Atomica is a simulation engine for compartmental models. It can be used to simulate disease epidemics, health care cascades, and many other things.
For detailed documentation, visit https://atomica.tools/docs
Atomica requires Python 3.11 or later and is distributed via PyPI. To install, run
pip install atomica
Atomica is mainly used as the modelling platform for downstream analyses. Therefore, the most common usage is to set up an analysis-specific repository, and simply include atomica as a dependency that will be installed automatically via PyPI. We recommend using uv to manage Python environments, in which case atomica can be added as a dependency using
uv add atomica
Atomica ships an MCP server (atomica.mcp) that exposes tools for querying framework and databook files and a set of built-in workflow skills (MCP prompts). The tools let Claude read compartments, parameters, transitions, and variable metadata directly from .xlsx framework files, as well as read and write data in .xlsx databook files. The skills guide Claude through multi-step workflows such as producing a structured summary of a framework.
To register the server with Claude Code, run the following from within your project directory:
claude mcp add atomica -- uv run python -m atomica.mcp
Once added, the tools and skills are available automatically in any Claude Code session for that project. If you use Atomica across multiple projects and want the MCP server available in all of them without repeating the claude mcp add step, register it at the user level instead:
claude mcp add -s user atomica -- uv run python -m atomica.mcp
The Atomica MCP will then be used whenever you are working within a project that has atomica as a dependency.
To use an Atomica branch in a downstream project with uv, you can add the Git repository directly as a dependency
uv add git+https://github.com/atomicateam/atomica --branch <branch name>
For more information on this usage, see https://docs.astral.sh/uv/concepts/projects/dependencies/.
If you want to install a different branch of Atomica, or plan to make changes to the Atomica source code, you will need to install Atomica via Git rather than via PyPI.
git clone https://github.com/atomicateam/atomica.git
cd atomica
pip install -e .
If using uv, simply cloning the repository is sufficient, and scripts can be run with uv run. If you are developing Atomica in parallel with your own analysis repository, it would be recommended to clone atomica and then install it in your analysis repository as an editable package with uv
uv add --editable ../<path to atomica>
In which case you can edit your local copy of Atomica and have it reflected in your analysis code.
Atomica includes a suite of tests. The automated test suite can be executed with pytest.
uv run --extra test pytest
Note the inclusion of the extra test dependencies that are not installed by default. Many of the tests open matplotlib figures as part of the test. If the test script is run on a machine without a display available, the error
_tkinter.TclError: couldn't connect to display "localhost:0.0"
will be raised. In that case, simply set the matplotlib backend to agg which allows the calls to succeed with a display present. For example, run
export MPLBACKEND=agg
uv run --extra test pytest
To also validate the example and tutorial notebooks, include the nbval options:
uv run --extra test pytest --nbval-lax --current-env --nbval-cell-timeout=600 --dist loadscope
This will reproduce the automated testing that is executed on GitHub as part of the CI workflow.
The MCP server publishes a set of skills, which are plain Markdown files in atomica/mcp/skills/. To add a new workflow, add a .md file into that directory — it is registered as an MCP prompt automatically when the server starts, with no code changes required. The first # Heading line becomes the prompt description shown in the MCP client.
If running pip install -e . in a new environment, numpy must be installed prior to scipy. In some cases,
installing numpy may fail due to missing compiler options. In that case, you may wish to install numpy via Anaconda
(by installing Python through Anaconda, and using conda install numpy scipy matplotlib). In general, our experience
has been that it is easier to set up the C binaries for numpy and the QT dependencies for matplotlib via Anaconda
rather than doing this via the system, which involves different steps on every platform.
On some systems, the default matplotlib backend may hang - this is not an issue with atomica. To resolve, try changing the backend by including
import matplotlib
matplotlib.use("Qt5Agg")
at the very start of your script. A different backend may be required for your system. You can make the change persistant by setting the backend in your matplotlibrc file.