Every run keeps a full record: the exact prompt the model saw, each tool call with its result, every memory change, and who wrote each note. A feature called Studio reads the live team's runs, notes, memory and schedules, finds why something keeps failing, and drafts a fix that the user reviews and publishes, with nothing changing until then. Model choice is per agent. Options include Claude, GPT, Gemini and Grok, plus open-weight models such as Qwen and Llama, and agents in one team can use different ones. Users bring their own keys, and platform keys on the hosted version are listed as coming soon. The site offers both self-hosting and a hosted option. It also positions the product as an open-source alternative to Grok bots. The space is crowded. Frameworks such as CrewAI and the OpenAI Agents SDK already cover multi-agent orchestration, and the OpenAI SDK includes built-in tracing and guardrails. One industry write-up notes that open-core projects sometimes gate orchestration, observability or fine-grained control behind commercial licenses. For founders shipping agent products, the pitch is auditability. When a multi-agent run fails, teams need to know which agent, prompt and step caused it. A free, self-hostable trace layer with mixed-model support lowers the cost of testing that. Its reliance on user-supplied keys and the early state of the hosted tier are the limits to check.