skore-cli

Command-line interface for skore.

skore-cli installs a single skore command with three areas:

• skills — discover, install and manage Agent Skills from the probabl-ai/skills catalog
• agent — connect a project to the Skore Hub agent, write harness config and launch a local coding agent
• sync — synchronize report projects across local storage, Skore Hub, and MLflow

Installation

pip install skore-cli

The base install is batteries-included: it bundles the agent feature (so it pulls in skore). No extras are required.

Usage

Skills

Install skills into the current project by default. Pass --global/-g for a user-wide install and --agent/-a to target specific agents (agents, claude-code, cursor, codex, gemini, windsurf, cline, roo, amp, github-copilot, bob, bob-ide). Pass --repo owner/name on install to pull from another GitHub catalog that ships .catalog.json (the default is probabl-ai/skills). Interactive skore skills install also lets you edit that owner/name in the wizard; --repo only pre-fills the field. list and update reuse the stored source and show it in their output; they do not take --repo.

agents is the cross-client default (.agents/skills). Other names write the tool's native SKILL.md directory. Most global installs mirror that folder under your home directory; Windsurf uses ~/.codeium/windsurf/skills, Amp uses ~/.config/agents/skills, and GitHub Copilot has no user-level skills directory (--global is rejected). Bob Shell and Bob IDE share .bob/skills.

skore skills list          # list installed skills (with source)
skore skills install       # install skills (interactive or by id)
skore skills install --repo acme/skills gamma
skore skills update        # update from each skill's recorded source
skore skills remove        # remove installed skills

Agent

On the first run, skore agent logs in when needed, lets you pick a workspace and harness, creates a workspace API key, writes the harness configuration and launches the agent. Supported harnesses: Bob Shell, Bob IDE, Claude, Cursor, OpenCode, Pi, GitHub Copilot and Codex (all must be on PATH; on macOS, Bob IDE is found via its application bundle). Bob IDE also accepts --harness bobide — the name of the command its installer puts on PATH — as an alias for --harness bob-ide. Later runs reuse .skore in the project directory (gitignored). Use SKORE_HUB_URI (or --hub-url) to point at a non-default hub.

Launching a harness exports the .skore credentials as SKORE_HUB_API_KEY and SKORE_HUB_URI, so skore.login() in the scripts the agent runs authenticates with that key instead of opening a browser. Values already set in your environment are left untouched.

skore agent
skore agent --harness claude    # non-interactive harness choice
skore agent --workspace ./myapp # configure another project directory

Sync

Synchronize a source project to a destination mode. The source defaults to local when only --to is set; the destination defaults to local when only --from is set.

# Local -> Hub (add --both or --dry-run as needed)
SKORE_HUB_API_KEY=... skore sync experiment --to=hub --to-workspace=team

# Hub -> local with a different project name
SKORE_HUB_API_KEY=... skore sync production \
  --from=hub --from-workspace=team --to-project=downloaded

# Local -> MLflow
skore sync experiment --to=mlflow --tracking-uri=http://localhost:5000

Hub synchronization requires SKORE_HUB_API_KEY. Use --hub-url to target a custom Hub API. Install skore[mlflow] to synchronize with MLflow.

Agent detection

When skore is run inside a coding agent, it detects the agent from environment variables and adapts its behavior:

• skore (no args) shows an agent-specific quick-start with the detected agent's skill directory and harness
• skore skills install (no args, non-interactive) prints the catalog and the detected agent's skill directory — no --agent flag needed
• skore skills install all installs all skills into the detected agent's directory (also works with --all)
• skore skills install <ids> installs specific skills into the detected agent's directory
• skore agent (no --harness, non-interactive) auto-selects the detected agent's harness and skips the launch step (the agent is already running)

Agent	Env Var
Claude Code	CLAUDECODE
Cursor	CURSOR_AGENT
Gemini CLI	GEMINI_CLI
Codex CLI	CODEX_SANDBOX
Pi	PI_CODING_AGENT
OpenCode	OPENCODE_CLIENT

Any non-empty value triggers detection.

License

MIT