so call it from your service’s server.

Before connecting a real service

After confirming that test questions succeed, choose the model, limits, and retention period you need. Check each provider’s guidance to confirm that your account’s terms allow processing other people’s requests. See the agent and service guidance on the support page for related explanations and sources.

Examples of HTTPS proxy configuration for external access are available in orke support settings. By default, the listener opens only on 127.0.0.1:48700. Entering the platform request URL in Settings also updates the address in Connect tab examples. Prepare the AI and listener on the PC or server that will actually run them.

Search documentation

The page could not be displayed. Reload ×