Command-line interface

The skore command synchronizes report projects, installs open-source agent skills, and connects a coding agent to Skore Hub. Every option is listed in the command reference.

Hub commands need the hub extra (pip install skore[hub]). Synchronizing with MLflow needs the mlflow extra (pip install skore[mlflow]).

Synchronizing projects

skore sync copies reports from one project to another. Start from a local project named my-xp and send it to a Hub workspace:

$ skore sync my-xp --to hub --to-workspace my-workspace

The name after skore sync is the source project. --to hub makes the Hub the destination, and --to-workspace names that workspace. The source stays local, so its workspace is the one Project would discover.

To see what would come back from the Hub before anything is copied, reverse the direction and stop after the plan:

$ skore sync my-xp --from hub --from-workspace my-workspace --dry-run

Here the Hub project is the source and the local project is the destination. --dry-run prints the plan and leaves both projects unchanged.

The same command can target MLflow. This copies missing reports in both directions between the local project and an MLflow tracking server:

$ skore sync my-xp --to mlflow --tracking-uri http://localhost:5000 --both

--tracking-uri is the MLflow server. --both fills in whichever side is missing a report. How reports are matched is described in Synchronizing projects.

At least one of --from or --to is required; an omitted endpoint defaults to local storage. Use --to-project when the destination project has a different name. Hub endpoints also accept --hub-url, while --from-workspace and --to-workspace select endpoint-specific workspaces. See Command-line interface for the complete option list.

Installing agent skills

skore skills installs Agent Skills from the latest release of probabl-ai/skore-skills. To install every skill into the current project:

$ skore skills install --all

That writes them to .agents/skills, or to the project directory of the agent that is currently running. To install one skill or one workflow:

$ skore skills install <skill-or-workflow-id>

A workflow id installs every skill that workflow includes. Running skore skills install with no arguments opens a picker for the source, the skills, the agent, and the project or user directory. Without a terminal, that same command prints the catalog instead.

To see what is installed:

$ skore skills list

To bring those skills up to the latest release:

$ skore skills update --all

To delete one skill:

$ skore skills remove <skill-id>

The skore-skills package also provides deterministic helpers used by the installed skills. They are available through skore skills tools as an alias for the standalone skore-skills command:

$ skore skills tools --help

Using Skore Hub

Managing credentials

A Hub project and a Hub agent both use a workspace API key stored on your machine. Create one for a workspace:

$ skore hub api-key generate --workspace my-workspace

This opens a device login in the browser, then stores the key for my-workspace. Key generation also accepts options for the Hub host, key name, expiration, login timeout, and replacing an existing key; see Command-line interface. See which keys are already stored:

$ skore hub api-key list

And drop the key for a workspace, locally and on the Hub:

$ skore hub api-key revoke --workspace my-workspace

Project in hub mode still prefers SKORE_HUB_API_KEY over this registry. See Storing data science artifacts.

Connecting a coding agent

skore agent signs in, chooses a workspace and a harness, and launches that harness with the Hub configured. From the project directory:

$ skore agent

The first run asks you to pick the workspace and the harness, saves that choice in .skore, and adds .skore to .gitignore. It also creates an API key when that workspace does not have one yet. Later runs reuse .skore and launch the harness.

If you already know which harness to open, name it:

$ skore agent --harness cursor

Supported harnesses include Bob, Claude, Cursor, Codex, OpenCode, Pi, and Copilot, with CLI and IDE variants where available. The accepted --harness values are listed in Command-line interface. When this command is already running inside the selected harness, it writes the configuration and asks you to start a new session.