What is ocel-generator?

ocelgen is a tool for generating realistic synthetic workflow traces from multi-agent AI systems. A trace here is a detailed log of everything that happened during a task: which AI agents were involved, what prompts they received, what responses the language model returned, which tools were called, how long each step took, and what it cost. These traces mirror the observability data you would capture from a real production AI system. The tool is designed for teams who need trace data for testing, training, or benchmarking without running actual AI workloads at scale. It can produce over 1,500 events in under two seconds with no API key required. When you connect an AI model through an OpenAI-compatible endpoint, it fills each trace with realistic generated content, including actual LLM prompts and responses, tuned to the chosen scenario. Three workflow structures are supported: sequential, where one agent hands off to the next, supervisor, where one coordinator delegates to multiple workers, and parallel, where multiple workers run at the same time. Traces can also include intentional deviations such as wrong tool calls, skipped steps, and timeouts, each with ground-truth labels, which makes the data useful for training anomaly detection systems. Ten deviation types are available in total. Ten built-in domains are included, covering scenarios like customer support triage, code review, incident response, and financial analysis. Custom domains can be defined in a YAML configuration file and merged with the built-in set. Every generated trace passes through five validation layers checking things like schema compliance and temporal ordering, and follows the OCEL 2.0 standard so it loads directly into PM4Py, a process mining library. A pre-built dataset of over 17,000 events is also available on Hugging Face for teams who prefer not to generate their own. The full README is longer than what was shown.