Responses
The OpenAI Responses API, for agents built on it. The OpenAI Agents SDK and Codex CLI use it by default. It runs on the same path as chat completions, and nothing is stored.
Request
curl https://uncensoredllm.xyz/v1/responses \ -H "Authorization: Bearer $UNCENSOREDLLM_API_KEY" \ -H "Content-Type: application/json" \ -d '{ "model": "venym/chimaera-1", "instructions": "You are blunt and concise.", "input": "Pros and cons of SQLite in production?", "max_output_tokens": 2000 }'Each request is turned into a chat completion and the answer is turned back, so limits, billing, retries and errors match the chat path exactly.
Fields
message items (roles user, assistant, system, developer; text parts, and images by URL), function_call and function_call_output items (tool results reach the model as text; an image in one becomes a note), and the reasoning items earlier responses returned.{"type": "function", "name", "description", "parameters"}. namespace tools (a named group of functions, as Codex sends) work too: a call into one comes back with its name and namespace. web_search turns on web search, unless it says external_web_access: false (Codex's default), which leaves it off. OpenAI's other hosted tools (file_search, computer_use_preview, code_interpreter, image_generation, mcp) and custom tools return 400 unsupported_tool.auto, none, required, {"type": "function", "name"} or {"type": "allowed_tools", "mode", "tools"}.false asks for one call at a time.max_tokens.reasoning_effort: low, medium or high.json_object or json_schema (name, schema, strict), sent as response_format.store: false.web_search, search_results.Engine extras the chat endpoint passes through (seed, stop, presence_penalty, frequency_penalty, top_k, min_p, repetition_penalty, chat_template_kwargs) pass through here too. Other unknown fields are ignored.
Response
{ "id": "resp_...", "object": "response", "created_at": 1767225600, "status": "completed", "model": "venym/chimaera-1", "output": [ { "type": "reasoning", "id": "rs_...", "summary": [], "content": [{ "type": "reasoning_text", "text": "..." }] }, { "type": "message", "id": "msg_...", "role": "assistant", "status": "completed", "content": [{ "type": "output_text", "text": "...", "annotations": [] }] } ], "usage": { "input_tokens": 42, "output_tokens": 380, "total_tokens": 422, "input_tokens_details": { "cached_tokens": 0 }, "output_tokens_details": { "reasoning_tokens": 0 } }, "incomplete_details": null, "error": null, "store": false}outputholds, in order: aweb_search_callwhen web search grounded the answer, areasoningitem with the model's thinking as written (Chimaera 1 thinks first), themessage, and anyfunction_callitems.statusiscompleted, orincompletewithincomplete_details.reason: "max_output_tokens"when the reply hit the cap. What was written is kept.usageis what you are billed on. Output tokens include the thinking.- The id carries the request id:
resp_Xis the call sent back withx-request-id: req_X.
Tool calls
When the model wants a function run, output has a function_call item with a call_id, a name and JSON arguments. Send the conversation back with the result as a function_call_output item:
import json, osfrom openai import OpenAI client = OpenAI(base_url="https://uncensoredllm.xyz/v1", api_key=os.environ["UNCENSOREDLLM_API_KEY"])tools = [{ "type": "function", "name": "get_weather", "description": "Current weather for a city.", "parameters": {"type": "object", "properties": {"city": {"type": "string"}}, "required": ["city"]},}] def get_weather(city): return {"city": city, "temp_c": 11, "sky": "rain"} # your real lookup here items = [{"role": "user", "content": "Do I need an umbrella in Bristol?"}]r = client.responses.create(model="venym/chimaera-1", input=items, tools=tools) items += r.output # the whole turn goes back: reasoning, message, callsfor call in (o for o in r.output if o.type == "function_call"): result = get_weather(**json.loads(call.arguments)) items.append({"type": "function_call_output", "call_id": call.call_id, "output": json.dumps(result)}) final = client.responses.create(model="venym/chimaera-1", input=items, tools=tools)print(final.output_text)Streaming
With stream: true the answer arrives as Responses events: an event: line naming the type, then data: with the event, numbered by sequence_number.
response.createdresponse.in_progressresponse.output_item.added reasoning itemresponse.content_part.addedresponse.reasoning_text.delta ... the thinking, in piecesresponse.reasoning_text.doneresponse.content_part.doneresponse.output_item.doneresponse.output_item.added message itemresponse.content_part.addedresponse.output_text.delta ... the answer, in piecesresponse.output_text.doneresponse.content_part.doneresponse.output_item.doneresponse.completed the whole Response, with usage- A function call is
response.output_item.added,response.function_call_arguments.deltaevents,response.function_call_arguments.done, thenresponse.output_item.done. - A reply that hits
max_output_tokensends withresponse.incompleteinstead ofresponse.completed. - A failure mid-reply sends an
errorevent, thenresponse.failedwithresponse.error.code:upstream_timeout(the model ran out of time),upstream_error(the connection dropped) orinference_error. The OpenAI SDKs raise anAPIError. - Refusals that come before the first event (a bad request, a context that is too long, no credit, a rate limit) are ordinary HTTP errors, stream or not. See errors.
- While the model is quiet (a long prompt, long thinking) the stream sends
: keepalivecomments, which SSE clients skip.
Nothing is stored
UncensoredLLM keeps no responses and no conversations, so each call carries the whole conversation in input. The Agents SDK and Codex already work this way.
previous_response_idreturns400 previous_response_not_found.conversation,backgroundand storedprompttemplates return400 unsupported_parameter.GET,DELETEandcancelon/v1/responses/{id}return404 response_not_found.reasoningitems you send back are accepted but not shown to the model again: they would be billed as prompt, and the model's own template drops old thinking.web_search_callitems are skipped too.
Agents SDK and Codex
OpenAI Agents SDK
import osfrom openai import AsyncOpenAIfrom agents import Agent, Runner, set_default_openai_client, set_tracing_disabled set_default_openai_client(AsyncOpenAI(base_url="https://uncensoredllm.xyz/v1", api_key=os.environ["UNCENSOREDLLM_API_KEY"]))set_tracing_disabled(True) # traces would go to OpenAI, with your UncensoredLLM key agent = Agent(name="Assistant", instructions="Be concise.", model="chimaera-1")print(Runner.run_sync(agent, "Write a haiku about recursion.").final_output)Codex CLI
# ~/.codex/config.tomlmodel = "venym/chimaera-1"model_provider = "uncensoredllm"model_context_window = 131072 # so Codex compacts the conversation before it outgrows the model [model_providers.uncensoredllm]name = "UncensoredLLM"base_url = "https://uncensoredllm.xyz/v1"env_key = "UNCENSOREDLLM_API_KEY"wire_api = "responses"Codex warns that it has no metadata for the model and uses defaults; the context window line is the one that matters. Its own tools work as sent, sub-agents included. Its web search is in cached mode by default, which UncensoredLLM leaves off. With web_search = "live" every call is grounded in web search, billed per search.