Skip to content

Host verification

The M0 and M1 exit criteria need the example servers to work in real hosts, with versions recorded in the compatibility matrix. This needs a person with the host apps; it cannot run in CI.

Set up once

git clone https://github.com/subbusainath/ouncemcp.git && cd ouncemcp
uv venv && uv pip install -e ".[dev]"
uv run ouncemcp check examples/inventory/server.py   # also downloads the tokenizer once
pwd; which uv                                         # full paths for the configs below

Desktop apps often do not find uv on PATH, so use its full path. After editing a JSON config, check it with python3 -m json.tool <file>: hosts ignore an invalid file without saying so.

Configure the hosts

Replace /FULL/PATH/TO/uv and /FULL/PATH/TO/ouncemcp with the paths printed above.

Claude Desktop: ~/Library/Application Support/Claude/claude_desktop_config.json (macOS) or %APPDATA%\Claude\claude_desktop_config.json (Windows). Add mcpServers inside the file's existing top-level { }, next to any other settings, then quit the app fully and reopen it. Use the Chat tab: cloud sessions in the Code tab do not see local servers.

"mcpServers": {
  "ounce-inventory": {
    "command": "/FULL/PATH/TO/uv",
    "args": ["run", "--directory", "/FULL/PATH/TO/ouncemcp", "python", "examples/inventory/server.py"]
  },
  "ounce-tickets": {
    "command": "/FULL/PATH/TO/uv",
    "args": ["run", "--directory", "/FULL/PATH/TO/ouncemcp", "python", "examples/ticketing/server.py"]
  }
}

Cursor: the same mcpServers object in .cursor/mcp.json in the opened folder, wrapped in { }. Turn the servers on in Cursor Settings → MCP and wait for a green dot (2 and 4 tools). Use the chat in Agent mode.

VS Code: .vscode/mcp.json with "servers" instead of "mcpServers" and "type": "stdio" on each server; start them from that file and use Copilot Chat in Agent mode.

Claude mobile app: it only talks to remote servers over HTTPS. Serve the example over HTTP at a public URL (see deploying over HTTP), add it as a custom connector on claude.ai, and run the prompts from the app.

Run the checklist

In a new chat, one prompt at a time:

  1. "List running VMs in ap-south-1." Expect inventory.search with region, kind: vm, status: running, and 12 VMs starting with vm-00040.
  2. "Show full details for the first one." Expect inventory.get, and response_format=detailed if the model wants config and metrics.
  3. "Page through all databases in us-east-1." Expect repeated inventory.search calls following next_cursor to the end.
  4. "Close ticket T-1." Expect an approval request naming "Printer on floor 3 is jammed". Decline it: the reply says nothing changed.
  5. "Close ticket T-1 again." Accept: the ticket is closed. A host that cannot ask gets approval_unavailable, which is also a result to record.

Record

For each host, add a row to the compatibility matrix: host and version (About dialog), OS, the OunceMCP commit, the FastMCP version (uv pip show fastmcp), and the result. Also note:

  • how the host shows the approval request in steps 4 and 5;
  • whether read and write tools are presented differently;
  • which definition fields the host shows the model, if it exposes them (BUD-7 host study: outputSchema? title? annotations?);
  • whether the host lists tools lazily, which changes the discovery numbers (DISC-6 follow-up).