cwltool: The reference reference implementation of the Common Workflow Language standards

Linux Status Coverage Status Documentation Status

PyPI: PyPI Version PyPI Downloads Month Total PyPI Downloads

Conda: Conda Version Conda Installs

Debian: Debian Testing package Debian Stable package (Docker): Container

This is the reference implementation of the Common Workflow Language open standards. It is intended to be feature complete and provide comprehensive validation of CWL files as well as provide other tools related to working with CWL.

cwltool is written and tested for Python 3.x {x = 6, 8, 9, 10, 11}

The reference implementation consists of two packages. The cwltool package is the primary Python module containing the reference implementation in the cwltool module and console executable by the same name.

The cwlref-runner package is optional and provides an additional entry point under the alias cwl-runner, which is the implementation-agnostic name for the default CWL interpreter installed on a host.

cwltool is provided by the CWL project, a member project of Software Freedom Conservancy and our many contributors.


cwltool packages

Your operating system may offer cwltool directly. For Debian, Ubuntu, and similar Linux distribution try

sudo apt-get install cwltool

If you encounter an error, first try to update package information by using

sudo apt-get update

If you are running macOS X or other UNIXes and you want to use packages prepared by the conda-forge project, then please follow the install instructions for conda-forge (if you haven’t already) and then

conda install -c conda-forge cwltool

All of the above methods of installing cwltool use packages that might contain bugs already fixed in newer versions or be missing desired features. If the packaged version of cwltool available to you is too old, then we recommend installing using pip and venv

python3 -m venv env      # Create a virtual environment named 'env' in the current directory
source env/bin/activate  # Activate environment before installing `cwltool`

Then install the latest cwlref-runner package from PyPi (which will install the latest cwltool package as well)

pip install cwlref-runner

If installing alongside another CWL implementation (like toil-cwl-runner or arvados-cwl-runner) then instead run

pip install cwltool

MS Windows users

  1. Install Windows Subsystem for Linux 2 and Docker Desktop.

  2. Install Debian from the Microsoft Store.

  3. Set Debian as your default WSL 2 distro: wsl --set-default debian.

  4. Return to the Docker Desktop, choose SettingsResourcesWSL Integration and under “Enable integration with additional distros” select “Debian”,

  5. Reboot if you have not yet already.

  6. Launch Debian and follow the Linux instructions above (apt-get install cwltool or use the venv method)

Network problems from within WSL2? Try these instructions followed by wsl --shutdown.

cwltool development version

Or you can skip the direct pip commands above and install the latest development version of cwltool:

git clone # clone (copy) the cwltool git repository
cd cwltool           # Change to source directory that git clone just downloaded
pip install .[deps]  # Installs ``cwltool`` from source
cwltool --version    # Check if the installation works correctly

Remember, if co-installing multiple CWL implementations, then you need to maintain which implementation cwl-runner points to via a symbolic file system link or another facility.

Run on the command line

Simple command:

cwl-runner my_workflow.cwl my_inputs.yaml

Or if you have multiple CWL implementations installed and you want to override the default cwl-runner then use:

cwltool my_workflow.cwl my_inputs.yml

You can set cwltool options in the environment with CWLTOOL_OPTIONS, these will be inserted at the beginning of the command line:

export CWLTOOL_OPTIONS="--debug"

Use with boot2docker on macOS

boot2docker runs Docker inside a virtual machine, and it only mounts Users on it. The default behavior of CWL is to create temporary directories under e.g. /Var which is not accessible to Docker containers.

To run CWL successfully with boot2docker you need to set the --tmpdir-prefix and --tmp-outdir-prefix to somewhere under /Users:

$ cwl-runner --tmp-outdir-prefix=/Users/username/project --tmpdir-prefix=/Users/username/project wc-tool.cwl wc-job.json

Using uDocker

Some shared computing environments don’t support Docker software containers for technical or policy reasons. As a workaround, the CWL reference runner supports using the udocker program on Linux using --udocker.

udocker installation:

Run cwltool just as you usually would, but with --udocker prior to the workflow path:

cwltool --udocker

As was mentioned in the Recommended Software section,

Using Singularity

cwltool can also use Singularity version 2.6.1 or later as a Docker container runtime. cwltool with Singularity will run software containers specified in DockerRequirement and therefore works with Docker images only, native Singularity images are not supported. To use Singularity as the Docker container runtime, provide --singularity command line option to cwltool. With Singularity, cwltool can pass all CWL v1.0 conformance tests, except those involving Docker container ENTRYPOINTs.


cwltool --singularity

Running a tool or workflow from remote or local locations

cwltool can run tool and workflow descriptions on both local and remote systems via its support for HTTP[S] URLs.

Input job files and Workflow steps (via the run directive) can reference CWL documents using absolute or relative local filesystem paths. If a relative path is referenced and that document isn’t found in the current directory, then the following locations will be searched:

You can also use cwldep to manage dependencies on external tools and workflows.

Overriding workflow requirements at load time

Sometimes a workflow needs additional requirements to run in a particular environment or with a particular dataset. To avoid the need to modify the underlying workflow, cwltool supports requirement “overrides”.

The format of the “overrides” object is a mapping of item identifier (workflow, workflow step, or command line tool) to the process requirements that should be applied.

          MESSAGE: override_value

Overrides can be specified either on the command line, or as part of the job input document. Workflow steps are identified using the name of the workflow file followed by the step name as a document fragment identifier “#id”. Override identifiers are relative to the top-level workflow document.

cwltool --overrides overrides.yml my-tool.cwl my-job.yml
input_parameter1: value1
input_parameter2: value2
          MESSAGE: override_value
cwltool my-tool.cwl my-job-with-overrides.yml

Combining parts of a workflow into a single document

Use --pack to combine a workflow made up of multiple files into a single compound document. This operation takes all the CWL files referenced by a workflow and builds a new CWL document with all Process objects (CommandLineTool and Workflow) in a list in the $graph field. Cross references (such as run: and source: fields) are updated to internal references within the new packed document. The top-level workflow is named #main.

cwltool --pack my-wf.cwl > my-packed-wf.cwl

Running only part of a workflow

You can run a partial workflow with the --target (-t) option. This takes the name of an output parameter, workflow step, or input parameter in the top-level workflow. You may provide multiple targets.

cwltool --target step3 my-wf.cwl

If a target is an output parameter, it will only run only the steps that contribute to that output. If a target is a workflow step, it will run the workflow starting from that step. If a target is an input parameter, it will only run the steps connected to that input.

Use --print-targets to get a listing of the targets of a workflow. To see which steps will run, use --print-subgraph with --target to get a printout of the workflow subgraph for the selected targets.

cwltool --print-targets my-wf.cwl

cwltool --target step3 --print-subgraph my-wf.cwl > my-wf-starting-from-step3.cwl

Visualizing a CWL document

The --print-dot option will print a file suitable for Graphviz dot program. Here is a bash onliner to generate a Scalable Vector Graphic (SVG) file:

cwltool --print-dot my-wf.cwl | dot -Tsvg > my-wf.svg

Modeling a CWL document as RDF

CWL documents can be expressed as RDF triple graphs.

cwltool --print-rdf --rdf-serializer=turtle mywf.cwl

Environment Variables in cwltool

This reference implementation supports several ways of setting environment variables for tools, in addition to the standard EnvVarRequirement. The sequence of steps applied to create the environment is:

  1. If the --preserve-entire-environment flag is present, then begin with the current environment, else begin with an empty environment.

  2. Add any variables specified by --preserve-environment option(s).

  3. Set TMPDIR and HOME per the CWL v1.0+ CommandLineTool specification.

  4. Apply any EnvVarRequirement from the CommandLineTool description.

  5. Apply any manipulations required by any cwltool:MPIRequirement extensions.

  6. Substitute any secrets required by Secrets extension.

  7. Modify the environment in response to SoftwareRequirement (see below).

Leveraging SoftwareRequirements (Beta)

CWL tools may be decorated with SoftwareRequirement hints that cwltool may in turn use to resolve to packages in various package managers or dependency management systems such as Environment Modules.

Utilizing SoftwareRequirement hints using cwltool requires an optional dependency, for this reason be sure to use specify the deps modifier when installing cwltool. For instance:

$ pip install 'cwltool[deps]'

Installing cwltool in this fashion enables several new command line options. The most general of these options is --beta-dependency-resolvers-configuration. This option allows one to specify a dependency resolver’s configuration file. This file may be specified as either XML or YAML and very simply describes various plugins to enable to “resolve” SoftwareRequirement dependencies.

Using these hints will allow cwltool to modify the environment in which your tool runs, for example by loading one or more Environment Modules. The environment is constructed as above, then the environment may modified by the selected tool resolver. This currently means that you cannot override any environment variables set by the selected tool resolver. Note that the environment given to the configured dependency resolver has the variable _CWLTOOL set to 1 to allow introspection.

To discuss some of these plugins and how to configure them, first consider the following hint definition for an example CWL tool.

  - package: seqtk
    - r93

Now imagine deploying cwltool on a cluster with Software Modules installed and that a seqtk module is available at version r93. This means cluster users likely won’t have the binary seqtk on their PATH by default, but after sourcing this module with the command modulecmd sh load seqtk/r93 seqtk is available on the PATH. A simple dependency resolvers configuration file, called dependency-resolvers-conf.yml for instance, that would enable cwltool to source the correct module environment before executing the above tool would simply be:

- type: modules

The outer list indicates that one plugin is being enabled, the plugin parameters are defined as a dictionary for this one list item. There is only one required parameter for the plugin above, this is type and defines the plugin type. This parameter is required for all plugins. The available plugins and the parameters available for each are documented (incompletely) here. Unfortunately, this documentation is in the context of Galaxy tool requirement s instead of CWL SoftwareRequirement s, but the concepts map fairly directly.

cwltool is distributed with an example of such seqtk tool and sample corresponding job. It could executed from the cwltool root using a dependency resolvers configuration file such as the above one using the command:

cwltool --beta-dependency-resolvers-configuration /path/to/dependency-resolvers-conf.yml \
    tests/seqtk_seq.cwl \

This example demonstrates both that cwltool can leverage existing software installations and also handle workflows with dependencies on different versions of the same software and libraries. However the above example does require an existing module setup so it is impossible to test this example “out of the box” with cwltool. For a more isolated test that demonstrates all the same concepts - the resolver plugin type galaxy_packages can be used.

“Galaxy packages” are a lighter-weight alternative to Environment Modules that are really just defined by a way to lay out directories into packages and versions to find little scripts that are sourced to modify the environment. They have been used for years in Galaxy community to adapt Galaxy tools to cluster environments but require neither knowledge of Galaxy nor any special tools to setup. These should work just fine for CWL tools.

The cwltool source code repository’s test directory is setup with a very simple directory that defines a set of “Galaxy packages” (but really just defines one package named random-lines). The directory layout is simply:


If the galaxy_packages plugin is enabled and pointed at the tests/test_deps_env directory in cwltool’s root and a SoftwareRequirement such as the following is encountered.

    - package: 'random-lines'
      - '1.0'

Then cwltool will simply find that file and source it before executing the corresponding tool. That script is only responsible for modifying the job’s PATH to add the required binaries.

This is a full example that works since resolving “Galaxy packages” has no external requirements. Try it out by executing the following command from cwltool’s root directory:

cwltool --beta-dependency-resolvers-configuration tests/test_deps_env_resolvers_conf.yml \
    tests/random_lines.cwl \

The resolvers configuration file in the above example was simply:

- type: galaxy_packages
  base_path: ./tests/test_deps_env

It is possible that the SoftwareRequirement s in a given CWL tool will not match the module names for a given cluster. Such requirements can be re-mapped to specific deployed packages or versions using another file specified using the resolver plugin parameter mapping_files. We will demonstrate this using galaxy_packages, but the concepts apply equally well to Environment Modules or Conda packages (described below), for instance.

So consider the resolvers configuration file. (tests/test_deps_env_resolvers_conf_rewrite.yml):

- type: galaxy_packages
  base_path: ./tests/test_deps_env
  mapping_files: ./tests/test_deps_mapping.yml

And the corresponding mapping configuration file (tests/test_deps_mapping.yml):

- from:
    name: randomLines
    version: 1.0.0-rc1
    name: random-lines
    version: '1.0'

This is saying if cwltool encounters a requirement of randomLines at version 1.0.0-rc1 in a tool, to rewrite to our specific plugin as random-lines at version 1.0. cwltool has such a test tool called random_lines_mapping.cwl that contains such a source SoftwareRequirement. To try out this example with mapping, execute the following command from the cwltool root directory:

cwltool --beta-dependency-resolvers-configuration tests/test_deps_env_resolvers_conf_rewrite.yml \
    tests/random_lines_mapping.cwl \

The previous examples demonstrated leveraging existing infrastructure to provide requirements for CWL tools. If instead a real package manager is used cwltool has the opportunity to install requirements as needed. While initial support for Homebrew/Linuxbrew plugins is available, the most developed such plugin is for the Conda package manager. Conda has the nice properties of allowing multiple versions of a package to be installed simultaneously, not requiring evaluated permissions to install Conda itself or packages using Conda, and being cross-platform. For these reasons, cwltool may run as a normal user, install its own Conda environment and manage multiple versions of Conda packages on Linux and Mac OS X.

The Conda plugin can be endlessly configured, but a sensible set of defaults that has proven a powerful stack for dependency management within the Galaxy tool development ecosystem can be enabled by simply passing cwltool the --beta-conda-dependencies flag.

With this, we can use the seqtk example above without Docker or any externally managed services - cwltool should install everything it needs and create an environment for the tool. Try it out with the following command:

cwltool --beta-conda-dependencies tests/seqtk_seq.cwl tests/seqtk_seq_job.json

The CWL specification allows URIs to be attached to SoftwareRequirement s that allow disambiguation of package names. If the mapping files described above allow deployers to adapt tools to their infrastructure, this mechanism allows tools to adapt their requirements to multiple package managers. To demonstrate this within the context of the seqtk, we can simply break the package name we use and then specify a specific Conda package as follows:

    - package: seqtk_seq
      - '1.2'

The example can be executed using the command:

cwltool --beta-conda-dependencies tests/seqtk_seq_wrong_name.cwl tests/seqtk_seq_job.json

The plugin framework for managing the resolution of these software requirements as maintained as part of galaxy-tool-util - a small, portable subset of the Galaxy project. More information on configuration and implementation can be found at the following links:

Use with GA4GH Tool Registry API

Cwltool can launch tools directly from GA4GH Tool Registry API endpoints.

By default, cwltool searches . Use --add-tool-registry to add other registries to the search path.

For example

cwltool test.json

and (defaults to latest when a version is not specified)

cwltool test.json

For this example, grab the test.json (and input file) from


Running MPI-based tools that need to be launched

Cwltool supports an extension to the CWL spec When the tool definition has this in its requirements/hints section, and cwltool has been run with --enable-ext, then the tool’s command line will be extended with the commands needed to launch it with mpirun or similar. You can specify the number of processes to start as either a literal integer or an expression (that will result in an integer). For example:

#!/usr/bin/env cwl-runner
cwlVersion: v1.1
class: CommandLineTool
  cwltool: ""
    processes: $(inputs.nproc)
    type: int

Interaction with containers: the MPIRequirement currently prepends its commands to the front of the command line that is constructed. If you wish to run a containerized application in parallel, for simple use cases, this does work with Singularity, depending upon the platform setup. However, this combination should be considered “alpha” – please do report any issues you have! This does not work with Docker at the moment. (More precisely, you get n copies of the same single process image run at the same time that cannot communicate with each other.)

The host-specific parameters are configured in a simple YAML file (specified with the --mpi-config-file flag). The allowed keys are given in the following table; all are optional.








The primary command to use.




Flag to set number of processes to start.




Default number of processes.




A list of any other flags to be added to the runner’s command line before the baseCommand.




A list of environment variables that should be passed from the host environment through to the tool (e.g., giving the node list as set by your scheduler).




A list of python regular expressions that will be matched against the host’s environment. Those that match will be passed through.




A dictionary whose keys are the environment variables set and the values being the values.

Enabling Fast Parser (experimental)

For very large workflows, cwltool can spend a lot of time in initialization, before the first step runs. There is an experimental flag --fast-parser which can dramatically reduce the initialization overhead, however as of this writing it has several limitations:

  • Error reporting in general is worse than the standard parser, you will want to use it with workflows that you know are already correct.

  • It does not check for dangling links (these will become runtime errors instead of loading errors)

  • Several other cases fail, as documented in


Running tests locally

  • Running basic tests (/tests):

To run the basic tests after installing cwltool execute the following:

pip install -rtest-requirements.txt
pytest   ## N.B. This requires node.js or docker to be available

To run various tests in all supported Python environments, we use tox. To run the test suite in all supported Python environments first clone the complete code repository (see the git clone instructions above) and then run the following in the terminal: pip install "tox<4"; tox -p

List of all environment can be seen using: tox --listenvs and running a specific test env using: tox -e <env name> and additionally run a specific test using this format: tox -e py310-unit -- -v tests/

  • Running the entire suite of CWL conformance tests:

The GitHub repository for the CWL specifications contains a script that tests a CWL implementation against a wide array of valid CWL files using the cwltest program

Instructions for running these tests can be found in the Common Workflow Language Specification repository at .

Import as a module


import cwltool

to your script.

The easiest way to use cwltool to run a tool or workflow from Python is to use a Factory

import cwltool.factory
fac = cwltool.factory.Factory()

echo = fac.make("echo.cwl")
result = echo(inp="foo")

# result["out"] == "foo"

CWL Tool Control Flow

Technical outline of how cwltool works internally, for maintainers.

  1. Use CWL load_tool() to load document.

    1. Fetches the document from file or URL

    2. Applies preprocessing (syntax/identifier expansion and normalization)

    3. Validates the document based on cwlVersion

    4. If necessary, updates the document to the latest spec

    5. Constructs a Process object using make_tool()` callback. This yields a CommandLineTool, Workflow, or ExpressionTool. For workflows, this recursively constructs each workflow step.

    6. To construct custom types for CommandLineTool, Workflow, or ExpressionTool, provide a custom make_tool()

  2. Iterate on the job() method of the Process object to get back runnable jobs.

    1. job() is a generator method (uses the Python iterator protocol)

    2. Each time the job() method is invoked in an iteration, it returns one of: a runnable item (an object with a run() method), None (indicating there is currently no work ready to run) or end of iteration (indicating the process is complete.)

    3. Invoke the runnable item by calling run(). This runs the tool and gets output.

    4. An output callback reports the output of a process.

    5. job() may be iterated over multiple times. It will yield all the work that is currently ready to run and then yield None.

  3. Workflow objects create a corresponding WorkflowJob and WorkflowJobStep objects to hold the workflow state for the duration of the job invocation.

    1. The WorkflowJob iterates over each WorkflowJobStep and determines if the inputs the step are ready.

    2. When a step is ready, it constructs an input object for that step and iterates on the job() method of the workflow job step.

    3. Each runnable item is yielded back up to top-level run loop

    4. When a step job completes and receives an output callback, the job outputs are assigned to the output of the workflow step.

    5. When all steps are complete, the intermediate files are moved to a final workflow output, intermediate directories are deleted, and the workflow’s output callback is called.

  4. CommandLineTool job() objects yield a single runnable object.

    1. The CommandLineTool job() method calls make_job_runner() to create a CommandLineJob object

    2. The job method configures the CommandLineJob object by setting public attributes

    3. The job method iterates over file and directories inputs to the CommandLineTool and creates a “path map”.

    4. Files are mapped from their “resolved” location to a “target” path where they will appear at tool invocation (for example, a location inside a Docker container.) The target paths are used on the command line.

    5. Files are staged to targets paths using either Docker volume binds (when using containers) or symlinks (if not). This staging step enables files to be logically rearranged or renamed independent of their source layout.

    6. The run() method of CommandLineJob executes the command line tool or Docker container, waits for it to complete, collects output, and makes the output callback.

Extension points

The following functions can be passed to main() to override or augment the listed behaviors.

executor(tool, job_order_object, runtimeContext, logger)
  (Process, Dict[Text, Any], RuntimeContext) -> Tuple[Dict[Text, Any], Text]

An implementation of the top-level workflow execution loop should synchronously run a process object to completion and return the output object.

  () -> Text

Return version string.


Handler object for logging.

The following functions can be set in LoadingContext to override or augment the listed behaviors.

fetcher_constructor(cache, session)
  (Dict[unicode, unicode], requests.sessions.Session) -> Fetcher

Construct a Fetcher object with the supplied cache and HTTP session.

resolver(document_loader, document)
  (Loader, Union[Text, dict[Text, Any]]) -> Text

Resolve a relative document identifier to an absolute one that can be fetched.

The following functions can be set in RuntimeContext to override or augment the listed behaviors.

construct_tool_object(toolpath_object, loadingContext)
  (MutableMapping[Text, Any], LoadingContext) -> Process

Hook to construct a Process object (eg CommandLineTool) object from a document.

  (Dict[str, int], RuntimeContext) -> Dict[Text, int]

Take a resource request and turn it into a concrete resource assignment.

  (Text) -> StdFsAccess

Return a file system access object.

In addition, when providing custom subclasses of Process objects, you can override the following methods:

  (RuntimeContext) -> Type[JobBase]

Create and return a job runner object (this implements concrete execution of a command line tool).

make_workflow_step(toolpath_object, pos, loadingContext, parentworkflowProv)
  (Dict[Text, Any], int, LoadingContext, Optional[ProvenanceProfile]) -> WorkflowStep

Create and return a workflow step object.

Indices and tables