Creating Custom Tools and Toolkits

MTP supports custom tools via plain Python functions and toolkit loaders.

This guide mirrors common patterns used in agent frameworks:

• Python functions as tools
• Grouping tools into a toolkit
• Registering toolkit into a ToolRegistry

1) Python functions as tools

Use @mtp_tool to attach metadata and keep definitions explicit:

pythonfrom mtp import Agent

@Agent.mtp_tool(
    description="Add two integers.",
    risk_level=Agent.ToolRiskLevel.READ_ONLY,
    cache_ttl_seconds=60,
)
def add(a: int, b: int) -> int:
    return a + b

If you omit input_schema, MTP infers a schema from function signatures/type hints.

2) Build a toolkit from functions

pythonfrom mtp import Agent

toolkit = Agent.toolkit_from_functions("custom", add)

This produces tool names like:

• custom.add

3) Register toolkit in registry

pythonfrom mtp import Agent

registry = Agent.ToolRegistry()
registry.register_toolkit_loader("custom", toolkit)

4) Use with agent

pythonfrom mtp import Agent
from mtp.providers import Groq

provider = Groq(model="openai/gpt-oss-120b")
agent = Agent.MTPAgent(provider=provider, tools=registry)
print(agent.run("Use custom.add with a=20 and b=22"))

Manual ToolSpec generation (advanced)

If you need explicit control:

pythonfrom mtp import Agent

spec = Agent.tool_spec_from_callable(add, namespace="custom")

Notes for robust tools

• Write precise docstrings/descriptions. The model reads them.
• Keep parameter names explicit and stable.
• Mark risk level correctly (read_only, write, destructive).
• For side effects, prefer explicit user confirmation policy.

Related:

• Storage and Sessions