Skip to main content
Poixe supports tool calling (Tools / Function calling) in the mainstream protocols. You can think of it this way: while processing a request, the model triggers a “function call”, the system executes a tool, and the result is fed back to the model to continue reasoning. Note that some of the built-in tools offered by providers have their own separate billing mechanisms. Some of these tools are difficult or impossible to bill in Poixe’s aggregation gateway scenario, so we restrict them.
Principle: enable whatever can be enabled; only disable what cannot be billed or controlled.Function tools you define yourself (custom function calling) are not subject to these restrictions.

​
Scope of restrictions: provider built-in tools only

  • Available: your own function / custom tools (Poixe does not restrict them)
  • Restricted: provider built-in tools (Poixe may disable some of them)
The reasons are straightforward:
  • Upstream built-in tools are usually billed separately by number of calls / duration / resource consumption
  • An aggregation gateway can hardly provide stable, accurate, and explainable cost allocation
  • To avoid uncontrollable costs and billing disputes, we disable some tools

​
Current disabled list (by protocol)

Below are the built-in tool restrictions Poixe currently applies to each protocol/provider (excerpted from system configuration; may change as policies are updated).

​
1. OpenAI Responses: some built-in tools disabled

Under the openai_responses protocol, Poixe disables the following by default:
  • code_interpreter
  • computer_use_preview
  • file_search
  • image_generation

OpenAI Responses Official Documentation

Instructions for function calling with the OpenAI API
Tip: Under OpenAI’s Responses protocol, the commonly used web_search tool is supported. The system has adapted billing for this tool, with the same rules as the official ones.

​
2. Anthropic Messages: Code Execution disabled

Under the anthropic_messages protocol, Poixe disables Claude’s built-in Code Execution tool, specifically:
  • code_execution_20250522
  • code_execution_20250825
For Anthropic’s official documentation on the Code Execution tool, see:

Anthropic Official Documentation

Instructions for the code execution tool
Tip: Under Anthropic Claude’s Messages protocol, the commonly used web_search tool is supported. The system has adapted billing for this tool, with the same rules as the official ones.

​
3. Google Gemini Content: Code Execution / File Search disabled

Under the google_gemini_content protocol, Poixe disables:
  • code_execution
  • file_search
For Google Gemini’s official documentation on the Code Execution / File Search tools, see:

Google Gemini Official Documentation

Instructions for function calling with the Gemini API
Tip: Under Google Gemini’s Content protocol, the commonly used google_search and google_maps tools are supported. The system has adapted billing for these tools, with the same rules as the official ones.

​
Which tools can I use?

​
1. Tools not on the disabled list

Poixe supports normal use of all built-in tools that are not on the disabled list (billing rules are the same as the official ones). This includes, but is not limited to:
  • Web Search (OpenAI / Azure / Claude, etc.)
  • Google Search, Google Maps (Google Gemini)
Although mainstream tools are also billed separately upstream, Poixe has adapted them and allows their use.
In other words: mainstream search tools are currently available, while tools whose costs are harder to define, such as “execution / file / image generation / computer use”, are disabled by default.

​
2. Custom Function

Even if some built-in tools are disabled, you can still achieve the same functionality through custom function tools, for example:
  • Your own search service / internal knowledge base retrieval
  • Your own code execution sandbox
  • Your own image generation/processing service
  • Your own file lookup / database queries
Poixe places no restrictions on “functions you define yourself” (as long as your tool service itself is accessible and secure).