Opinionated agentic stack, one script deep, bakes in most efficient (and safe) token-reduction tools around.

Install

Requirements

radin itself only copy files. Companion tools pull own stacks, each gated behind y prompt.

For You need
radin core (agents + skills) curl, tar, bash
Claude plugins (caveman, ponytail) claude CLI
rtk Homebrew, or curl for rtk's own installer
code-review-graph, headroom python3 with pip3 or pipx

Homebrew optional. When present, radin use it for rtk. Not required on Linux.

Homebrew Python note (macOS). Broken [email protected] bottle can make every pip/pipx install fail with pyexpat / libexpat symbol error. install.sh detect this, print fix: brew reinstall --build-from-source [email protected]. Plain brew reinstall reinstall same broken bottle — --build-from-source flag relinks Python against Homebrew's expat.

# macOS · Linux · WSL
curl -fsSL https://raw.githubusercontent.com/shortcuts/radin/main/install.sh | bash

Update

# macOS · Linux · WSL
curl -fsSL https://raw.githubusercontent.com/shortcuts/radin/main/install.sh | bash -s -- --force

The backlog lifecycle

Repo's backlog live inside repo, at .claude/.radin/backlog/ from repo root: index file plus one markdown file per task. Every radin tool read from or write to it through radin's own CLI, never need look inside. Run /radin-show read it as plain markdown. Commit .claude/.radin/ share backlog with team, or add to .gitignore keep private. Radin never touch your .gitignore either way.

Typical flow:

  1. Capture. Something come up mid-session — bug, idea, feedback from teammate. Run radin-record turn it into backlog entry.
  2. Plan (optional). Point radin-plan at one entry write step-by-step plan for it, without touching any code. Repeat per entry you want planned ahead of time.
  3. Execute. Run radin-execute work through backlog, entry by entry, committing as it goes.
  4. Review. Run radin-review against commit, PR, or directory. Every finding become new backlog entry, ready for next pass of step 3.

Tools you get

Homemade

Tool What it does
radin-execute Chews through backlog, one task at time, commits as it goes
radin-plan Writes plan for one backlog entry you point it at, instead of touching code
radin-review Strict code-quality pass, findings logged straight back into backlog
radin-record Logs feedback/bugs/ideas raised mid-session as backlog entries
radin-show Prints current project's backlog
radin-doctor Checks radin's own install complete, reports which companion tools reachable
radin-setup-hooks Wires up per-repo hooks/MCP config for companion tools
radin-stats Shows each installed companion tool's own stats/gain output, side by side
radin-uninstall Removes everything install.sh added to ~/.claude

Some delegate to other skills instead of reimplementing review or style logic themselves:

Tool Delegates to
radin-execute /ponytail (plan-or-skip gate, per-task implementation), /radin-plan (only for tasks judged complex enough), /caveman-commit (commit message), /radin-review (optional end-of-session review)
radin-plan /ponytail (split judgment and plan writing), /thermo-nuclear + /ponytail-review (reviewing plan itself before handoff)
radin-review /thermo-nuclear (code-quality pass), /ponytail-review or /ponytail-audit (over-engineering pass)

radin-record

Log something raised mid-conversation, before lost.

/radin-record log the auth timeout bug we just found

Result: new fix-classified task added to backlog, bug described enough detail future session act on it, no other context needed.

radin-plan

Write plan for one backlog entry, without writing any code. Judges whether entry's scope should split into multiple independent plans, confirms with you before splitting.

/radin-plan the auth timeout bug

Result: one plan file per plan under .claude/.radin/plans/ at repo root (more than one if entry split). Each plan reviewed with /thermo-nuclear and /ponytail-review before handoff, findings fixed directly in plan file, **Plan:** <path> line appended to task's own backlog file per plan produced.

radin-execute

Work through backlog end to end: prioritize, implement, test, commit — one entry at time. Uses existing plan from radin-plan if entry already has one. If not, asks /ponytail whether task straightforward enough implement directly. Only tasks judged genuinely complex go through /radin-plan first.

/radin-execute

Result: each entry implemented and committed in its own commit. Finished entries removed from backlog; failed ones stay, marked for retry. At end can optionally run /thermo-nuclear review of session, log findings back to backlog as new entries.

radin-review

Run strict quality review over chosen scope, log findings to backlog instead of printing to terminal.

/radin-review #123

Also accepts commit hash, directory path, or natural-language range like "commits since Monday". Result: one backlog task per finding, classified as fix (real bug) or refactor (structural).

radin-setup-hooks

Wire up per-repo config for companion tools — currently just code-review-graph's MCP registration and hooks.

/radin-setup-hooks

Run once per project, right after install.sh, in repo you want wired. Previews exact files it will touch, asks confirmation before writing anything.

Vendored in (optional)

install.sh just asks if you want them — never forked, never vendored, own repo stays source of truth.

Tool What it does
rtk CLI proxy reduces LLM token consumption 60-90% on common dev commands. Single Rust binary, zero dependencies
caveman Why use many token when few token do trick — Claude Code skill cuts 65% of tokens by talking like caveman
ponytail Makes AI agent think like laziest senior dev in room. Best code is code never written
code-review-graph Local-first code intelligence graph for MCP and CLI. Builds persistent map of codebase so AI coding tools read only what matters
headroom Local-first context-compression stack — proxy/MCP/wrap layer with cross-agent memory and CLAUDE.md-learning. Complements rtk (whole-session wrap vs. rtk's per-command compression), not replacement. Python/pip footprint — install prompts extra confirmation
thermo-nuclear Code quality review skill, vendored from cursor/plugins at install time via vercel-labs/skills CLI
mattpocock-skills Engineering skills plugin (claude-plugins-official marketplace) — radin-plan delegates interview step to /grilling, API/library fact-checking to /research instead of reimplementing them

Maintaining or hacking on radin itself? See AGENTS.md and CONTRIBUTING.md.