Onboard your agent to Andrey

Accelerated, accurate, agentic causal discovery.

uv add andrey-core

pip install andrey-core

Get started

• AcceleratedMore than 5,000x faster
• AccurateMore than 80% fewer errors
• AgenticBuilt for agents
• All-in-one44 methods, five kinds of graph

Accelerated

More than 5,000x faster than other popular packages.

Constraint-basedScore-basedPermutation-basedFunctional causal modelsTemporal causal discoveryLatent variablesPairwise direction

more than5,000xfaster

AllPCFCIGFCICD-NODInter-IAMBIAMBMMPCICPFCI-JCI

Andreycausal-learngCastleTetradbnlearnInvariantCausalPredictionpcalgMost attractive quadrant

Each line runs from a method's smallest graph to its largest. Down and left is better. Details in the release cards

more than900xfaster

AllXGESTabu searchHill climbingGESGIESExact searchCALMMMHCRSMAX2

Andreyxgesbnlearncausal-learnpcalgMost attractive quadrant

Each line runs from a method's smallest graph to its largest. Down and left is better. Details in the release cards

more than200xfaster

AllBOSSGRaSPBOSS-FCIGRaSP-FCI

Andreycausal-learnTetradMost attractive quadrant

Each line runs from a method's smallest graph to its largest. Down and left is better. Details in the release cards

more than10xfaster

AllDirectLiNGAMMultiGroupDirectLiNGAMCAMRESIT

Andreycausal-learnlingamCAMauthors' R RESITMost attractive quadrant

Each line runs from a method's smallest graph to its largest. Down and left is better. Details in the release cards

more than100xfaster

AllPCMCIPCMCI+LPCMCIVARMA-LiNGAMLongitudinalLiNGAMGrangerJ-PCMCI+

Andreytigramitelingamcausal-learnMost attractive quadrant

Each line runs from a method's smallest graph to its largest. Down and left is better. Details in the release cards

more than100xfaster

AllRCDBottomUpParceLiNGAMGINCAM-UV

AndreylingamTetradcausal-learnMost attractive quadrant

Each line runs from a method's smallest graph to its largest. Down and left is better. Details in the release cards

more than1,000xfaster

AllANMPNL

Andreycausal-learnMost attractive quadrant

Each line runs from a method's smallest graph to its largest. Down and left is better. Details in the release cards

Explore why Andrey is so fast

Accurate

More than 80% fewer errors than other popular packages.

• Fewer errors than the other package

  Each family's largest cut, summed over every size both ran.

  1. PNL89% fewer errors

     Andrey6
     causal-learn55

  2. RESIT70% fewer errors

     Andrey19
     lingam RESIT63

  3. PC on mixed data, degenerate Gaussian test69% fewer errors

     Andrey30
     Tetrad96

  4. BOSS50% fewer errors

     Andrey43
     causal-learn86

  5. CALM, paper schedule32% fewer errors

     Andrey27
     causal-learn40

• Tested against the true graphs

  Every release is tested against the true graphs that generated the data. Its release card gives every method's errors beside the other package's.

  Release cards

Agentic

A command line that never prompts and answers in JSON, so an agent can run it alone.

• Find a method, read it, run it

  An agent lists the methods, reads one's parameters as JSON Schema, and gets the graph back as JSON, its nodes named by the data's header row.

  $ andrey list | jq -r '.[] | [.name, .output, .status] | @tsv'
  pc	cpdag	supported
  fci	pag	supported
  gfci	pag	supported
  ... 43 more
  $ andrey run pc --help --json | jq -c '.method.params.properties.alpha'
  {"type":"number","default":0.05,"exclusiveMinimum":0.0,"exclusiveMaximum":1.0,"description":"Significance level of the conditional-independence test.","concept":"alpha","aliases":[],"choice_aliases":{}}
  $ andrey run pc --data sachs.csv | jq -c '.structure | .labels as $n | .edges[] | [$n[.source], $n[.target], .type]'
  ["raf","mek","undirected"]
  ["plc","pip3","undirected"]
  ["pip2","pip3","undirected"]
  ["erk","akt","undirected"]
  ... 4 more

All-in-one

44 supported methods for every kind of data, and five kinds of causal graph.

• Compared with other packages

  Scroll sideways for every package.

  	Andrey	causal-learn	Tetrad	pcalg	bnlearn
  Coverage
  Distinct methods	42	21	21	8	10
  Temporal data	Yes	Yes	Yes	No	No
  Mixed data	Yes	Partly: No mixed-data test or score; only KCI and the generalized score take category codes, and neither is built for them.	Yes	No	Yes
  Nonlinear data	Yes	Yes	Yes	No	No
  Interventional data	Yes	Partly: CD-NOD only, for data from several domains, with no targets.	Partly: Intervention variables tagged in a metadata file join the first knowledge tier as ordinary variables; no method takes the targets.	Yes	Partly: The mbde score only, for categorical data in its score-based searches.
  Permutation invariance	Yes	No	Partly: PC and FCI follow the variables' names; BOSS, GRaSP, and GIN changed with the column order.	No	No
  Real data
  Missing values	Yes	Partly: PC and CD-NOD only, with the Fisher-z test.	Yes	No	Yes
  Background knowledge	Yes	Partly: PC, FCI, CD-NOD, the LiNGAM methods, and exact search; GES, BOSS, and GRaSP take none.	Yes	Partly: Fixed adjacencies only, with no directions or tiers.	Yes
  Edge confidence	Yes	Partly: Bootstrap edge probabilities for the LiNGAM methods only.	Yes	No	Yes
  Speed
  CPU parallelization	Yes	Partly: Several CPU cores for RLCD and the FastKCI test only.	Partly: Several threads for FGES, BOSS, GFCI, and PC's adjacency search.	Partly: Several cores for the skeleton step of its stable.fast method only.	Partly: A cluster for its constraint-based methods and DirectLiNGAM only.
  GPU acceleration	Yes	Partly: CALM only.	No	No	No
  Open benchmarks	Yes	No	No	No	No
  Agentic
  JSON command line	Yes	No	Partly: causal-cmd writes the graph, and only the graph, to a JSON file.	No	No
  Schema for every method	Yes	No	Partly: Typed parameter defaults and bounds, in an HTML manual.	No	No
  Agent skill	Yes	No	Partly: py-tetrad's development branch has a skill and guides; no release has them.	No	No
  Read from each package's released code: causal-learn 0.1.4.8, Tetrad 7.6.10, pcalg 2.7-12, and bnlearn 5.2.1.

• Five kinds of causal graph

  DAGs, CPDAGs, PAGs, cyclic graphs, and graphs over time.

  DAG

  CPDAG

  PAG

  Graph over time

  Cyclic graph

From data to graph

From a table of measurements to its causal graph, in a few lines of Python.

Sachs and colleagues measured 11 proteins and phospholipids in 853 human immune cells. PC finds 8 of the 17 links in the known signaling network, and none outside it.

The recovered graph, drawn with andrey.viz.draw.

More examples

example.py

import andrey
from andrey.data import load_dataset

df = load_dataset("sachs").data
out = andrey.pc(df)
print(out)

example.py

pc  cpdag  |  11 nodes  |  8 edges (6 undirected, 2 directed)

  raf -- mek
  plc -- pip3
  pip2 -- pip3
  erk -- akt
  erk -- pka
  akt -- pka
  p38 -> pkc
  jnk -> pkc

  indep_test: fisherz
  tie_break: rank

Five kinds of causal graph

A causal graph says which variables drive which, and how much the data can tell decides its kind. Four methods on data about a lawn, its sprinkler, and the rain: three on a table of measurements, and VARMA-LiNGAM on the same lawn day by day.

DAG, from DirectLiNGAM. One direction for every edge: non-Gaussian noise settles each one.

CPDAG, from PC. The data cannot say which way the two edges from season point, so they stay plain lines.

PAG, from FCI. Circles on both ends: either direction, or a hidden common cause, fits the data.

Graph over time, from VARMA-LiNGAM. Rain wets the lawn the same day; the sprinkler wets it the next day, and a dry lawn turns the sprinkler on the next day.

Cyclic graph, the same graph with time collapsed. The sprinkler and the lawn drive each other, though every arrow over time points forward.

What each kind of graph means

Agent support

The andrey command never prompts and answers in JSON, and it describes every method's parameters, so an agent can use it with no person watching. andrey --skill prints a guide the agent loads as a skill.

1. Find a method

   andrey list names all 46 methods, with the graph each returns.

2. Read its contract

   andrey run pc --help --json gives each parameter's type, default, and meaning as JSON Schema.

3. Run it

   The run prints the graph as JSON, its nodes named by the data's header row.

$ andrey list | jq -r '.[] | [.name, .output, .status] | @tsv'

pc cpdag supportedfci pag supportedgfci pag supported... 43 more

exit 0

$ andrey run pc --help --json | jq -c '.method.params.properties.alpha'

{"type":"number","default":0.05,"exclusiveMinimum":0.0,"exclusiveMaximum":1.0,"description":"Significance level of the conditional-independence test.","concept":"alpha","aliases":[],"choice_aliases":{}}

exit 0

$ andrey run pc --data sachs.csv | jq -c '.structure | .labels as $n | .edges[] | [$n[.source], $n[.target], .type]'

["raf","mek","undirected"]["plc","pip3","undirected"]["pip2","pip3","undirected"]["erk","akt","undirected"]["erk","pka","undirected"]["akt","pka","undirected"]["p38","pkc","directed"]["jnk","pkc","directed"]

exit 0

The graph in that JSON, drawn with andrey.viz.draw.

Command-line reference

Methods for every kind of data

Andrey is a Python package for learning causal graphs from observational data, with constraint-based, score-based, permutation-based, functional causal model, temporal, latent-variable, and pairwise methods. 46 methods in 7 families. Every method name links to its API reference. An experimental method has no published benchmark, and its API may change without deprecation.

Constraint-based

• andrey.pcBenchmarkedCPDAG

  General-purpose structure learning from i.i.d. observational data

• andrey.fciBenchmarkedPAG

  Structure learning that allows latent confounders and selection bias

• andrey.gfciBenchmarkedPAG

  Score-then-FCI hybrid

• andrey.cdnodBenchmarkedCPDAG

  Constraint-based discovery for data pooled from several domains or regimes

• andrey.fci_jciBenchmarkedPAG

  FCI for data pooled from several domains or regimes, allowing hidden common causes

• andrey.icpBenchmarkedDAG

  Invariant causal prediction

• andrey.iambBenchmarkedCPDAG

  One variable's Markov blanket (its parents, children, and children's other parents) by incremental association

• andrey.inter_iambBenchmarkedCPDAG

  One variable's Markov blanket by interleaved IAMB, which shrinks after every addition

• andrey.mmpcBenchmarkedUndirected graph

  One variable's parents and children, undirected, by max-min parents and children

Score-based

• andrey.gesBenchmarkedCPDAG

  Greedy equivalence search

• andrey.giesBenchmarkedCPDAG

  Greedy interventional equivalence search

• andrey.gniesBenchmarkedCPDAG

  Greedy noise-interventional equivalence search

• andrey.xgesBenchmarkedCPDAG

  Extremely greedy equivalence search

• andrey.hcBenchmarkedCPDAG

  Hill climbing over DAGs

• andrey.tabuBenchmarkedCPDAG

  Tabu search over DAGs

• andrey.exact_searchBenchmarkedCPDAG

  Globally optimal score search over DAGs (A* or dynamic programming)

• andrey.calmBenchmarkedDAG

  Continuous-optimization structure learning in the line of NOTEARS

• andrey.notearsDAG

  Experimental NOTEARS

• andrey.dagmaDAG

  Experimental DAGMA

• andrey.mmhcBenchmarkedCPDAG

  Max-min hill climbing

• andrey.rsmax2BenchmarkedCPDAG

  Restrict-maximize search

Permutation-based

• andrey.bossBenchmarkedCPDAG

  Best-order score search

• andrey.graspBenchmarkedCPDAG

  Greedy relaxation of the sparsest permutation

• andrey.boss_fciBenchmarkedPAG

  GFCI with BOSS in place of GES

• andrey.grasp_fciBenchmarkedPAG

  GFCI with GRaSP in place of GES

• andrey.igspBenchmarkedCPDAG

  Interventional greedy sparsest permutation

• andrey.ut_igspBenchmarkedCPDAG

  Unknown-target interventional greedy sparsest permutation

Functional causal models

• andrey.direct_lingamBenchmarkedDAG

  DirectLiNGAM for linear non-Gaussian data

• andrey.ica_lingamBenchmarkedDAG

  ICA-based LiNGAM for linear non-Gaussian data

• andrey.multi_group_direct_lingamBenchmarkedDAG

  DirectLiNGAM fit jointly across >= 2 datasets sharing the same variables

• andrey.camBenchmarkedDAG

  Causal additive model (CAM) for nonlinear additive effects with additive noise

• andrey.resitBenchmarkedDAG

  Regression with subsequent independence test (RESIT) for nonlinear effects with additive noise of any distribution

Temporal causal discovery

• andrey.varma_lingamBenchmarkedTemporal graph

  VARMA-LiNGAM for a single multivariate time series

• andrey.longitudinal_lingamBenchmarkedTemporal graph

  Longitudinal LiNGAM for panel data (the same units measured at several time points)

• andrey.pcmciBenchmarkedTemporal graph

  PCMCI for a single multivariate time series

• andrey.pcmciplusBenchmarkedTemporal graph

  PCMCI+ for a single multivariate time series

• andrey.jpcmciplusBenchmarkedTemporal graph

  J-PCMCI+ for time series from several datasets that differ through measured or hidden contexts

• andrey.lpcmciBenchmarkedTemporal graph

  LPCMCI for a single multivariate time series with possible hidden common causes

• andrey.grangerBenchmarkedTemporal graph

  Granger causality for a single multivariate time series

• andrey.dynotearsBenchmarkedTemporal graph

  DYNOTEARS for a single multivariate time series

Latent variables

• andrey.ginBenchmarkedDAG

  Generalized independent noise (GIN) for linear non-Gaussian models with latent variables

• andrey.rcdBenchmarkedPAG

  Repetitive causal discovery (RCD) for linear non-Gaussian data with hidden common causes

• andrey.bottom_up_parce_lingamBenchmarkedPAG

  BottomUpParceLiNGAM for linear non-Gaussian data with hidden common causes

• andrey.camuvBenchmarkedPAG

  Causal additive models with unobserved variables (CAM-UV) for nonlinear additive effects where some causes are hidden

Pairwise direction

• andrey.anmBenchmarkedTwo-node DAG

  Additive-noise-model pairwise direction test

• andrey.pnlBenchmarkedTwo-node DAG

  Post-nonlinear pairwise direction test

Documentation

Latest post · November 8, 2026Multi-domain dataWhat multi-domain data are, why data from several environments help causal discovery, and the methods Andrey has for them: CD-NOD, FCI-JCI, and J-PCMCI+.

• Getting startedInstall, then recover your first graph.
• Graph typesWhat a DAG, CPDAG, PAG, or temporal graph claims.
• API referenceEvery method, structure, and command.
• Release cardsTimings, accuracy, limits, and how to rerun them.