MCP server

Connect Claude, ChatGPT, Cursor and other MCP clients to your QuantCX workspace: search and read tickets, conversations and knowledge, and reply to or update tickets.

QuantCX runs a remote Model Context Protocol server, so a member can work with the Desk from the AI client they already use: Claude (claude.ai, Claude Desktop, Claude Code), ChatGPT, and developer tools such as Cursor. Nothing to install.

<code>https://app.quantcx.com/api/mcp</code>

MCP Streamable HTTP over one endpoint: POST with JSON-RPC, JSON responses. The server is stateless; GET and DELETE answer 405.

Requirements

The workspace needs API access (Premium plan or above). Tickets and conversations also need the Desk to be included in the plan.

Authentication

OAuth (people)

Clients that support OAuth discover QuantCX's own OAuth 2.1 server automatically (/.well-known/oauth-protected-resource). When you connect, you sign in to QuantCX, pick one workspace, and choose read or read and write. The AI app never receives your QuantCX password or session. Tools then run as you, with your role and permissions.

  • Access tokens last 30 minutes; refresh tokens 30 days and rotate on every use.
  • Disconnect any time under Connected apps in your account settings. The connection stops working from the next call.

API keys (scripts)

A workspace API key (Authorization: Bearer qcx_live_…, read scope) works too, for headless scripts and CI. A key has no person behind it, so it gets the read-only tools that the REST API already covers: search, fetch, helpdesk_overview, list_tickets, get_ticket and get_conversation. Knowledge, KPIs, knowledge gaps and all writes need OAuth.

Tools

ToolAccessAPI keyWhat it does
searchreadyesSearches tickets (subject, #number), conversations (subject, contact) and knowledge (OAuth only). Up to 20 results.
fetchreadyesReads ticket:<uuid>, conversation:<uuid> or knowledge:<uuid>. Text capped at 20,000 characters.
helpdesk_overviewreadyesOpen tickets by status and priority, unassigned, chats waiting, oldest open tickets.
kpisreadnoConversations, AI-resolved, hand-off, CSAT, median first response and resolution for 7d, 30d or 90d, with change vs the previous period.
list_ticketsreadyesFilter by status, priority, assignee (me, unassigned or an email), query; limit 1-50, cursor-paginated.
get_ticketreadyesA ticket by number or id, with its latest timeline events (1-100, default 30).
get_conversationreadyesA conversation with its latest messages (1-200, default 50).
search_knowledgereadnoThe passages the brand's AI would answer from. Uses AI credits like Try a question.
list_knowledge_gapsreadnoOpen knowledge gaps: question text and counts only.
reply_to_ticketwritenoReplies on a ticket as you (plain text, up to 20,000 characters). This sends to the customer.
update_ticketwritenoChanges status, priority or assignee (me, none or a member's email).
add_knowledge_snippetwritenoAdds a text or Q&A snippet to a brand's knowledge (owners, admins and managers).

Where a tool takes brand, pass a brand id or exact name; it is optional when the workspace has one brand. Every result is capped at 60,000 characters; cut lists say truncated: true. API keys, tokens and integration secrets are never returned.

Connecting a client

  • Claude.ai and Claude Desktop: Settings > Connectors > Add custom connector, name QuantCX, URL https://app.quantcx.com/api/mcp, then Connect, sign in, pick the workspace and access, and Allow. On Team and Enterprise an owner adds it once under Organization settings > Connectors.
  • ChatGPT: Settings > Apps & Connectors > Advanced > enable Developer mode (needed for write tools) > Create, URL https://app.quantcx.com/api/mcp, authentication OAuth. Deep research uses search and fetch.
  • Claude Code: claude mcp add --transport http quantcx https://app.quantcx.com/api/mcp, then /mcp > quantcx > Authenticate.
  • Cursor: add the server to ~/.cursor/mcp.json or .cursor/mcp.json; Cursor runs the OAuth flow on first use.
<code>{ "mcpServers": { "quantcx": { "url": "https://app.quantcx.com/api/mcp" } } }</code>

With an API key instead (read-only):

<code>claude mcp add --transport http quantcx https://app.quantcx.com/api/mcp \ --header "Authorization: Bearer qcx_live_…" curl -s -X POST https://app.quantcx.com/api/mcp \ -H "Authorization: Bearer $QCX_KEY" -H 'content-type: application/json' \ -H 'accept: application/json, text/event-stream' \ -d '{"jsonrpc":"2.0","id":1,"method":"tools/call","params":{"name":"helpdesk_overview","arguments":{}}}'</code>

Rate limits

Per connection or API key: 120 tool calls per minute, of which at most 20 writes. Over the limit, the tool answers with an error saying when to retry, so the model can back off. initialize and tools/list are not counted.

Errors

SituationHTTPBody
No or malformed bearer token401{"error":"invalid_request"} with WWW-Authenticate
Invalid, expired or revoked token401{"error":"invalid_token",…} with WWW-Authenticate
Plan without API access403{"error":"insufficient_plan",…}
Unknown tool200JSON-RPC error -32602
Tool refused (scope, role, plan, validation, rate limit, not found)200result.isError: true, with an explanation

Audit and data protection

  • Every tool call is logged with the workspace, the person or API key, the client, the tool, the outcome and its duration (ids only, never bodies or queries). Owners and admins can read their workspace's log. Connecting, disconnecting and every write are also recorded in the workspace audit log.
  • Data a member reads through MCP (tickets, conversations, contact names and emails, knowledge) is sent to the AI client they chose and processed under that provider's terms. The consent screen says so before anything is approved.