PyOD

PyOD is a Python library for detecting anomalies in tabular, time-series, graph, text, image, and audio data. Its documentation lists 61 detectors across those data types through one API. Users can choose the classic detector API or ADEngine, which profiles data, selects benchmark-backed detectors, runs detectors in parallel, calculates consensus scores, and reports diagnostics. ADEngine also works as a Python API without an LLM. PyOD includes an agentic investigation workflow, an od-expert skill for Claude Code and Codex, and an optional MCP server for compatible agents. Install it with pip, conda-forge, or from source; Python 3.9 or higher is required. Optional pip extras add capabilities such as PyTorch and graph detectors, embeddings, audio, and MCP support. PyOD is free, with optional capabilities requiring extras. Its documentation cautions that ADEngine’s quality verdict is heuristic, not a guarantee, and recommends checking results against held-out labels or through domain review. The model persistence guide also warns that pickle and joblib can run arbitrary Python code when loading artifacts; callers must pass trusted=True.

PyOD brings multiple anomaly detection methods and an orchestration workflow into a Python library. Validate results independently, and load only trusted persisted artifacts.

PyOD suits Python users in academic research or commercial products who need anomaly detection across varied data types. Its classic API and standalone ADEngine option can serve users who do not want to use an LLM.

PyOD is a Python library for building anomaly detection into custom analysis and software workflows. It suits developers and researchers who want multiple detection methods across varied data types. Its breadth and orchestration are useful, but the output still calls for independent validation.

Overview

Initialized in 2017, PyOD brings 61 detectors for tabular, time-series, graph, text, image, and audio data under one API. It supports hybrid and real-time detection, anomaly explanations, and self-hosted deployment. That makes it a flexible toolkit for building detection into a Python pipeline, rather than a ready-made monitoring service.

Users can work with the classic detector API, ADEngine lifecycle orchestration, or an agentic investigation workflow. Python 3.9 or higher is required, and some capabilities depend on optional pip extras. The breadth is most valuable when a project needs to compare methods or handle more than one data type; it also leaves users responsible for choosing suitable detectors and judging their output.

Key features

Detectors and lifecycle orchestration

The shared API spans 61 detectors across six data types. ADEngine profiles data, selects benchmark-backed detectors, runs multiple detectors in parallel, computes consensus scores, and reports diagnostics. This gives Python users a way to coordinate and compare detector runs without managing each step separately. Its quality verdict is a heuristic, not a correctness guarantee; validate results against held-out labels or through domain review.

Agent and Python workflows

ADEngine also works as a standalone Python API without an LLM. For assisted investigation, PyOD provides an od-expert skill for Claude Code and Codex, activation paths for those agents, and an optional MCP server for compatible agents. These options can fit teams already using agent workflows, while the standalone API remains suitable where an LLM is not wanted.

Extensible installation

PyOD can be installed through pip, conda-forge, or from source. Optional extras cover capabilities including PyTorch detectors, graph detectors, embeddings, audio, and MCP; the broader integrations include SUOD, XGBoost, model combination, thresholding, OpenAI embeddings, and Hugging Face encoders. This modularity lets teams add relevant components, but optional capabilities mean the base installation does not automatically provide every integration.

Persistence needs care: pickle and joblib artifacts can execute arbitrary Python code when deserialized. Load only trusted artifacts, and pass trusted=True as required by the persistence guide.

Pricing

PyOD is free. The PyOD plan costs nothing and is an open-source Python library; optional capabilities require pip extras. There is no free trial, since the library is already offered as a free plan. No paid tier or seat-based limit is part of this offering.

Platforms

PyOD is available for Linux, macOS, and Windows, with API access and self-hosted deployment. The Python 3.9-or-higher requirement makes it a fit for teams able to maintain a Python environment, rather than users seeking a standalone desktop app.

Who it's for

PyOD is aimed at researchers and product teams building anomaly detection into their own work. The common detector API and data-type coverage suit projects that need method choice across varied inputs; ADEngine suits users who want coordinated runs and diagnostics. It is less appropriate for buyers who need a turnkey service, guaranteed result quality, or a paid support plan. For proposed new detectors, contributors are expected to commit to at least two years of maintenance, a meaningful obligation for would-be maintainers.

Pros and cons

• Pros: 61 detectors across tabular, time-series, graph, text, image, and audio data share one API, giving developers a broad toolkit for mixed-data work.
• Pros: ADEngine handles profiling, parallel runs, consensus scoring, and diagnostics, reducing orchestration work for users comparing detectors.
• Pros: Free distribution through pip and conda-forge, plus source installation, lets teams integrate the library into their own Python environments.
• Cons: Python 3.9 or higher is required, and optional extras add setup choices for capabilities beyond the core installation.
• Cons: ADEngine's quality verdict is heuristic, so teams need held-out labels or domain review rather than treating its scores as proof.
• Cons: Persisted pickle and joblib artifacts require a trust decision because loading them can deserialize arbitrary code.

Alternatives

For a focused open-source Python option, Anomalib is another free library, installable from PyPI or source. PySAD is also a free open-source Python framework and may suit readers comparing Python-based frameworks. Choose ManageEngine NetFlow Analyzer instead when its interface-based Free Edition or paid Standard and Professional editions better match the deployment and budget. ObservabilityOS offers a free developer plan with one service, 500MB of logs per month, and seven-day retention for readers seeking a bounded observability service rather than a detector library.

Axomaly is another free option. Metaplane offers a free plan for up to 10 monitored tables, four users, and three custom SQL monitors. Soda has a free plan covering pipeline testing, metrics observability, and alerting and ticketing integrations. OpenSearch is a free Apache 2.0 open-source option. Browse more choices in Anomaly Detection Software.

Verdict

Choose PyOD if you need a free, self-hosted Python toolkit with a wide detector catalog and an orchestration layer for comparing results. Its main strength is combining broad method and data-type coverage under one API; its main caveat is that scores remain heuristic and require validation. Look elsewhere for a turnkey monitoring product or a result you can trust without domain review.