radar

Agent skill that tracks competitors, upstream tools and key dependencies of your repo — checks versions, reads changelogs, and tells you what matters for your project.

Works with any agent supporting the Agent Skills spec: Claude Code, Codex, Cursor, Gemini CLI, Copilot, Antigravity.

Install

Claude Code — as a plugin (central install, per-project enable):

/plugin marketplace add mralabs/claude-plugins
/plugin install radar@mralabs

Any agent — via the GitHub CLI:

gh skill install mralabs/radar radar                       # interactive — pick your agent

Or target an agent directly — the skill works anywhere the Agent Skills spec does (45+ hosts):

gh skill install mralabs/radar radar --agent claude-code
gh skill install mralabs/radar radar --agent codex
gh skill install mralabs/radar radar --agent opencode
gh skill install mralabs/radar radar --agent cursor
gh skill install mralabs/radar radar --agent gemini-cli
gh skill install mralabs/radar radar --agent github-copilot
# …plus Amp, Goose, Windsurf, Warp, Cline, Junie and more:
# gh skill install --help for the full list

Default scope is the current repo; add --scope user to install globally.

Requires gh ≥ 2.90 (the release that introduced gh skill) — newer recommended, gh ships frequent security fixes. The bundled CLI runs on bun or plain Node ≥ 22.18 — whichever is already installed. No gh? Cloning this repo's skills/radar/ into your agent's skills directory works the same.

Use

In any repo:

/radar init      # creates .radar/ (git-tracked), proposes what to track
/radar           # check → changelogs → analysis grounded in YOUR repo
/radar add <url> # track a new tool
/radar discover  # find new tools in your categories
/radar deep <x>  # research one tool in depth — tracked or not

Findings are reported, not filed. /radar init asks where 🔥/💡 items should go (a task board, GitHub issues, a spec file) and remembers it — but nothing gets created until you say so.

Deterministic work (version checks against GitHub/npm/PyPI/NuGet, changelog fetching, state) runs in the bundled CLI (scripts/radar.ts, zero deps, bun or node). The agent does the judgment: compares changes against your roadmap and code, produces recommendations. Changelogs come from GitHub releases, falling back to the repo's CHANGELOG.md, then to commits — always flagging when a range may be incomplete.

Optional weekly check via GitHub Actions: /radar init --workflow — keeps a rolling "Radar digest" issue and comments when new updates (or check errors — a silently failing source is a finding too) land. GitHub notifications = your alerting, no infra; run summary in the Actions tab. The installed workflow is two uses: lines running the radar composite action pinned to a commit SHA — nothing floating executes in your CI, and Dependabot's github-actions ecosystem will PR pin updates if enabled.

Want the analysis automated too? The installed .github/workflows/radar.yml ends with a commented-out job step running claude-code-action that reads the updates and comments its analysis on the digest issue. Uncomment it, add an ANTHROPIC_API_KEY repo secret, and pin the action to a reviewed commit SHA (same supply-chain policy as the rest of the workflow).

Data layout (in the consuming repo)

.radar/
├── registry.json   # tracked tools: categories, features, curated notes
├── versions.json   # last-known versions + history
└── config.json     # taskSink — where 🔥/💡 findings get filed (tokens are env-only: GITHUB_TOKEN)

Set GITHUB_TOKEN env to lift the anonymous 60 req/h GitHub API limit.

License

MIT