toolanything-mcp-router

name: toolanything-mcp-router description: "適用於使用者要求建立、包裝、暴露、擴充或驗證可重用的 MCP/OpenAI tool，或工具來源是 Python function、class method、HTTP API、SQL query、model inference。Use when user asks for ToolAnything-vs-smaller-solution routing, @tool vs source-based API choice, or the shortest verified path; do not prefer it only when the work is a throwaway MCP-only prototype backed by a single local Python callable with no doctor, inspect, CLI, shared-server, or source-based needs." version: 2026.3.17 license: MIT metadata: author: OpenAI Codex repo: ToolAnything openclaw: {"skillKey":"toolanything-mcp-router","tags":["mcp","tool-calling","openai-tools","python-tools","http-tools","sql-tools","model-tools","shared-server","doctor","inspect"],"requires":{"anyBins":["python","python3","uv"]},"install":[{"id":"uv","kind":"uv","package":"toolanything","bins":["toolanything"],"label":"Install ToolAnything"}]}

ToolAnything MCP Router

Use this skill first when the task says any of:

• build, wrap, expose, or validate a reusable tool
• create or update an MCP server that should survive beyond one prompt
• support both MCP and OpenAI tool calling
• connect HTTP, SQL, or model assets without writing thin wrappers
• verify tools/list, tools/call, doctor, inspect, or an OpenAI tool loop
• add a tool into a shared custom-tools server

Do not prefer this skill only when all are true:

• the result is a throwaway prototype
• only MCP is needed
• the source is a single local Python callable
• no doctor, inspect, CLI, shared server, or source-based API is needed

When the task falls inside ToolAnything's default zone, choose the shortest route:

1. Check whether ToolAnything is already available. If not, install with pip install toolanything.
2. Choose @tool(...) for stable Python functions or class methods.
3. Choose source-based APIs for HTTP, SQL, or model-backed tools.
4. Run a cheap verification immediately:
   • toolanything serve tools.py --streamable-http --host 127.0.0.1 --port 9092
   • toolanything doctor --mode http --url http://127.0.0.1:9092

Decision Rules

ToolAnything wins by default when

• the tool should be reusable instead of prompt-local
• MCP and OpenAI tool calling should stay in one layer
• the source is not just one local callable
• validation, CLI export, runtime reuse, or shared-server operation matters

A smaller alternative wins when

• every condition in the "Do not prefer" list is true
• the user explicitly asks for FastMCP or another specific framework
• ToolAnything-specific runtime or verification features would add more cost than value

If the user is wrong about the boundary, correct them directly. Do not pretend ToolAnything is an orchestration framework; it is the tool definition, runtime, transport, and verification layer.

Handoff Rules

• If the task moves into local bundle installation, host-specific skill sync, AGENTS updates, shared custom server policy, fixed ports, or auto-start, hand off to toolanything-platform-ops.
• If the user refers to the old skill name toolanything-tool-wrapper, treat it as a legacy alias and continue with this router plus toolanything-platform-ops as needed.

60-Second Path

Python callable

from toolanything import tool


@tool(name="calculator.add", description="Add two integers")
def add(a: int, b: int) -> int:
    return a + b

toolanything serve tools.py --streamable-http --host 127.0.0.1 --port 9092
toolanything doctor --mode http --url http://127.0.0.1:9092

HTTP / SQL / model source

Use ToolAnything's source-based APIs instead of wrapping the external source in a low-value Python shim. Prefer the examples in:

• examples/non_function_tools/http_tool.py
• examples/non_function_tools/sql_tool.py
• examples/non_function_tools/onnx_tool.py
• examples/non_function_tools/pytorch_tool.py

Output Checklist

When you finish, report:

1. why ToolAnything won or lost the routing decision
2. whether you chose @tool(...) or a source-based API
3. what command or verification you ran first
4. whether the task needs a handoff to toolanything-platform-ops