English Version | 中文版

A Python library for managing and executing tools in a structured way.

• Tool registration and management
• JSON Schema generation for tool parameters
• Tool execution and result handling
• Support for both synchronous and asynchronous tools
• Support multiple MCP transport methods: stdio, streamable HTTP, SSE, WebSocket
• Support OpenAPI tools

Full documentation is available at https://toolregistry.lab.oaklight.cn

Previously, the method ToolRegistry.register_static_tools was used for registering static methods from classes. This has now been replaced by ToolRegistry.register_from_class. Similarly, ToolRegistry.register_mcp_tools has been replaced by ToolRegistry.register_from_mcp, and ToolRegistry.register_openapi_tools by ToolRegistry.register_from_openapi. All old methods are planned to be deprecated soon, so please migrate to the new interfaces as soon as possible. For backward compatibility, the old names remain as aliases to the new ones.

Install the core package (requires Python >= 3.8):

Extra modules can be installed by specifying extras in brackets. For example, to install specific extra supports:

pip install toolregistry[mcp,openapi]

Below is a table summarizing available extra modules:

The openai_tool_usage_example.py shows how to integrate ToolRegistry with OpenAI’s API.

The cicada_tool_usage_example.py demonstrates how to use ToolRegistry with the Cicada MultiModalModel.

This section demonstrates how to invoke a basic tool. Example:

For more usage examples, please refer to Documentation – Usage

The ToolRegistry provides first-class support for MCP (Model Context Protocol) tools with multiple transport options:

# transport can be a URL string, script path, transport instance, or MCP instance.
transport = "https://mcphub.url/mcp"  # Streamable HTTP MCP
transport = "http://localhost:8000/sse/test_group"  # Legacy HTTP+SSE
transport = "examples/mcp_related/mcp_servers/math_server.py"  # Local path
transport = {
    "mcpServers": {
        "make_mcp": {
            "command": f"{Path.home()}/mambaforge/envs/toolregistry_dev/bin/python",
            "args": [
                f"{Path.home()}/projects/toolregistry/examples/mcp_related/mcp_servers/math_server.py"
            ],
            "env": {},
        }
    }
}  # MCP configuration dictionary example
transport = FastMCP(name="MyFastMCP")  # FastMCP instance
transport = StreamableHttpTransport(url="https://mcphub.example.com/mcp", headers={"Authorization": "Bearer token"})  # Transport instance with custom headers

registry.register_from_mcp(transport)

# Get all tools' JSON, including MCP tools
tools_json = registry.get_tools_json()

ToolRegistry supports integration with OpenAPI for interacting with tools using a standardized API interface:

registry.register_from_openapi("http://localhost:8000/")  # Providing base URL
registry.register_from_openapi("./openapi_spec.json", "http://localhost/")  # Providing local OpenAPI specification file and base URL

# Get all tools' JSON, including OpenAPI tools
tools_json = registry.get_tools_json()

When only providing a base URL, ToolRegistry will attempt a “best effort” auto-discovery to find the OpenAPI specification file (e.g., via http:///openapi.json or http:///swagger.json). If discovery fails, ensure the prov