MCP Server

The Qiscus Helpdesk MCP Server lets your team connect Qiscus Helpdesk to an AI assistant using the Model Context Protocol (MCP), an open standard for connecting AI assistants to external systems. Once connected, your agents and supervisors can work on tickets through natural language: summarise a conversation, draft a reply, triage a queue, update a status, or assign a ticket, without leaving the AI assistant they already use.

Unlike a shared API token, every MCP connection is tied to an individual Helpdesk login. The AI assistant acts as that person and sees exactly what that person is allowed to see, following the same role, division, and ticket visibility rules that apply in the dashboard. Every action it takes is recorded in the Ticket Log and in a dedicated audit trail.

A. Benefits of Using the MCP Server

• Work on tickets from the AI assistant your team already uses. Agents summarise conversations, draft replies, and make routine updates without switching tools.

• Ask about your queue in plain language. Supervisors can ask which tickets are unassigned, which are approaching an SLA breach, and how work is distributed across a division, then act on the answer immediately.

• No new permission model to manage. Access follows the roles, divisions, and visibility settings you already configured. An agent connecting an AI assistant cannot reach a ticket they could not open in the dashboard.

• Full visibility for Owners and Admins. Every tool call is recorded with the user, the AI client, the tool, and the result, and can be exported for review.

• Graduated control. You decide which categories of action are available, which roles may connect, and whether the whole workspace stays read-only.

B. What Your AI Assistant Can Do

Capabilities are grouped into four permission groups. A group is the unit of access: when a user connects, they grant the groups that are currently enabled on your workspace.

Inside those groups, the AI assistant is given a fixed set of tools:

Example requests

Once connected, you can ask your AI assistant things like:

• "Summarise ticket 4821 and tell me what the customer is actually blocked on."

• "Draft a reply to ticket 4821 in Bahasa Indonesia, apologetic but concise."

• "Which tickets in the Billing division have no assignee right now?"

• "Show me tickets that will breach their first response target in the next hour."

• "Set ticket 4821 to Pending, add the tag waiting-customer, and leave an internal note explaining why."

• "Assign ticket 4903 to Dina and raise the priority to High."

• "Create a follow-up ticket from closed ticket 4102 for the same customer."

• "Create a ticket from the Refund Request template for the customer on ticket 4821."

• "The vendor replied on the forwarded thread for ticket 4821. Reply to them that we will send the logs today."

C. Setting Up the MCP Server

Setup has two halves. An Owner or Admin configures the workspace once, then each person connects their own AI assistant.

1. Configure the workspace (Owner or Admin)

Go to Settings > Integration > MCP Server. From this page you control:

• Read-only mode. When enabled, only the Read group is offered, so connected AI assistants can look but cannot change or send anything. This is the recommended starting point for a first rollout.

• Permission groups. Turn Internal writes, Customer-facing writes, and Destructive on or off for the whole workspace. A group that is off is never offered to anyone.

• Who can connect. Choose which roles are allowed to connect an AI assistant.

• Connection URL. The MCP URL for your workspace, which you can copy and share with your team.

Changes take effect when you select Save changes. Reducing access applies immediately to everyone who is already connected, so if you turn a group off, remove a role, or enable read-only mode, connected assistants lose that access on their next request. Because this affects people who are already working, Helpdesk asks you to confirm before saving a change that reduces access. Widening access is not applied retroactively: users who are already connected keep the access they were granted until they connect again.

2. Find your connection URL

Your workspace has its own MCP URL in the form:

Owners and Admins find it on Settings > Integration > MCP Server. Agents and Supervisors find the same URL on Settings > Profile, in the Connected Apps tab, so they do not need access to the Integration settings to get started.

3. Connect your AI assistant

The MCP Server works with AI clients that support remote MCP servers over Streamable HTTP. The steps below cover the clients Qiscus supports.

Claude on claude.ai

1. Open Settings > Connectors.

2. Select Add custom connector.

3. Paste your workspace MCP URL and confirm.

4. Complete the authorisation described in step 4.

Claude Desktop

1. Open Settings > Connectors.

2. Select Add custom connector.

3. Paste your workspace MCP URL and confirm.

4. Complete the authorisation described in step 4.

Claude Code

1. Run the following in your terminal, replacing the URL with your own:

   claude mcp add --transport http qiscus-helpdesk https://<your-workspace>-helpdesk.qiscus.com/mcp

2. Start Claude Code and complete the authorisation described in step 4.

3. Run /mcp to confirm the server is connected.

4. Authorise the connection

When your AI client first reaches the MCP Server, it opens your Helpdesk login page in the browser.

1. Sign in with your Helpdesk email and password.

2. Review the consent screen. It shows which AI client is asking for access, your own identity, and the permission groups that will be granted.

3. Select Authorize.

Your AI client returns to its own window and the Helpdesk tools become available.

Access is granted as a whole: you accept the groups your workspace has enabled, and you cannot pick a subset. If you are signed in as the wrong person, use Not you? Switch account to sign out and sign in again in the same flow.

D. Managing Connections

Your own connections

Go to Settings > Profile and open the Connected Apps tab. You will see every AI client connected to your account, when it was connected, and when it was last active. Select Disconnect to end a connection.

You can have up to 10 active connections at the same time, which covers using several AI clients or several devices. If you reach the limit, disconnect one you no longer use, then start the connection again from your AI client.

Connections across the workspace (Owner or Admin)

On Settings > Integration > MCP Server, the Connections tab lists everyone in your workspace who has connected an AI assistant, which client they used, when they connected, when they were last active, and which permission groups they hold. You can revoke any connection from here.

Revoking removes the connection entirely. The person is not notified, and their AI assistant stops working against Helpdesk on its next request.

Connection status

A connection is either Active or Expired. This table shows what happens to a connection row for each event:

An Expired row stays visible with its connect date and last activity so you can still see the history, and it is cleaned up automatically 30 days later. Expired connections do not count towards your limit of 10.

To use an Expired connection again, start the connection from your AI client the same way you did the first time. There is no reconnect action inside Helpdesk, because only the AI client can begin an authorisation.

E. Audit Log and Ticket Attribution

Audit tab

On Settings > Integration > MCP Server, the Audit tab records every tool call made through the MCP Server, with the time, the user, the AI client, the tool, and whether the call succeeded or failed. You can filter the list and export the selected range as a CSV file.

Audit records are kept for 90 days.

Ticket Log attribution

Actions taken through an AI assistant also appear in the ticket's own history, attributed to the person who authorised the connection and marked as coming from an AI assistant. A reply drafted and sent through MCP is recorded as that agent's reply, so response metrics and reporting stay accurate.

F. Safety and Data Handling

• Permissions are checked on every request. A connection carries your identity, not a frozen copy of your access. If your role changes, your division assignment changes, or an Admin changes the workspace settings, the next request follows the new rules.

• Destructive actions are confirmed before they run. When the AI assistant calls merge, mark as spam, or delete, the server does not execute immediately. It first returns a summary of what would happen, including which tickets are affected and whether the action can be undone, and only runs after an explicit confirmation.

• Customer content is treated as data, not instructions. Text that came from a customer, such as an email body or a customer name, is labelled so the AI assistant does not follow instructions hidden inside it. Sending a message to an email address that has not appeared on the ticket requires the same confirmation as a destructive action.

• Usage is rate limited. Each user gets 60 read calls, 20 write calls, and 10 destructive calls per minute, with a daily ceiling of 5,000 calls, plus a workspace ceiling of 1,200 calls per minute. These limits exist to contain an AI client that retries in a loop. When a limit is reached, the AI assistant receives an error that says when it may try again, and the blocked call is recorded in the audit log.

• Some things are never exposed. The AI assistant can never read or change the MCP Server settings themselves, read or create API tokens, touch email integration credentials, change passwords or invitations, or permanently delete anything. The only deletion available through MCP is a move to Trash, which can be restored.

G. Limitations and Expected Behaviours

• Custom domains are not supported for the MCP URL. Use your *.qiscus.com subdomain.

• The connection uses your Helpdesk email and password. Signing in with Qiscus Omnichannel credentials is not supported, so accounts created automatically through Omnichannel must set a Helpdesk password first.

• Consent is all or nothing. A user accepts every permission group the workspace has enabled and cannot select a subset.

• The consent screen lists permission groups, not individual tools. Your AI client shows the full tool list after connecting.

• Knowledge Base articles are not available as tools.

• Ticket templates are available to read, not to apply. The AI assistant sees the templates you can use in Create Ticket, following the same visibility, and copies a template's values into the new ticket itself. Tickets created this way are not recorded as created from a template. If a template's form is inactive or deleted, the AI assistant needs to pick another form, just as you would in the dashboard.

• Bulk updates across many tickets are not available. Creating tickets in bulk is limited to 20 per call, and each item reports its own result.

• Analytics and workload aggregates are not available as tools. Ask about specific tickets and queues instead.

• Administrative configuration, such as SLA policies, automation, CSAT, user administration, and channel or email integration, cannot be changed through an AI assistant.

• Ticket data is not masked or redacted before it reaches the AI assistant.

H. Troubleshooting

I. Summary

The Qiscus Helpdesk MCP Server connects your workspace to the AI assistants your team already uses, with access bounded by each person's existing role and visibility, graduated controls for the Owner or Admin, and a complete audit trail of everything the assistant does. Enable it in read-only mode first, let your team connect their own clients, review the audit log, then widen access as confidence grows.