Generate a Function Calling Schema

Describe a tool in natural language and generate a Function Calling JSON Schema. Runs locally — good for OpenAI tools and Agent inputs.

When you write an Agent, tool input has to be validatable. Describe "search products by keyword and price range," pick a platform, and you get a draft.

The platform option changes the wrapper: OpenAI tools, Gemini, Claude, MCP, or bare JSON Schema. The core is still names, types, and required.

You don't need an API Key. You get a definition — it will not call your backend.

How to write a useful description

Say what the tool does, which parameters exist, which are required, and what the enum values are. "Search something" yields a schema that is too wide.

What you must edit after generate

Add descriptions, make price a number, turn sort into an enum. The model decides whether to call the tool — and how to fill it — from those descriptions.

How to use it

Describe the tool and its main parameters in one sentence, or load a sample.
Pick a platform: OpenAI / Gemini / Claude / MCP / generic Schema.
Click Generate Schema and check that parameters were not all flattened to string.
Fill in required, enums, and descriptions. Remove internal fields that should not be exposed.
Copy it into your Agent config. For an MCP list, refine it on the MCP page.

When to use it

Add tools like search_products to a support or commerce Agent.
Turn an internal API into a tools list the model can call.
Export the same capability as OpenAI and Claude wrappers.
Produce the input Schema first, then write the real handler.

Keep in mind

Local rules generate the draft. Complex auth and pagination still need to be written by hand.
Don't put secrets or internal id rules in descriptions the model can see.
When you have many tools, keep each one narrow. One function should not do everything.

FAQ

Do I need an API Key?

No. This page only generates a Schema. It does not call a model or hit your APIs.

Can I ship this as-is?

No. The draft still needs descriptions, enums, and auth-related constraints.

How is this different from Structured Output?

This constrains tool input. Structured Output constrains the model's final JSON reply.

Is the description uploaded?

No.

Are parameter types accurate?

Common types are inferred from names and descriptions. Recheck numbers, booleans, and enums.

Can I generate several tools at once?

Write one tool at a time. Generate separately, then assemble the list.

When should I use the MCP page instead?

Use this page for OpenAI-style tools. Use the MCP page for an MCP Tool list.

Should I add examples?

They help the model fill arguments. After generate, add one or two valid input examples by hand.