Tool calling

Every model on the line calls tools. This is what an agent harness relies on, so it is the part worth knowing exactly.

In the OpenAI format

curl

curl https://uncensoredopencode.com/v1/chat/completions \
  -H "Authorization: Bearer $UNCENSORED_OPENCODE_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "glm-5.3-flash-uncensored",
    "messages": [{"role": "user", "content": "What is in the current directory?"}],
    "tools": [{"type": "function", "function": {"name": "list_files", "description": "List files in a directory", "parameters": {"type": "object", "properties": {"path": {"type": "string"}}, "required": ["path"]}}}]
  }'

The reply's message.tool_calls holds the call. Run it, then send the result back as a tool message with the matching tool_call_id.

What to know

parallel_tool_calls is passed on. If a model cannot take it, the request is sent again without it rather than failing.
Some models only accept tool_choice: "auto" and answer anything else with a 400 and the code tool_choice_unsupported. On others, naming one tool is honoured by offering only that tool.
In the messages format your own tools are translated both ways and the stop reason is tool_use. In the responses format, function tools are supported.
Reasoning counts as output. A tight max_tokens on a reasoning model can spend the whole budget thinking and return an empty answer with finish_reason: "length". Leave the limit off, or set it generously, in an agent loop.