Your outlier library now takes orders from LLMs

What it does

PyOD is a mature Python library—established in 2017 with 38+ million downloads—that houses more than 60 anomaly detection algorithms for tabular, time series, graph, text, and image data. It exposes everything through a classic fit/predict API that remains unchanged in version 3. The new release adds ADEngine, an orchestration core that auto-selects and benchmarks detectors, and an agentic layer that lets Claude Code, Codex, or MCP-compatible LLMs drive multi-turn investigations via natural language.

The interesting bit

The library essentially split into three layers: the classic API for when you know exactly which detector you want, ADEngine for when you want the computer to choose, and an agentic front-end for when you want an LLM to run the whole conversation. That last layer is unusual for a statistical package; it turns PyOD from a model zoo into a kind of forensic assistant that can explain findings and suggest next steps.

Key highlights

• 60+ detectors across tabular, time series, graph, text, and image data, routed using benchmarks like ADBench, TSB-AD, BOND, and NLP-ADBench.
• ADEngine handles model selection, comparison, and assessment without breaking the legacy API.
• Agent integration includes an od-expert skill for Claude Code and Codex, plus an MCP server with ten stateless tools for knowledge queries, planning, and detection.
• Under the hood, training is accelerated by SUOD parallelization and numba JIT compilation.
• Enterprise track record includes Walmart pricing updates, Databricks insider-threat detection, and the European Space Agency’s OPS-SAT benchmark.

Caveats

• Stateful MCP tools for investigate and iterate are deferred, so the MCP server currently only exposes stateless operations.

Verdict

Data scientists and ML engineers who want a unified anomaly detection toolkit—and especially those already living in Claude Code or MCP agent stacks—will get the most from v3. If you just need one reliable outlier model and have no interest in LLM orchestration, the classic API still works exactly as it always has.