MCP integration

What the MCP server exposes

When you run qoc up, Qoc starts an MCP server alongside the desk process. The server registers two categories of objects: tools (callable actions) and resources (readable workspace files). Agents discover these through the standard MCP capability negotiation handshake.

Tools run inside the Qoc process with access to live venue connections. Resources are served directly from the workspace directory as UTF-8 text, giving the agent read access to the same files you see in your editor.

Exposed MCP tools

Exposed MCP resources

Pointing Claude Code at the Qoc MCP server

Claude Code reads MCP server configuration from its project or user-level mcp_servers block. Add a qoc entry pointing at the socket or HTTP address that qoc up prints on startup (default: http://localhost:4242/mcp).

After adding the config block and restarting Claude Code, the tools appear automatically in the agent's tool list. You can verify with /mcp in the Claude Code prompt.

Claude Code MCP config — project level

{
  "mcpServers": {
    "qoc": {
      "type": "http",
      "url": "http://localhost:4242/mcp",
      "description": "Qoc trading desk — tools and workspace resources"
    }
  }
}

Using Qoc tools inside an agent session

Once connected, the agent can call tools naturally in conversation. For example, asking "what is the current buying power?" will cause Claude Code to call snapshot_equity and return the live figure. Asking "place a limit buy for 5 shares of XYZ at 120" will cause it to call place_order, which writes the proposal file — it will not submit to the venue until you approve.

Because place_order is a write tool, Claude Code will prompt for your permission the first time it is called (subject to your allow-list settings). You can pre-approve it with /allowed-tools add qoc:place_order in Claude Code.

Using Qoc tools from a custom MCP client

import httpx, json

MCP_URL = "http://localhost:4242/mcp"

payload = {
    "jsonrpc": "2.0",
    "id": 1,
    "method": "tools/call",
    "params": {
        "name": "snapshot_equity",
        "arguments": {}
    }
}

resp = httpx.post(MCP_URL, json=payload)
result = resp.json()["result"]
print(json.dumps(result, indent=2))

Security scoping

The MCP server binds to localhost only by default and does not require authentication for local connections. It should not be exposed to the public internet without an authenticating reverse proxy.

Tool-level scoping is available in desk.toml: you can restrict which tools the MCP server advertises, for example to expose read-only tools (symbol_search, technical_analysis, snapshot_equity, list_positions) to an untrusted client while keeping place_order unexposed.

Restricting exposed tools

[mcp]
# Bind address. Never expose to 0.0.0.0 without an auth proxy.
bind = "127.0.0.1:4242"

# Allowlist of tools to advertise. Omit to expose all tools.
allowed_tools = [
  "symbol_search",
  "technical_analysis",
  "snapshot_equity",
  "list_positions",
  # "place_order" — omit to prevent order proposals via MCP
]

Related pages