Sends the conversation to a model that operates the computer named by computer_id (clicks, types, runs commands) until it answers. The request and response follow OpenAI chat completions, plus computer_id and an orgo block, so the OpenAI SDKs work with base URL https://www.orgo.ai/api/v1. messages[].content is a string or OpenAI content parts (text, and image_url as an https or base64 data URL in user messages). Roles other than user and assistant are dropped, and the first remaining message must be from the user. With stream: true the answer arrives as chat.completion.chunk Server-Sent Events ending in data: [DONE]; only text streams, and the thread id is in the X-Thread-Id header. A stream that ends without a chunk carrying a finish_reason failed: it ends with an error chunk (message, type, and code for a refused own key) before data: [DONE]. Closing the connection stops the run at its next step; steps already taken are billed.

A run takes up to 250 model steps. max_steps raises that, up to 500; it never lowers it. A run that reaches its limit finishes with finish_reason: "length".

Billed from AI credits by the workspace owner, unless you pay with your own model key: x-openai-key for a model OpenAI serves (orgo-agent, which runs GPT-6.1 Terra), x-anthropic-key for one Anthropic serves (the claude-* models and orgo-agent-opus-5.5). Then no credits are used, orgo.cost_cents is 0, and orgo.own_key says whose key paid. With no model, the key picks it: x-openai-key runs orgo-agent (also when you send both keys), x-anthropic-key alone runs claude-sonnet-5. On a computer orgo-agent does not run on, both keys run claude-sonnet-5 and x-openai-key alone is refused (400 own_key_unsupported). orgo-agent asks OpenAI for its fast tier, so OpenAI bills your key at its fast tier rates. A key that can't pay for the model you name is refused too (400 own_key_model_mismatch), so a request that sends a key is never billed to credits. When the provider turns your key down, the error says so: 401 invalid_own_key (a bad or unauthorized key), 429 own_key_rate_limited (its rate limit or quota), 400 own_key_model_unavailable (a model the key can't use). Keys are used for the request only, never stored.