Public tools Kun owns that meet eligibility: owned, public, not archived, not a fork, at least 80 stars, and self-starred. Use this file to know what exists and what it is for before reaching for something else.
A few captain-requested entries that are not GitHub-star eligibility (for example the Firstmate Grok Bot template) are kept here on purpose. Daily tools sync must not remove them.
https://github.com/kunchenguid/no-mistakes
no-mistakes is a local git proxy and validation gate. You push to a no-mistakes remote instead of origin; it spins up a disposable worktree, runs an AI-driven review/test/docs/lint pipeline, forwards the branch only after checks pass, and opens a clean PR. It is agent-agnostic (claude, codex, grok, and others) and also exposes a TUI plus a /no-mistakes agent skill.
It solves the gap between faster agent/human change production and the need for clean, evidence-backed PRs. Isolation keeps your working tree free while the gate runs; safe mechanical fixes apply automatically, and judgment calls park for you.
Follow the installation guide at https://kunchenguid.github.io/no-mistakes/, run no-mistakes init in a repo, then git push no-mistakes, bare no-mistakes, or /no-mistakes from a coding agent.
https://github.com/kunchenguid/firstmate
firstmate is an agent distro, not a model, harness, skill, MCP server, or installable app. The cloned repo (AGENTS.md, bundled skills, helpers) turns a general-purpose coding agent into a first mate that dispatches a crew of autonomous agents in visible session backends (tmux by default, plus herdr, zellij, cmux, or Orca), each in an isolated git worktree.
It solves tab-juggling when you want several project tasks in parallel. You talk to one liaison; it spawns crewmates, supervises them with event-driven zero-token watchers, and returns finished PRs, approved local merges, or scout investigation reports, with optional persistent secondmates and Relay for X/Discord mentions.
Clone the repo, authenticate gh, launch a verified primary harness inside it (Claude Code, Grok, or Pi are co-primary recommendations; Codex, OpenCode, and Cursor Agent CLI are also supported), and give the first mate work in chat. Git, GitHub CLI, and the chosen session backend are required.
https://x.ai/bot/__4FfrkUdvpdMk6-LKg5r
firstmate-grok-bot is the public Firstmate Grok Bot template ("Firstmate by Kun"). It is a click-import bot on Grok Bot, not a GitHub repo. Once imported, the Firstmate agent is the single agent you talk to; it manages crewmates and the rest of the factory for you.
It solves the gap for people who want the Firstmate experience in Grok Bot without standing up the open-source firstmate repo themselves.
Open the template URL while signed into Grok Bot, click import/add to install Firstmate into your bot list, then talk to that agent in chat. No separate Grok Ship install steps; GitHub gh is not required just to start using the imported bot.
https://github.com/kunchenguid/gnhf
gnhf (good night, have fun) is a ralph/autoresearch-style overnight orchestrator. One command starts an autonomous loop where each iteration has a coding agent make one small, committed, documented change toward an objective, with retries, rollback on failure, live terminal status, and an exit summary.
It solves waking up empty-handed after leaving agents unsupervised. Successful iterations land as commits on a branch with a log of what happened; failures reset or repair instead of leaving a dirty mess, and caps on iterations or tokens keep runs bounded.
Install with npm install -g gnhf or build from source with pnpm. Run from a clean git repo with gnhf and an objective string; optional flags cover max iterations, max tokens, worktrees, and push after each successful commit. Needs a supported coding-agent CLI; macOS, Linux, and Windows. Agent skill at skills/gnhf/SKILL.md covers Hands-Off and Companion modes.
https://github.com/kunchenguid/lavish-axi
lavish-axi (Lavish Editor) is an AXI CLI that opens agent-generated HTML artifacts in a local browser for collaborative review. You can pinpoint elements and selected text, edit Mermaid whiteboard diagrams the agent authored, and send feedback back to the agent without falling into screenshots and long prose change requests.
It solves the weak human-agent loop on rich HTML: agents already produce interactive artifacts, but collaboration usually loses that interactivity. Lavish keeps the loop local-first, token-efficient (TOON, long polling, contextual disclosure), and playbook-backed for common visualization tasks.
Recommended setup is npx skills add kunchenguid/lavish-axi --skill lavish; the skill teaches agents to run npx -y lavish-axi on demand. Alternatives: tell the agent to use the CLI directly, or npm install -g lavish-axi plus lavish-axi setup hooks / setup plugin for ambient session context. macOS, Linux, and Windows.
https://github.com/kunchenguid/axi
axi is the design-principles and catalog repo for AXI (Agent eXperience Interface): ten principles for agent-ergonomic CLIs that treat token budget as a first-class constraint, plus benchmarks and an official/community catalog of AXI tools. Site: https://axi.md/.
It solves the overhead of human-shaped CLIs and heavy MCP tool schemas when agents call external services. Published browser and GitHub benchmarks show reference AXIs beating raw MCP and plain CLI setups on success, cost, duration, and turns.
There is nothing to run as a product binary here; read the principles, then adopt reference AXIs such as gh-axi and chrome-devtools-axi (and point agents at them from AGENTS.md / CLAUDE.md). Use this repo when designing or evaluating agent-facing CLIs.
https://github.com/kunchenguid/treehouse
treehouse is a Go CLI that manages a pool of reusable, isolated git worktrees so each agent session gets a clean environment instantly. Worktrees return to the pool with dependencies and build cache intact instead of being thrown away after every session.
It solves slow, collision-prone parallel agent work: cloning repeatedly, sharing one dirty tree, or recreating worktrees that lose dependency folders and build caches. Conflict detection keeps leased worktrees from stepping on each other.
Follow the project docs for setup (platform script, Nix, Go toolchain, or build from source). From a repo, run treehouse to enter a pooled worktree subshell and exit to return it to the pool. Default root is ~/.treehouse/; --root . keeps the pool in-project. macOS, Linux, and Windows.
https://github.com/kunchenguid/backpass
backpass treats AGENTS.md / project skills as weights and closed agent transcripts as the loss signal. It collects local sessions from supported harnesses, proposes evidence-backed edits (adds, rewrites, removes, skill extractions) under a token budget, and never writes until you run backpass apply.
It solves the open loop where agent memory only improves when a human remembers a failure and edits by hand. Every mutating proposal needs verbatim quotes from at least two distinct sessions; analysis is local-first with secret redaction and no upload except into an agent you already authenticated.
Requires Node >= 22.5 and acpx on PATH. npm install -g backpass or npx backpass, then backpass init, backpass, and backpass apply in a repo (or --scope user for user-level memory). No API keys of its own; model calls go through acpx. macOS and Linux.
https://github.com/kunchenguid/dotfiles
Kun's personal Mac setup managed with nix-darwin and home-manager: system settings, Homebrew casks/CLIs, Nix user packages, zsh/starship, Neovim and WezTerm (rose-pine moon), and shared agent configs so Claude, Codex, and OpenCode load one AGENTS.md.
It solves reproducible machine bootstrap for agentic engineering on Apple Silicon (Intel needs one hostPlatform line change). One repo and one switch path recreate the same environment instead of hand-tuning each Mac.
Clone, review username/host/Homebrew zap settings in "Make it yours", run ./bootstrap.sh on a fresh machine, then ./rebuild.sh for daily applies. Feature PRs are not accepted; bug issues are. This is a personal template to fork and adapt, not a multi-user product.
https://github.com/kunchenguid/chrome-devtools-axi
chrome-devtools-axi wraps chrome-devtools-mcp as an AXI-compliant CLI for browser automation: navigate, click, fill, extract, with combined operations, TOON output, stale-ref checks, and contextual next-step hints after every response.
It solves expensive, turn-heavy browser work through raw MCP or human-shaped browser CLIs. On the axi.md browser benchmark it posts the lowest tokens, cost, duration, and turns among compared setups at 100% task success.
Recommended: npx skills add kunchenguid/chrome-devtools-axi --skill chrome-devtools-axi -g so agents use npx -y chrome-devtools-axi on demand; or global install plus chrome-devtools-axi setup hooks for SessionStart ambient context. Point agents at real-browser tasks; simpler fetch/curl stays preferred for static pages.
https://github.com/kunchenguid/vision
vision is an Agent Skill that mines a repo's merged PRs or commit history, drafts a testable VISION.md acceptance policy, stress-tests it with hard hypotheticals, and iterates with you on a Lavish-backed interactive review board until the document stands alone for accept/resist decisions.
It solves values that live only in shipped history: contributors and coding agents otherwise re-litigate what the project refuses to become. Principles must cite real evidence; generic engineering virtues are banned, and unreadable history makes the skill refuse rather than invent.
Add the skill with the skills CLI from kunchenguid/vision (global recommended), then /vision (or /vision owner/repo) inside the target repo. Needs gh or gh-axi for PR history when available, git history as fallback, and npx -y lavish-axi for the review board. Harness-agnostic.
https://github.com/kunchenguid/gh-axi
gh-axi is the GitHub AXI: a token-efficient wrapper around the official gh CLI for agents, with TOON output, live dashboard defaults, contextual next steps, and structured errors across issues, PRs, stacks, workflow runs, Projects, secrets/variables, gists, and more.
It solves agents burning tokens and turns on raw gh JSON or eager GitHub MCP schemas. On the axi.md GitHub benchmark it is the only compared condition at 100% success, at lower cost than raw gh and far cheaper than GitHub MCP variants.
Install the skill with npx skills add kunchenguid/gh-axi --skill gh-axi -g (agents then use npx -y gh-axi), or install globally and optionally gh-axi setup hooks. Requires Node 20+, gh authenticated via gh auth login; stacked PRs need gh extension install github/gh-stack. macOS, Linux, and Windows.
https://github.com/kunchenguid/whathappened
whathappened is a Grok Build Agent Skill that turns an X firehose around a breaking topic into a short neutral briefing: what happened, where the conversation is, a public opinion map with camps, live debates, and receipt links. Adaptive time windows favor freshness.
It solves dunks, quote-tweet chains, and conflicting "official" takes with no single place to read them. Briefings stay X-first (one optional web lookup only to resolve who/what), with steelmanned camps instead of a vibes paragraph.
Add into Grok Build with the skills CLI from kunchenguid/whathappened (global recommended), or copy skills/whathappened to ~/.grok/skills/whathappened, then /whathappened <topic>. Grok Build only: it needs Grok's native X tools; other agents can add the package but cannot run it for real.
https://github.com/kunchenguid/baby-menu
baby-menu is a macOS menu-bar Electron app whose popover can run your coding agent to edit the menu on the fly. Ask in plain English for a widget or layout; the agent writes an extension under ~/.baby-menu/extensions, hot-reloads it, and offers Keep/Undo based on the real diff.
It solves fixed menu-bar apps that never match the exact mix of system and agent metrics you want (CPU next to Claude usage next to calendar, and so on). Personal software evolves by conversation instead of waiting for someone else's release.
Requires macOS 13+, Homebrew, and an authenticated agent CLI such as claude or codex on PATH. Install with brew install --cask kunchenguid/tap/baby-menu, open the app, and ask in the composer. Settings cover launch-at-login and embedded/custom ACP agents.
SSHHIP is an SSH client app on iOS to help developers control the same terminal sessions on their workstations from their mobile phone.
It makes the control extremely easy and natural by featuring a signature "command dial" that allows the user to use a single finger drag motion to trigger all the common operations and keybinds that are hard to do on mobile. It has native voice input support to help people avoid typing on the mobile keyboard as well.
https://github.com/kunchenguid/quota-axi
quota-axi is a data-only AXI CLI that reports local quota/usage windows for Claude, Codex, Cursor, GitHub Copilot, Grok, Kimi, Z.AI, Alibaba, OpenCode Go, and Antigravity in one TOON-shaped call. It never routes, recommends, proxies, logs in, imports browser cookies, or mints credentials (beyond optional delegated vendor CLI refresh).
It solves agents choosing work without knowing remaining subscription capacity, and humans parsing vendor dashboards that are not shaped for shell automation. Normalized scopes, runway, and attention blocks give routing-aware callers evidence without inventing numbers.
Run npx -y quota-axi (or install globally). On macOS, Claude/Cursor Keychain credentials may need a one-time quota-axi --allow-keychain-prompt with Always Allow. Use --provider, --json, and --full as needed. Local-first against first-party endpoints; macOS, Linux, and Windows.
https://github.com/kunchenguid/compact-adviser
compact-adviser is an agent plugin that judges whether an AI coding session has reached a safe checkpoint for /compact, with hint or opt-in automatic modes on supported hosts.
It solves the problem of compacting too early or too late by using TypeSafe's Jev model to assess whether work is finished and whether the session is hands-on or coordination work, with a usage-aware threshold.
Install it with the plugin flow for Pi, Claude Code, Codex CLI, or Grok Build, and use Node 22+ plus a TypeSafe API key. Pi and Claude Code can opt into automatic compaction; Codex and Grok are hint-only, and platform-specific setup is required.