Connect an AI client over MCP

Dhanu ships a remote MCP server at https://api.dhanu.ai/api/pub/v1/mcp. Any client that speaks the Model Context Protocol over Streamable HTTP can connect with an API key and use the workspace as tools: hand work to a staff member, read the reply, run a workflow, search what the workspace already knows.

It is the same surface as the REST API, not a second one: every tool calls the service its REST endpoint calls, is gated by the same scope (tasks.create, staff.read, brain.search, …) and is metered and logged like any other public call, a tool the key may not call is simply not listed.

Connect

The server is stateless: send JSON-RPC requests to the URL with the key as a bearer token and no session id.

{
  "mcpServers": {
    "dhanu": {
      "type": "http",
      "url": "https://api.dhanu.ai/api/pub/v1/mcp",
      "headers": { "Authorization": "Bearer dhanu_sk_live_…" }
    }
  }
}

That shape works for Claude Desktop and Claude Code (claude mcp add --transport http dhanu https://api.dhanu.ai/api/pub/v1/mcp --header "Authorization: Bearer …"), Cursor, Windsurf and most other clients; use a test key while wiring it up and the tools answer with fixtures instead of touching the workspace.

What the client sees

initialize negotiates protocol version 2025-06-18 (or 2025-03-26 / 2024-11-05 if the client asks for one of those) and announces the tools capability. tools/list returns the tools the key’s scope allows, each with a JSON-schema input. tools/call returns the result as text and as structuredContent; a failed call, a task that does not exist, a scope the key lacks, invalid arguments, comes back as a tool result with isError: true, so the model can read the reason and recover. There are no resources, prompts or server-initiated streams, so a GET to the URL answers 405.

Safety

The key is the boundary. A tool never widens what the key can do, and everything a client sends becomes an ordinary task the workspace can see, with a staff run behind it, a create_task from an MCP client is treated exactly like a message from the API channel: non-interactive, and its text is framed as untrusted data. Scope the key to tasks.read and brain.search for a client that should only look.