Skip to main content

Intelligent Tools

CaseDesk deployments can call built-in tools automatically — web search and URL reading. When enabled, the model uses them during inference without any changes to your client code.

How it works​

When a request arrives at any of the three endpoints (OpenAI, Anthropic, or Gemini), the proxy checks whether tools are enabled. If they are, it runs an agentic loop: the model can call tools, receive results, and reason further before producing the final response. Your client sees only the final answer.

The loop runs for a maximum of 5 iterations per request.

Enabling tools​

Go to your deployment detail page → Intelligent Tools section. Toggle the tools you want and click Save.

ToggleToolBest for
Web Searchweb_searchCurrent events, updated guidelines, statistics, anything that may have changed since training
Web Readerweb_readerWhen the user asks about a specific URL or document

All three endpoint formats benefit from the same tools:

  • POST /proxy/{id}/v1/chat/completions (OpenAI)
  • POST /proxy/{id}/anthropic/v1/messages (Anthropic)
  • POST /proxy/{id}/gemini/v1beta/models/{model}:generateContent (Gemini)

Streaming behaviour​

When tools are enabled, responses are always buffered — even if your client sends stream: true. The model may need several roundtrips to produce the final answer, so the proxy buffers the result before returning it. Streaming applies to the final response only when no tool calls are needed.

MCP server​

For MCP-native clients (Claude Desktop, Cursor, Continue.dev), CaseDesk ships a standalone FastMCP server at mcp_server.py. Run it alongside the app:

# HTTP transport (default, port 8002)
python mcp_server.py

# stdio transport — for Claude Desktop config
MCP_TRANSPORT=stdio python mcp_server.py

Add to your Claude Desktop claude_desktop_config.json:

{
  "mcpServers": {
    "casedesk-tools": {
      "command": "python",
      "args": ["/path/to/byoc/mcp_server.py"],
      "env": {
        "MCP_TRANSPORT": "stdio"
      }
    }
  }
}

Available MCP tools: web_search, web_reader.

Custom tools​

You can register your own tools in code. They appear automatically in all three proxies and the MCP server:

from app.tools.registry import register_tool

register_tool(
    name='company_lookup',
    description='Look up a company by name in the internal CRM.',
    parameters={
        'type': 'object',
        'properties': {
            'name': {'type': 'string', 'description': 'Company name to look up.'},
        },
        'required': ['name'],
    },
    fn=lambda name: crm_client.lookup(name),
)

Call register_tool at application startup (e.g. in app/__init__.py after create_app). The tool is then available to any deployment on the same instance.