MCP servers

An MCP server in Zorvik answers like a real one, from settings instead of code. Use it to:

• try tools with a real AI app before you build them: write a description and a sample answer, connect Claude, Cursor or VS Code, and see whether the model picks the tool and fills in the arguments as you hoped,
• test an MCP client or an agent against answers you control, including failed calls and slow tools,
• stand in for a server that is down, not built yet, or costs money per call.

Create one

Open the Servers sidebar, click + (New server) and choose MCP server. It listens on 127.0.0.1, port 3004 (or the next port no other saved server uses), and comes with an example tool, resource and prompt. Press Start: the address it shows, such as http://127.0.0.1:3004/mcp, is the one to give MCP clients.

It speaks every transport:

Transport	Address	For
Streamable HTTP	http://127.0.0.1:3004/mcp	Current clients. Sessions (Mcp-Session-Id), answers as JSON, a GET stream for the server’s notifications, DELETE to end a session.
HTTP+SSE	http://127.0.0.1:3004/sse	Clients of the 2024-11-05 protocol: GET /sse, then POST /messages?sessionId=…
stdio	zorvik serve <workspace> "<server>" --stdio	AI apps that start servers as programs

It answers protocol versions 2025-11-25, 2025-06-18, 2025-03-26 and 2024-11-05 (the client’s, when it’s one of these), and ping, logging levels, resource subscriptions and completions (with no suggestions).

Tools

Part	In the file	Description
Name	name	What clients call. Unique among the enabled tools.
Description	description	What the tool does and when to use it: the model reads it to decide.
Title	title	A display name for people (optional).
Input schema	inputSchema	JSON Schema of the arguments, as JSON text (empty: any object). A call without a required argument is answered as a failed call that names it, like a careful server would.
Result	result	The answer’s text. JSON text is fine.
Output schema	outputSchema	JSON Schema of a structured result (optional). Then the result must be JSON: it is sent as text and as structuredContent.
Answer as a failed call	isError	Send the result with isError: true: the call worked, the tool “couldn’t do it”.
After	delayMs	Wait before answering (at most an hour), to test timeouts and progress.
On/off	enabled	Switched-off tools aren’t offered.

An unknown tool gets a JSON-RPC error (-32602 Unknown tool), as the protocol says.

Resources

Part	In the file	Description
URI	uri	Such as docs://readme. With {parts} it’s a template (users://{id}): it’s listed under resource templates, and any URI that fits is read, with the parts available as {{params.id}}.
Name, description	name, description	What clients list.
MIME type	mimeType	Default text/plain.
Text	text	The content.

Prompts

A prompt has a name, a description, arguments (each with a name, a description and whether it’s required) and messages: each a role (user or assistant) and a text. A missing required argument is answered with a JSON-RPC error.

Answers are templates

Results, resource texts and prompt messages can use:

Placeholder	Is
{{args.city}}	One argument of the call. Text is inserted as it is; numbers, objects and arrays as JSON.
{{args}}	All the arguments, as JSON.
{{params.id}}	A part of a resource template’s URI.
{{$uuid}}, {{$randomFirstName}}, …	Dynamic variables.
{{token}}	The active environment’s and the workspace’s variables.

The client’s values go in after variables are filled in, so a client sending {{token}} gets that text back, never the variable’s value.

Introduction and clients

Setting	In the file	Default	Description
Name clients see	serverName	the server’s name	serverInfo.name in the initialize answer
Version	version	1.0.0	serverInfo.version
Instructions	instructions	none	Sent with initialize: how to use the server. AI apps pass them to the model.
Endpoint path	path	/mcp	The Streamable HTTP endpoint (not /sse or /messages, which the older transport uses)
CORS	cors	off	Answer browsers on other origins, for web-based MCP clients and inspectors

The server’s Clients section shows its address and the zorvik serve … --stdio command to copy.

Connect an AI app

Most AI apps take MCP servers in a JSON configuration. Give them the running server’s address:

{
  "mcpServers": {
    "weather-mock": { "url": "http://127.0.0.1:3004/mcp" }
  }
}

or let them start it as a program over stdio (the zorvik command line comes with every download; see Install):

{
  "mcpServers": {
    "weather-mock": {
      "command": "zorvik",
      "args": ["serve", "/path/to/workspace", "Weather mock", "--stdio"]
    }
  }
}

Over stdio, only MCP messages go to standard output; the traffic log goes to standard error, as JSON lines with --json. See zorvik serve.

Traffic and live changes

Every message shows in the server’s Traffic panel: sessions opening and closing, each request with a short summary (tools/call get_weather, resources/read docs://readme) and each answer (result of tools/call get_weather); click one for its JSON. Settings that would confuse clients, such as two tools with the same name or a schema that isn’t a JSON object, show there as errors.

Edits reach a running server at once. When its tools, resources or prompts change, clients with an open stream are told (notifications/tools/list_changed, and the same for resources and prompts) and list them again; Zorvik’s own MCP client does.

Saved format

servers/Weather mock.yaml

name: Weather mock
kind: mcp
port: 3004
mcp:
  instructions: Use get_weather for the weather in a city.
  tools:
    - name: get_weather
      description: The current weather in a city.
      inputSchema: '{"type": "object", "properties": {"city": {"type": "string"}}, "required": ["city"]}'
      result: '{"city": "{{args.city}}", "forecast": "sunny", "temperatureC": 21}'
    - name: slow_report
      description: A report that takes a while.
      result: Done.
      delayMs: 5000
  resources:
    - uri: docs://readme
      name: readme
      mimeType: text/markdown
      text: "# Weather service"
    - uri: stations://{name}
      name: station
      text: "Station {{params.name}}: 21 °C"
  prompts:
    - name: plan_trip
      description: Plan a day out
      arguments:
        - { name: city, required: true }
      messages:
        - { role: user, text: "Plan a day in {{args.city}} that suits the weather." }

AI agents can build, start and read the traffic of MCP servers too (save_server with kind mcp): see Agent tools.