Sklm

Skills manager for AI agents

Why Sklm? • Quickstart • Usage • Updating • Supported Agents • How it Works • Development

Why Sklm?

Sklm lets you install AI agent skills once and activate them per project. No more copy-pasting SKILL.md files between directories.

• Install a skill globally and every project sees it, even when it's irrelevant.
• Drop it in one project and it's invisible to others.

Sklm keeps a global library in ~/.sklm/, then lets you pick which skills each project sees.

Note

Sklm only supports skill resources for now. More resource types may come later.

Features

• Works with 30 AI agents — from OpenCode and Claude Code to Codex CLI, GitHub Copilot, and beyond.
• Install once, scope per project — a global store at ~/.sklm/ holds your skills; per-project symlinks activate only what you need.
• Auto-sync — sklm add and sklm rm automatically update the agent's skills directory. No manual copying.
• Registry discovery — index local folders or git repos as searchable skill catalogs.
• Git repo installation — sklm add --from clones a repo and figures out where the skill lives.
• Per-agent skill variants — a single skill can ship agent-specific file overrides in a variants/ subdirectory. Each agent receives the version tuned for it.

Installation

pip install sklm-cli

Tip

For development, clone the repo and use pip install -e . for an editable install.

Quickstart

pip install sklm-cli                   # install globally
sklm                               # interactive wizard opens — detects your setup

That's it. The CLI's interactive wizard detects your AI agents, initializes the workspace, and guides you through adding your first skill — no flags needed.

Tip

Run sklm init --agent opencode to skip the wizard and set a specific agent. Pass --agent multiple times for multiple agents.

Usage

Workspace setup

sklm init                          # auto-detect or prompt for agent(s)
sklm init --agent opencode         # force a specific agent
sklm init --agent claude --agent cursor   # multiple agents at once
sklm status                        # show workspace health
sklm status --repair               # fix broken symlinks

If no agent directory is detected, Sklm shows an interactive prompt. Select one or more agents (e.g. 1,3,5) or press c to skip.

Global store (install once, activate anywhere)

sklm install skill find-skills \
  --from https://github.com/vercel-labs/skills
sklm uninstall skill find-skills               # remove from global store
sklm uninstall skill find-skills --force       # skip confirmation

sklm migrate                                   # import all from ~/.agents/skills/
sklm migrate skill find-skills                 # import a single skill
sklm migrate --from-registry my-reg            # import from a local registry
sklm migrate --force-cleanup                   # delete sources after import

Project resources (activate per project)

sklm add skill my-skill                        # resolve → store → link → sync
sklm add skill my-skill \
  --from https://github.com/user/skills        # install from git + activate
sklm ls                                        # list active resources
sklm ls --json                                 # machine-readable output
sklm info skill my-skill                       # origin, path, link status
sklm rm skill my-skill                         # unlink + clean agent config

Skill variants (authoring)

A skill can ship agent-specific overrides using a variants/ subdirectory inside the skill. When synced, the base skill is copied first, then any files from variants/<agent-id>/ are merged on top.

my-skill/
  SKILL.md                 # fallback for any agent
  references/
    tools.md
  variants/
    opencode/
      SKILL.md             # overrides root SKILL.md for OpenCode
    claude/
      SKILL.md             # overrides it for Claude Code
      references/
        claude-only.md     # additional file, only for Claude

• Files in the variant override same-named files from the base.
• Files only in the variant are added.
• Files only in the base pass through untouched.
• variants/ itself is never copied to the agent's config directory.
• If no variant exists for an agent, the base skill is used as-is.

Variant directory names match agent IDs (opencode, claude, cursor, windsurf, gemini, cline, amazon-q, codex, github-copilot, and all others listed in Supported Agents).

sklm info skill <name> lists available variants when present.

Registry discovery

sklm registry add ~/my-skills                           # local folder
sklm registry add https://github.com/org/skills.git     # git repo
sklm registry ls                                        # list registries
sklm registry search scraper                            # search all registries
sklm registry search scraper --registry my-skills       # within one registry

You can also reference skills by registry when adding:

sklm add skill my-registry:my-skill

Agent management

sklm agent detect                    # identify the active agent
sklm agent list                      # list all known agents
sklm agent add opencode              # add an agent post-init (syncs skills)
sklm agent remove claude             # remove an agent (cleans skills)
sklm agent sync                      # force re-sync all linked skills
sklm agent sync --dry-run            # preview without applying

Telemetry

Anonymous usage data via Umami. Opt out anytime.

sklm telemetry status               # check if enabled
sklm telemetry off                   # disable
sklm telemetry on                    # re-enable

Telemetry is also disabled by setting SKLM_TELEMETRY=0 in your environment.

Updating

sklm checks for new versions automatically after every command (once per day). When a new release is available, a notice is shown with upgrade instructions.

sklm update                         # upgrade to latest version via pip
sklm update --check                 # check only, no upgrade
sklm update --force                 # bypass 24h cache

Disable the automatic check by setting:

export SKLM_NO_UPDATE_CHECK=1

Updates are installed via pip install -U sklm-cli. The version check uses the GitHub Releases API.

Supported Agents

Agent	Config dir	Skills path	Auto-detected
OpenCode	.opencode/	.opencode/skills/	✅
Claude Code	.claude/	.claude/skills/	✅
Cursor	.cursor/	.cursor/skills/	✅
Windsurf	.windsurf/	.windsurf/skills/	✅
Gemini CLI	.gemini/	.gemini/skills/	✅
Cline	.cline/	.cline/skills/	✅
Amazon Q	.amazonq/	.amazonq/skills/	✅
Bob Shell	.bob/	.bob/skills/	✅
CodeBuddy	.codebuddy/	.codebuddy/skills/	✅
Codex CLI	.codex/	.codex/skills/	✅
Continue	.continue/	.continue/skills/	✅
Crush	.crush/	.crush/skills/	✅
Factory Droid	.factory/	.factory/skills/	✅
iFlow	.iflow/	.iflow/skills/	✅
Junie	.junie/	.junie/skills/	✅
Kilo Code	.kilocode/	.kilocode/skills/	✅
Kimi CLI	.kimi/	.kimi/skills/	✅
Kiro	.kiro/	.kiro/skills/	✅
Lingma	.lingma/	.lingma/skills/	✅
Pi	.pi/	.pi/skills/	✅
Qoder	.qoder/	.qoder/skills/	✅
Qwen Code	.qwen/	.qwen/skills/	✅
Trae	.trae/	.trae/skills/	✅
Mistral Vibe	.vibe/	.vibe/skills/	✅
Auggie	.augment/	.augment/skills/	✅
CoStrict	.cospec/	.cospec/skills/	✅
ForgeCode	.forge/	.forge/skills/	✅
RooCode	.roo/	.roo/skills/	✅
Antigravity	.agent/	.agent/skills/	— (explicit only)
GitHub Copilot	.github/	.github/skills/	— (explicit only)

Antigravity and GitHub Copilot require sklm init --agent <name> because .agent/ and .github/ exist in many projects unrelated to those tools.

Tip

sklm init without --agent shows an interactive prompt if no agent directory is found. Use --agent for non-interactive setups.

How it Works

Sklm manages three locations to keep skills organized:

~/.sklm/                 # global store (user-wide)
  store/skills/          #   installed skill directories
  config.yaml            #   resource catalog
  registries.yaml        #   registry sources
  cache/                 #   cloned git repos

./.sklm/                 # per-project workspace (gitignored)
  sklm.yaml              #   project config (agents, links, resources)
  links/skills/          #   symlinks → ~/.sklm/store/skills/

<agent-dir>/skills/      # agent-visible copies (auto-synced)
                         # e.g., .opencode/skills/

Running sklm add skill my-skill does four things in sequence:

1. Resolve — finds the skill in the global store, a registry, or a local path
2. Store — copies it into ~/.sklm/store/skills/ if it wasn't there already
3. Link — creates a symlink in ./.sklm/links/skills/
4. Sync — copies the linked skill into the agent's config directory, applying any variants/<agent>/ overlay automatically

Removal (sklm rm) reverses steps 3 and 4. The global store is untouched, so skills stay available for other projects.

Development

pip install -e .                    # editable install
pip install -r requirements.txt     # dev dependencies (pytest, pytest-cov)
python3 -m pytest tests/            # run the test suite
python3 -m pytest tests/ -k <pattern>   # run a subset
sklm --version                      # check installed version

Troubleshooting

"No Sklm workspace found" Run sklm init first. It creates the .sklm/ directory and configures your agent.

"No agent configured — not synced" The skill is installed and linked, but no agent is set up to receive it. Run sklm init --agent <name>.

"Broken symlinks" Run sklm status --repair to re-create links that point to missing targets.

"Skill not found in git repo" Some repos use non-standard layouts. Use --subdir to point to the exact directory:

sklm add skill my-skill --from https://github.com/user/repo --subdir custom/path

"GitHub Copilot isn't detected" That's expected. Copilot requires explicit setup: sklm init --agent github-copilot.

"sklm registry add fails" For git registries, make sure git is installed and the URL is accessible. For local paths, use an absolute or ~-expanded path.