How does OpenAI-compatible function calling work on Plugsky?

Key facts

Shape	OpenAI-compatible tools array and tool_calls response
Loop	The model returns tool calls, your code executes them and returns role=tool results
Parallel calls	Multiple tool calls in one response, dispatched concurrently
tool_choice	Force a named tool, auto, or none
JSON mode	response_format set to json_object for guaranteed JSON output
Strict schemas	strict:true forces the exact schema with no extra fields
Frameworks	LangChain, LlamaIndex, Vercel AI SDK and the OpenAI SDKs
Product status	Live

TL;DR

• Tools are JSON Schema objects with descriptions and required fields.
• Parallel tool calls and tool_choice are supported with the OpenAI shape.
• Strict mode forces the model to match your schema exactly.
• Count on 5-20 tools in practice; more makes selection unreliable.
• Combine with the agents API when you want retries, memory and audit logs.

How it works, step by step

1. Define tools as JSON Schema objects with clear field descriptions.
2. Send them in the tools array on /v1/chat/completions.
3. Detect tool_calls in the response and parse the JSON arguments.
4. Execute your function and append a role=tool message with the result.
5. Loop until the model returns a normal message with no tool calls.
6. Add tool_choice or strict mode where you need determinism.
7. Return tool errors as payloads the model can see and react to.

Try it yourself

Open the function calling tester →

The function calling loop

The loop has five steps: send the conversation plus your tools array, receive either normal text or one or more tool calls, execute the requested functions, append a role=tool message with each result, and continue until the model answers without asking for tools.

Because the shape matches OpenAI, the same loop works with the official SDKs and with frameworks such as LangChain, LlamaIndex and the Vercel AI SDK. Plugsky validates arguments against your schema before returning the call to your code.

Tool definition rules

Tools are JSON Schema objects. Use type: object with properties and required, and support string, number, integer, boolean, array, enum and nested objects. Descriptions on every field matter because the model uses them to decide what to call and when.

Keep the tool count deliberate. The live guidance is that 5-20 tools is the sweet spot; beyond that the model struggles to choose correctly, and every definition consumes context on each request.

The same platform exposes 30+ models, so you can run cheap models for routine tool calls and stronger models for planning steps.

Parallel calls, JSON mode and agents

Plugsky supports parallel function calls, so the model can return several tool calls in one response and your code can dispatch them concurrently. tool_choice lets you force a specific tool, while response_format set to json_object guarantees JSON output. With strict: true on a tool, the model must use exactly the schema you provide.

When you want managed retries, memory and audit logs, combine function calling with the AI agents API. Otherwise the loop above scales to any number of tools and steps.

Honest comparison

Feature	Plugsky	Roll-your-own dispatch	Non-compatible providers
Tool call shape	OpenAI tool_calls	You define the protocol	Varies
Parallel calls	Supported	You dispatch	Rare
JSON mode	response_format json_object	Prompt engineering	Varies
Strict schema	strict:true supported	You validate	Rare
Frameworks	Drop-in with OpenAI SDKs	Custom glue	Per-vendor SDK
Managed agents	Agent runtime with memory and logs	You build retries and logging	Varies