Guards

Deterministic local compaction for Claude Code: replaces the stock compaction summarizer with the local UltraCompress binary (no model call, no API cost per compaction) and serves a recall tool over the raw session transcript.

claude-hud 0.10.0 as a mod: model, project, git, context, usage, tools, agents and todos, plus alerts, budget, history, a task summary, a detail pane and switchable themes

What each turn did, in a band above the prompt once it ends (files changed with lines, commands and failures, reads, subagents), and a toast when the model goes in circles; read from the turn's own tool calls: no tool, no prompt, no tokens

Animations and a companion pet while the model works: a mascot beside the spinner, a scene above the prompt (a shoot-em-up, Clawd, pixel homages and more), a finale and a pet that levels up, in fifteen themes, one of them a visualizer of the sound the computer plays

The task list's progress in a band above the prompt, read from the todo and task tools' own calls: no tool, no prompt, no tokens

Tailscale nodes' state in a band above the prompt, with a toast when a node comes up or goes down

Upgrade Claude from text search to code understanding.

A Design workspace in Claude Code: see design options for a UI as images, select one, ask for more, ask for a change with a note and hand an option off to be built. Your agent runs the 12ui command line once per request, which draws every option at once with your own Codex sign-in.

Routes the WebSearch tool to Perplexity (pplx CLI) via a function hook; falls back to native search on failure

A persistent pane of actions Claude needs you to run: tick them off, pick an outcome, or reply, and Claude is told

Live panel of what Claude is doing: running tools with real commands and timers, what is waiting for your approval, subagents, and its current todo plan.

Every file Claude created or edited this session, with +/- line counts from git and one-click 'why?', summary and self-review.

Clickable safety rules with presets: block rm -rf, force-push, destructive git, secret files, sudo, installs, network; lock Claude to the project folder or read-only; add your own block patterns.

Notices when Claude is going in circles (the same command failing again and again, the same code rewritten over and over) and tells Claude to step back and rethink, and tells you too.

Notification inbox + native OS notifications when a long turn ends, a subagent or background task finishes, or Claude is stuck waiting for your approval.

Every tool Claude can use (built-in and MCP) in plain language, how often each was used, and suggestions tailored to the project you are in.

Keeps a ledger of what Claude left running or lying around on your machine and servers

Live progress bars above the prompt for multi-step work: one row per plan, with stages, percent, waiting and failed colors, and sounds

Mask token-like strings in tool output before they reach the conversation

Cat mode: an ASCII cat above the prompt that reacts to what Claude is doing, plus status line and toasts.

A small box above the prompt that always says what the agent is doing right now (reading, thinking, writing, running a tool, waiting for you) and expands to show what it has done.

A Minecraft-inspired HUD for Claude Code: hearts for context, armor for safety nets, XP for verified work, and 60 advancements that persist across sessions

Edit .env files in a pane inside Claude Code. Claude manages keys and asks you for values without ever seeing them.

Wakes an idle Claude Code session when its local file mailbox receives an addressed instruction, and records the session's replies to files the sender can read. Independent and experimental.