kasperjunge/agr-cli

agr CLI

agr is the package manager for AI agent skills. It installs, shares, and syncs SKILL.md folders across Claude Code, Cursor, Codex, OpenCode, GitHub Copilot, and Antigravity. This skill helps you operate the agr CLI on the user's behalf — set up new repos, install and sync skills, manage agr.toml / agr.lock, and scaffold in-repo skills under skills/.

When to use

Use this skill when the user wants to:

• Install or remove a remote skill (agr add anthropics/skills/pdf).
• Sync agr.toml after pulling teammates' changes (agr sync).
• Upgrade an installed skill past its pinned commit (agr upgrade).
• Run a skill once without persisting it (agrx ...) or invoke an installed one (agr run).
• Scaffold a new skill (agr init my-skill) — especially as in-repo skills under skills/.
• Set up a repo from scratch to use agr (multi-tool config, instruction syncing, sources).
• Add a local in-repo skill to the project's dependency list.
• Inspect, edit, or troubleshoot agr.toml / agr.lock / agr config.
• Or whenever agr.toml is present and the user is doing resource-management work.

Do NOT use this skill for:

• Authoring the body of a SKILL.md. This skill scaffolds and registers; for writing effective skill instructions, defer to the user or to a dedicated authoring skill such as anthropics/skills/skill-creator.
• Iterating on an existing skill based on session feedback or a retrospective. That's a separate workflow — listen, propose, edit, commit, re-install. If the user has a dedicated debrief / improvement skill installed (e.g. skill-debrief), use it; otherwise just do the workflow directly.

Prerequisites

Before running anything, verify agr is installed:

agr --version

If missing, the standard install is uv tool install agr. Do not install agr without checking with the user first — they may prefer pipx, brew, or a pinned version in CI.

Mental model — read first

• A skill is a folder containing a SKILL.md. agr copies the whole folder into each configured AI tool's skills directory (.claude/skills/, .cursor/skills/, .opencode/skills/, .agents/skills/, etc.).
• A handle identifies a remote skill: user/skill (assumes a repo named skills), user/repo/skill (any repo), or ./local/path (on disk).
• agr.toml is the manifest (hand-edited or written by agr add). agr.lock is the lockfile (auto-generated, pins commit SHAs and content hashes — never edit by hand).
• Project-local vs global scope: every command takes -g to operate on ~/.agr/agr.toml and global tool dirs.
• agr sync only installs missing deps. To move past a pinned commit, use agr upgrade.
• Resource types: skill (consumed by AI tools), ralph (autonomous agent loop), package (a folder with its own agr.toml, expanded transitively). agr add auto-detects the type.

For the full conceptual model: references/handles.md.

Workflows

1. Set up a new repo for agr

When the user wants to start using agr in a project that has none configured:

1. Run agr init. agr auto-detects which AI tools the repo uses (.claude/, CLAUDE.md, .cursor/, .cursorrules, .codex/, .opencode/, .github/copilot-instructions.md, .gemini/) and writes agr.toml.
2. Confirm or adjust the detected tools. For multi-tool: agr config set tools claude codex opencode.
3. If the user maintains a canonical instruction file (e.g. CLAUDE.md) and wants it mirrored to others (AGENTS.md, GEMINI.md), enable instruction syncing: agr config set sync_instructions true then agr config set canonical_instructions CLAUDE.md.
4. Commit agr.toml. (agr.lock will appear after the first agr add or agr sync — commit it too.)

Full details: references/setup.md.

2. Install a skill

agr add anthropics/skills/pdf

Notes:

• Multi: agr add user/skill1 user/skill2 (skills from the same repo are batched into one download).
• --overwrite / -o to replace.
• -g for global (available in every project).
• agr add auto-creates agr.toml if missing — no need to agr init first.

For handle formats and private repos: references/handles.md and references/installing-skills.md.

3. In-repo skills (the recommended workflow for skills shipping with the codebase)

Place project-specific skills under skills/ at the repo root (or at the root of a relevant submodule). This keeps the skill in version control, reviewable in PRs, and sharable across the team via agr.toml.

agr init my-skill                  # scaffolds my-skill/SKILL.md in CWD
mkdir -p skills && mv my-skill skills/
agr add ./skills/my-skill          # records {path = "./skills/my-skill", type = "skill"} in agr.toml

Iterate: edit skills/my-skill/SKILL.md, then agr add ./skills/my-skill --overwrite to reinstall into each configured tool.

Teammates pick it up with agr sync after pulling. The local path dependency travels with the repo, so contributors don't need network access to use it.

For structuring (scripts/, references/, assets/), iteration patterns, and testing with agrx: references/in-repo-skills.md.

4. Sync after pulling teammates' changes

agr sync

This (1) syncs instruction files, (2) runs directory migrations, (3) installs any deps missing for any tool, (4) refreshes agr.lock.

CI flags:

• agr sync --frozen — install exactly what agr.lock says; fail if lock is missing or doesn't cover all deps. Use in deploy pipelines for byte-identical installs.
• agr sync --locked — fail if agr.lock is stale vs agr.toml. Use in PR checks to enforce lockfile hygiene.
• --frozen and --locked are mutually exclusive.

Full lifecycle: references/syncing.md.

5. Upgrade past pinned commits

agr sync only installs what is missing. To pull the latest upstream code for a skill that is already installed, use agr upgrade:

agr upgrade                                # everything in scope
agr upgrade pdf                            # short-name match (errors on ambiguity)
agr upgrade anthropics/skills/pdf          # full handle

Gotcha: upgrading one skill from a multi-skill repo (e.g. anthropics/skills/pdf) does NOT refresh siblings. Run agr upgrade with no args, or name each sibling explicitly, to refresh the whole repo together.

6. Run a skill

agr run pdf                               # invoke installed skill in default tool
agr run pdf -- "summarise report.pdf"     # extra prompt
agr run pdf -i                            # interactive
agr run pdf --tool cursor                 # override tool
agrx anthropics/skills/pdf -p "Extract"   # ephemeral, no install

agr run requires the skill to be already installed. agrx downloads, runs, and cleans up. Difference and tool-resolution rules: references/running-skills.md.

7. Configure agr

Keys: tools, default_tool, default_owner, default_source, sync_instructions, canonical_instructions, sources.

agr config show
agr config set tools claude codex opencode
agr config add tools cursor
agr config remove tools cursor      # ⚠ DELETES that tool's skills directory

Private repos / custom Git hosts:

agr config add sources gitlab --url "https://gitlab.com/{owner}/{repo}.git"
export GITHUB_TOKEN="..."           # auth for private GitHub repos

Add -g to operate on ~/.agr/agr.toml instead of the project file.

All keys with defaults: references/configuration.md.

8. List and inspect

agr list           # status of every dep: installed / partial / not synced / invalid
agr list -g        # global skills
cat agr.toml       # the manifest
cat agr.lock       # the lockfile (READ ONLY)

partial (claude, cursor) means a skill is installed in some tools but not others — agr sync to fan it out to the rest.

Boundaries

• Never edit agr.lock by hand. It's regenerated by agr add/agr sync/ agr remove/agr upgrade. If it looks wrong, run agr sync (or agr sync --locked to confirm hygiene).
• agr config remove tools <name> deletes that tool's skills directory. Confirm with the user before running it. The skills can be re-installed by re-adding the tool.
• Don't author the body of a new skill on the user's behalf unless they explicitly asked. This skill scaffolds and registers; SKILL.md content authoring is a separate task.
• Always confirm before --overwrite on a skill the user has been editing locally — overwriting will replace their working copy.
• Don't install agr globally without checking — the user may have a pinned version, alternative install method, or CI constraints.
• Don't push or git tag as part of these workflows. Commit changes locally; let the user push.

Troubleshooting pointers

• "Skill not found" → check the handle format (2 vs 3 parts), see references/handles.md.
• "agr.lock out of date" in CI → user added a dep but didn't commit the lockfile; run agr sync locally and commit.
• partial install status → agr sync to fan out to all configured tools.
• Auth errors on private repos → check GITHUB_TOKEN is exported.
• Type errors after manually editing agr.toml → run agr add to let agr rewrite the entry with the correct type field.

Full list: references/troubleshooting.md.

References

• setup.md — repo bootstrap, multi-tool, instruction syncing
• handles.md — handle formats, resolution, resource types, scopes
• installing-skills.md — agr add / agr remove, sources, private repos
• syncing.md — agr sync / agr upgrade, lockfile, CI patterns
• in-repo-skills.md — skills/ workflow, scaffolding, iterating
• configuration.md — all agr config keys, sources, scopes
• running-skills.md — agr run vs agrx
• troubleshooting.md — common errors and fixes

For the canonical CLI reference, point the user at https://computerlovetech.github.io/agr/reference/ or run agr <command> --help.

Related capabilities (separate skills if installed)

• Greenfield skill authoring — scaffolding a brand-new SKILL.md from a blank slate. Canonical option: anthropics/skills/skill-creator.
• Retrospective-driven skill improvement — listening to feedback after a session and updating an existing skill. If the user has a dedicated workflow skill for this, use it; otherwise the workflow is editing the source under skills/<name>/, committing, and agr upgrade <name>.