npx kimi-boost init → detects your stack → installs guardrails + conventions.

If it's ever saved you from a bad rm -rf or a push to main, a ⭐ tells others it's worth finding.

English · 中文文档

⚡ Why

Your AI coding agent only knows what you teach it. Unguided, it writes generic code, pushes straight to main, and runs rm -rf without flinching. Hand-configuring skills, hooks and agents takes hours — so nobody does it.

kimi-boost installs a complete, opinionated workflow in seconds:

| You get | What it does | |---|---| | 🧠 Skills | Best-practice rules your agent auto-loads — no prompting required | | 🔍 Reviewer agents | Read-only subagents your agent delegates to before committing | | ⚡ Slash commands | One-shot workflows like /<preset>:review — installed to ~/.claude/commands/ (Claude Code) or pre-positioned for Kimi Code | | 🛡️ Guards | Cross-platform Node hooks: dangerous commands, trunk protection, secret scanning | | 🔄 One-command updates | kimi-boost update keeps every preset current, even on forks |

$ kimi-boost install vue3
✓ [kimi] Installed preset 'vue3' into Kimi Code
  /Users/you/.kimi-boost/hooks/vue3
  config.toml[extra_skill_dirs], config.toml[extra_agent_dirs], config.toml[[hooks]] (+1)
  Run /reload or start a new session.

$ kimi-boost doctor
✓ kimi: detected (version 0.36.1)
✓ kimi: config.toml parses
✓ kimi: hook script valid
✓ kimi: mounted dir present
All checks passed.

🛡️ Guardrails — the seatbelt you can see

Most agent safety is invisible: a hook silently blocks something and you never learn your agent almost ran rm -rf /. kimi-boost makes the seatbelt visible.

Nine guards in the core preset (included by default in kimi-boost init, or /plugins install …/kimi-boost-core with no CLI at all):

| Guard | What it blocks | |---|---| | protect-main | Direct pushes to main/master | | block-dangerous | rm -rf /, mkfs, dd to disk, curl \| sh | | git-destructive | reset --hard, clean -f, checkout -- . (workspace-wiping) | | secret-scan | Hardcoded secrets in written files (AWS keys, private keys, tokens) | | protect-credentials | Reading ~/.ssh, ~/.aws/credentials, *.pem… into model context | | protect-paths | Hand-editing lockfiles; writes into node_modules/, dist/, .git/ | | protect-guards | The agent disabling guardrails or rewriting their config | | secret-scan-post | A PostToolUse re-scan that catches secrets written via shell redirection | | injection-scan | Prompt-injection patterns in tool output (warn-only — alerts, never blocks) |

Visible, tunable, honest:

• Every block lands in ~/.kimi-boost/guard-log.jsonl (previews redacted — never the secret itself). kimi-boost guard shows status + block counts; guard --log lists recent blocks; kimi-boost stats reports 🛡️ N blocks on your share card.
• Tune live without reinstalling: guard --disable/--enable <name> toggles, --warn/--block <name> softens to warn-only, --explain <name> tells you what it blocks and the risk of disabling it, --add-pattern <regex> teaches block-dangerous your own patterns.
• Same guards in CI: guard --ci [--staged] scans changed files for secrets and exits non-zero on findings; guard --install-git-hook wires that into a pre-commit hook.

Two layers, not one. Hooks are the soft layer — scripts we install. On Claude Code, core additionally writes a curated permissions.deny list into ~/.claude/settings.json: the hard layer, enforced by the harness itself so the agent can't route around it. And injection-scan watches tool output for prompt-injection patterns (warn-only — it alerts, never blocks, since false positives are inherent there).

[!IMPORTANT] What the guardrails don't promise. Pattern matching is not a security boundary — sandboxing is. These hooks raise the cost of accidents; a determined adversary (or a prompt-injected agent) can obfuscate around them (base64 | sh, scripts, runtime-assembled secrets). Everything fails open on purpose: a hook error never blocks your work. secret-scan-post informs — the write already happened. Guards are per-tool. Layer them with host-side secret scanning, review, and a sandbox.

📦 Presets

Each preset is one directory — a valid kimi.plugin.json plugin AND a kimi-boost preset:

presets/<id>/
├── preset.json          # kimi-boost metadata
├── kimi.plugin.json     # Kimi Code plugin manifest (native marketplace form)
├── skills/<name>/SKILL.md
├── agents/<name>-reviewer.md
└── hooks/<name>.mjs     # cross-platform Node, fail-open by design

By stack:

| Preset | Stack | Reviewer | Guard | Mirror | |---|---|---|---|---| | vue3 | Vue 3 + TypeScript | ✅ | 🛡️ main | ✅ | | react | React + TypeScript | ✅ | 🛡️ main | ✅ | | go | Go | ✅ | 🛡️ main | ✅ | | python | Python | ✅ | 🛡️ dangerous shell | ✅ | | nextjs | Next.js (fullstack) | ✅ | 🛡️ main | via CLI | | react-native | React Native | ✅ | 🛡️ main | via CLI | | flutter | Flutter / Dart | ✅ | 🛡️ main | via CLI | | uniapp | uni-app (cross-platform) | ✅ | 🛡️ main | via CLI | | weapp | WeChat Mini Program | ✅ | — | via CLI | | nestjs | NestJS backend | ✅ | 🛡️ main | via CLI | | express | Express (Node.js) | ✅ | 🛡️ main | via CLI | | fastapi | FastAPI | ✅ | 🛡️ main | via CLI | | rust | Rust | ✅ | 🛡️ main | via CLI | | java | Java (Spring Boot) | ✅ | 🛡️ main | via CLI |

Special:

| Preset | What it gives you | Mirror | |---|---|---| | core 🛡️ | The baseline insurance every project should have — the 9 guardrails above (injection-scan is warn-only). Default in init | ✅ | | usage | Sessions / prompts / tool calls tracked into ~/.kimi-boost/usage.json; daily limit hint via KIMI_BOOST_DAILY_LIMIT; view with kimi-boost stats | ✅ | | security | Deeper guard pack: scans Write/Edit for secrets, blocks git push --force/--delete + security-reviewer agent | via CLI | | git-workflow | Conventional commits, branch naming & PR discipline (skill) + reviewer agent | via CLI | | mcp-tools | Zero-config MCP servers: fetch + time | via CLI |

[!NOTE] Every preset bundles a best-practice SKILL.md (auto-loaded) + a reviewer agent. New stacks are vote-driven — issue #1. "via CLI" presets get their own official repo as demand grows.

🔧 Commands

Install & manage

| Command | What it does | |---|---| | init | Detect this project's stack and install matching presets — always includes the core guardrails (--yes, --dry-run, --project) | | install [preset] | Install by id — or from any GitHub repo: install github:owner/repo (--dry-run, --with-hooks, --project, --yes) | | inspect <preset> | Preview a project-level install (components, hosts, hooks/MCP/deny/scripts, network, skipped) without writing | | verify --project | Compare .kimi-boost/installed.json with installed files; non-zero on drift; does not delete edits | | remove <preset> | Uninstall cleanly | | list | Show available + installed presets | | status | Detect installed CLIs & platform | | doctor [--fix] | Diagnose config, hooks, mounted dirs, manifest consistency, duplicate hooks |

Guardrails

| Command | What it does | |---|---| | guard | Guardrail status + block history | | guard --log [-n N] | Recent blocked actions | | guard --disable/--enable <name> | Turn a guard off/on — live, no reinstall | | guard --warn/--block <name> | Soften to warn-only / restore blocking | | guard --explain <name> | What it blocks, how to bypass, the risk of disabling | | guard --add-pattern <regex> | Teach block-dangerous a custom pattern | | guard --ci [--staged] | Scan changed/staged files for secrets; non-zero exit on findings | | guard --install-git-hook | Wire --ci --staged into a pre-commit hook (--uninstall-git-hook removes it) |

Updates

| Command | What it does | |---|---| | update | Pull latest versions and re-apply (works on forks; community presets update from source repo) | | update --dry-run | Preview the version + file-level diff without writing | | update --check | Check without installing; notifies + exits non-zero if found | | update --watch [--interval h] [--uninstall] | Periodic background check (LaunchAgent/cron/schtasks) | | outdated [--project] [--json] | Installed presets with newer registry versions |

Stats & sharing

| Command | What it does | |---|---| | stats [-d N] [--share] | Usage report with bar chart, streak & top tools; --share exports an SVG card (alias: usage) | | badge [preset] | Print a README badge (markdown) | | export [file] | Export your preset setup to a shareable file (--embed-content, --include-usage) | | import <file> | Restore a previously exported setup (--yes, --dry-run) |

Authoring & team

| Command | What it does | |---|---| | create <id> | Scaffold a new preset under presets/ (--shape skill\|mcp\|command) | | validate <dir> | Validate a preset directory | | dev <dir> | Validate + preview-install a local preset (dry-run) | | package <dir> | Validate + pack a preset into <id>-<version>.zip | | marketplace | Generate a Kimi Code custom marketplace JSON | | bootstrap [--makefile] | Generate a setup.sh / Makefile target for team onboarding |

🌐 Install channels

① Official Kimi Code plugin channel — no CLI required. Flagship presets are mirrored to single-plugin repos:

/plugins install https://github.com/shidesheng0218/kimi-boost-vue3

Mirrors: vue3 · react · go · python · usage · core — the guardrails, installable in Kimi Code with no CLI at all.

Or browse them in the /plugins panel via our marketplace feed:

export KIMI_CODE_PLUGIN_MARKETPLACE_URL=https://raw.githubusercontent.com/shidesheng0218/kimi-boost/main/marketplace.json

② kimi-boost CLI — all 19 presets, three harnesses (Kimi Code, Claude Code, Codex CLI), with updates, doctor checks and clean uninstalls:

npx kimi-boost install

| | Official channel | kimi-boost CLI | |---|---|---| | Presets | 6 flagship (mirrored) | All 19 | | Harnesses | Kimi Code | Kimi Code · Claude Code · Codex | | Needs | Just Kimi Code | Node.js | | Extras | — | update · doctor · stats · dry-run |

③ Community presets — anyone can publish a preset as a plain GitHub repo (preset.json at the root):

kimi-boost install github:owner/repo          # or the full URL
kimi-boost install github:owner/[email protected]   # pin a branch or tag

kimi-boost shows exactly what the preset registers (especially its hooks, which execute when the agent runs) and asks before installing. Community presets are tracked in update / outdated against their source repo. Only install presets from authors you trust — hooks run shell commands with your agent's permissions.

📊 Shareable stats — your AI-coding wrapped

The usage preset quietly counts your sessions / prompts / tool calls. stats turns it into a report worth screenshotting:

$ kimi-boost stats -d 7
📊 kimi-boost stats · last 7 days
  128 prompts  ·  12 sessions  ·  856 tool calls  ·  5 active days  ·  streak 3 🔥

$ kimi-boost stats --share   # → kimi-boost-stats.svg

--share exports a self-contained SVG card (no server, no upload — your data never leaves your machine):

📤 More

• Project-level installs (team sharing) — install <preset> --project writes skills/agents into .agents/ + .claude/ and a lock at .kimi-boost/installed.json (version, source, content hash, targets, skipped). Commit it. Hooks land in Claude project settings only; Codex is skipped (no project-level config); MCP servers and Claude permissions.deny stay user-level and are recorded as skipped. injection-scan in core is warn-only.

kimi-boost inspect vue3          # preview, writes nothing
kimi-boost install vue3 --project
kimi-boost verify --project      # non-zero if files changed or the lock is stale

• Export & import — kimi-boost export captures presets + versions + community sources + the update watch into one shareable file; import restores it on another machine. Dotfiles-friendly.
• Authoring presets — validate / dev / package cover the loop; push the directory to GitHub and anyone can install github:you/my-preset.
• doctor — kimi-boost doctor diagnoses config, hooks, mounted dirs and manifest consistency; --fix restores what's missing.

🏗️ How it works

flowchart TD
    REG["<b>presets/ registry</b> (19 presets)<br/>skills · agents · hooks · mcp"]

    subgraph CI["Monorepo CI"]
      VERIFY["verify.yml<br/>live-installs every preset<br/>into a real Kimi Code CLI"]
      SPLIT["split-presets.yml<br/>git subtree mirror"]
    end

    REG --> CI
    SPLIT -->|"auto-mirror"| M1["kimi-boost-vue3"]
    SPLIT -->|"auto-mirror"| M2["react · go · python · usage · core"]
    M1 & M2 -->|"/plugins install"| KP["Kimi Code /plugins"]
    REG -->|"builds (repo sources)"| MKT["marketplace.json"]
    MKT -->|"KIMI_CODE_PLUGIN_MARKETPLACE_URL"| KP

    KB["<b>kimi-boost CLI</b><br/><i>install · remove · doctor · update</i>"]
    REG --> KB
    KB -->|"text-level edit"| KCF["~/.kimi-code/config.toml"]
    KB -->|"manifest-driven"| CCS["~/.claude/settings.json"]
    KB -->|"manifest-driven"| CXC["~/.codex/config.toml"]

    classDef cli fill:#7c3aed,color:#fff,font-weight:bold;
    classDef tool fill:#1e293b,color:#e2e8f0;
    classDef out fill:#064e3b,color:#a7f3d0;
    class KB,REG cli;
    class M1,M2,MKT tool;
    class KCF,CCS,CXC,KP out;

• Single source of truth — presets live in this monorepo; the six flagship mirrors are read-only CI artifacts, re-synced on every push.
• Kimi Code — the CLI edits ~/.kimi-code/config.toml at text level (a managed # >>> kimi-boost managed >>> block plus in-place array merge). Your comments and formatting survive untouched.
• Claude Code & Codex — manifest-driven install into ~/.claude / ~/.codex; agent files use the native frontmatter format.
• Hooks are plain Node .mjs — identical behavior on macOS / Windows / Linux.
• Compatibility is tested, not assumed — CI live-installs all 19 presets into a real Kimi Code CLI on every PR, and weekly against upstream drift.

🛡️ Safety by default

• 🔒 Never touches your config beyond its own managed section — comments, ordering, formatting all preserved
• 🗄️ Backed up to <config>.kboost.bak before every change
• 🚧 Managed-roots whitelist — refuses to delete anything outside ~/.kimi-boost, ~/.kimi-code, ~/.claude, ~/.codex
• 🛡️ Dual-channel guard — already installed via Kimi /plugins? No duplicate hooks.
• ♻️ Content-aware hook dedup — presets bundling the same guard script share a single config entry; uninstalling one preset retargets the entry to the next owner. doctor flags redundant or diverging copies.
• ⚡ Fail-open hooks — a crashing hook never blocks your work (exit 0 allow · exit 2 block)

🗺️ Roadmap

• [x] MCP server presets
• [x] Token/cost usage guard hooks
• [x] Official-channel distribution (single-plugin repo mirrors)
• [x] Per-project (.kimi-boost/) presets
• [ ] More stack presets (vote in issue #1)
• [ ] Mirror every preset to its own official repo

Contributing

The preset catalog is PR-driven: add a directory under presets/, CI validates schema, hook events and file existence — then live-installs your preset into a real Kimi Code CLI. See CONTRIBUTING.md.

MIT · built with ⚡ by kimi-boost contributors