Skip to main content
The MCP server lets an AI assistant use your workspace: answer from your agents’ knowledge, look through conversations, find leads, and add contacts or knowledge. It offers the REST API’s operations as tools.
It uses the streamable HTTP transport and authenticates with an API key sent as a bearer token. Make a key for the assistant under Developers → API keys. A Read only key gives the assistant only the read tools, which suits an assistant that should look but not change anything.

​
Connect a client

​
Claude Code

​
Cursor

In .cursor/mcp.json:

​
Other clients

Any client that supports remote MCP servers over streamable HTTP with a custom Authorization header can connect. Clients that only connect with OAuth, such as custom connectors in the Claude and ChatGPT apps, cannot connect yet.

​
Tools

Read tools are marked read-only and the three delete tools are marked destructive, so clients that ask before running a tool know which ones change things. A tool’s result is the same JSON the REST API returns. A failure comes back as a tool error the assistant can read, such as not_found: no agent with that id. Rate limits are the API’s: 120 requests a minute per key. Every response has an x-request-id header. Each call the assistant makes shows under Developers → Logs, named by its tool, with the arguments it sent. See Request logs.

​
Example prompts

  • “Using Wiredesk, what does our support agent say about refunds?”
  • “Find this week’s Wiredesk conversations about delivery delays and summarise what customers asked.”
  • “List the leads Wiredesk captured yesterday as a table.”
Starting an onboarding is not a tool: it is how a caller with no key gets one, so it is a REST call. See Agent-first onboarding. The key it returns works here.