From zero to a live trace, in a few minutes

1. Open the demo at /try

Go to /try. It mints an anonymous, ephemeral session on the spot: no email, no password, no credit card.

Today that session points at the HubSpot mock: contacts, companies, deals, tickets and notes, through the same tools with the same names, descriptions and JSON schemas as the real HubSpot MCP server. The exact tool list is in the catalog. Every response is synthetic and deterministic. There's no real HubSpot account or data behind it.

2. Copy the connection config

The demo hands you a ready-to-paste snippet, shaped like a claude_desktop_config.json entry:

{
  "mcpServers": {
    "mockd-hubspot": {
      "url": "https://mock.mock-d.com/s/<session-token>/mcp"
    }
  }
}

• "url" is your personal Mockd endpoint for this run. It speaks MCP exactly like a real server would.
• The token in the address identifies your anonymous session. It's not an API key to any real service, just a pointer to your sandboxed session — treat it like a secret anyway.
• The entry name is yours to pick. It only shows up inside your client.

3. Paste it into your MCP client and run your agent

Drop the snippet into the mcpServers section of whatever MCP client you already use: Claude, Cursor, the OpenAI Agents SDK, LangGraph, or any other MCP client. No SDK to install, no code to change.

Then run your agent as usual. Every tool call it makes (search contacts, create a deal, add a note) gets routed to the Mockd endpoint and answered with realistic, deterministic data instead of touching a real HubSpot account.

4. Watch the trace, then inject a failure

Open the trace viewer and watch calls land as your agent makes them: the exact arguments it sent, the response it got back, the latency and the status of each call.

To see how your agent handles the rough edges, inject a failure on demand: a rate limit (HTTP 429), a timeout, an empty result, or extra latency. Watch whether it retries and recovers, right there in the same trace.

What's next: your own server, same trace viewer

Already have your own MCP server? Bring Your Own MCP will point the same endpoint at it and record every call in the trace viewer you just used. That one is being built and isn't usable yet. The roadmap has the current status.

What already works today: the trace viewer is the one place to see exactly what your agent did with every tool, and the mocks it runs against are live.