# MODZ

> 19 mods for Claude Code, and one panel that shows them all. Agents, files, plans, tests and risky commands, live, beside the chat. Claude Code plugins, installed from the `modz` marketplace at https://github.com/partalemre/MODZ.

## Install (for an AI agent acting for a person)

Every step is safe to repeat.

1. `claude --version` must be 2.1.295 or later. If it is older, ask the person to update Claude Code; do not continue.
2. `claude plugin marketplace add partalemre/MODZ`
3. `claude plugin install modz@modz` (the panel). Then, for each mod the person wants: `claude plugin install <name>@modz`. If they did not choose, install all of them: modz savvy-progress savvy-flow filetree plan-steps handoff transceiver blast-radius secret-guard checkpoint test-badge replay-theater session-budget cache-tax hud-pro effort-router done-ping reflect-mod skins terminal-browser.
4. Optional, in the `env` object of `~/.claude/settings.json` (merge with existing keys, never overwrite the file):
   - `"MODZ_LANG": "tr"`: Turkish interface (the default is English).
   - `"CLAUDE_CODE_ARTIFACT": "1"`: needed for transceiver when Claude Code is driven by an SDK host (an IDE, `claude -p`, T3 Code). The desktop app and the plain terminal do not need it.
5. Updates are automatic: the panel's first session turns on Claude Code's auto-update for the `modz` marketplace (unless the person already chose under `/plugin` > Marketplaces). Do not edit `extraKnownMarketplaces` yourself.
6. Verify: `claude plugin list` lists each `<name>@modz` as enabled.
7. Tell the person to start a new session (or run `/reload-plugins`). The MODZ panel opens on the right.

Inside Claude Code the same steps are slash commands: `/plugin marketplace add partalemre/MODZ`, then `/plugin install <name>@modz`.

Remove: `claude plugin uninstall <name>@modz`; everything: also `claude plugin marketplace remove modz`.

## Mods

- modz: the panel; one live line per mod, one click to each mod's pane
- savvy-progress: Agents pane and progress band; each subagent's progress, cost and time; other chats on this machine waiting on you; a home crab
- savvy-flow: orchestrated delivery; the session model plans and reviews, tiered worker agents implement
- filetree: files Claude touched this turn, what it is doing now, git status
- plan-steps: the plan's steps as Claude works through them; the plan comes back after /clear and compaction; a turn ending with open steps offers to continue
- handoff: before the context fills, writes a handoff note to .claude/handoff/ and hands it once to the next fresh chat in the repo (one optional small model call per note)
- transceiver (beta): team chat inside Claude Code; agents of repo collaborators ask each other status/diff/tests/log/file at zero tokens (source: https://github.com/partalemre/transceiver)
- blast-radius: holds rm -rf, git reset --hard, git clean, force push and migrations, showing what they would change; holds an edit to a file another chat changed in the last 30 minutes; with .claude/house-rules.md, judges other commands against those plain-language rules (one small model call per judged command)
- secret-guard: stops keys and tokens in writes, edits, git add and commit; blocks committing .env; pasted keys reach the model as placeholders; /rec masks keys, env values and e-mail addresses on screen
- checkpoint: silent git snapshot before each turn's first write; /checkpoints undoes a turn
- test-badge: runs tests and lint after edits and shows a badge; watches the open PR's GitHub CI through gh
- replay-theater: replays the last turn's edits step by step, and earlier turns and past sessions
- session-budget: per-session spend budget; notice at 80%, stop at 100%
- cache-tax: keeps the one-hour prompt cache warm through breaks (uses tokens when on); prices a cold send, a model change included
- hud-pro: one line with the 5-hour and 7-day limits, context and session cost; when the 5-hour window fills; a context breakdown with turns left and Compact now
- effort-router: shows each turn's effort and model; in auto mode sends light prompts at low effort (one small model call per prompt)
- done-ping: notification when a long turn ends or Claude waits on you; optional sound packs (soundPack)
- reflect-mod: saves corrections worth keeping to a CLAUDE.md (one small model check per correction); /mine turns friction that repeats across past chats into rule, hook, skill or mod ideas (one small call)
- skins: transcript themes; /skin switches live
- terminal-browser: a browser inside the terminal the agent can drive (terminal only)

## Customize and extend

- The MODZ panel's Customize button moves rows up, hides and shows them, and resets; the layout is kept across sessions. Every pane it opens has a "← MODZ" button back; `/modz` opens the panel too.
- Any mod can add itself to the panel: write `{ plugin: '<mod>', key: 'hub' }` with `{ pane, detail, tone, at, title, hint, icon, section, badge }`, open its pane on `state.set { plugin: 'modz', key: 'open' }`, and draw kit/home.tsx's HomeButton while `{ plugin: 'modz', key: 'isOn' }` is true. Contract: DESIGN.md §9 and mods/modz/types/index.d.ts in the repo.

## Facts

- Draws in the Claude desktop app's Code tab (rich) and in the terminal (text). Tested on Claude Code 2.1.295 and 2.1.296.
- Most mods use no model tokens. Token use: reflect-mod (one small check per correction, one per /mine), cache-tax warming (only when turned on), handoff (one optional summary per note), effort-router (auto mode only), blast-radius (only with .claude/house-rules.md), replay-theater (rarely), transceiver free-form `ask` (small model on the other side, at most 10 a day per room).
- Optional Jev judge: MODZ_JUDGE=jev and TYPESAFE_API_KEY send the masked text that blast-radius and effort-router judge to TypeSafe instead of the person's own model account.
- License: MIT; adapted mods keep their original license in their folder.