claude-launchpad — плагин для Claude Code: что делает и как установить
Плагин claude-launchpad проверяет конфигурацию проекта для Claude Code, ставит оценку качеству инструкций и исправляет хуки с правами доступа.
MIT ★ 6 обновлён 2026-09-03
/plugin marketplace add anthropics/claude-plugins-community
/plugin install claude-launchpad@claude-communityЧто делает claude-launchpad
Launchpad сканирует файлы проекта: CLAUDE.md, правила из .claude/rules/ и хуки. Диагностика проверяет объём инструкций. Модель начинает пропускать правила, если файл длиннее примерно 200 строк. Утилита выявляет утечки секретов и неограниченный доступ к ~/.ssh или ~/.aws.
Для порядка в проекте используется трёхфайловая структура. Контекст делится между CLAUDE.md, текущими задачами TASKS.md и бэклогом BACKLOG.md. Запуск команды claude-launchpad doctor --fix автоматически создаёт недостающие правила, восстанавливает хук workflow-check.sh и защищает файл .env.
Для работы сканера сеть не нужна. Вся проверка запускается локально на машине и не тратит токены API.
Какие задачи решает
- разгрузить переполненный
CLAUDE.mdс помощью разделения задач и бэклога - найти утечки секретов и опасный доступ к папкам
~/.sshили~/.aws - настроить хуки автоформатирования и защиты файла
.env - исправить ошибки в настройках транспортов серверов MCP
Примеры запросов агенту
- Выполни команду /lp-enhance, чтобы обновить CLAUDE.md под проект
- Запусти npx claude-launchpad doctor --fix для починки конфигурации
- Инициализируй проект и хуки через npx claude-launchpad init
Как установить claude-launchpad
Через маркетплейс плагинов
/plugin marketplace add anthropics/claude-plugins-community/plugin install claude-launchpad@claude-communityПо шагам, с командами для каждой платформы: Плагины Claude Code: маркетплейсы, установка и как собрать свой
Вопросы о claude-launchpad
Как установить claude-launchpad в Claude Code?
Сначала подключите репозиторий плагинов командой /plugin marketplace add anthropics/claude-plugins-community, затем установите сам плагин через /plugin install claude-launchpad@claude-community.
Что делает claude-launchpad?
Плагин оценивает настройки проекта по шкале до 100 баллов и устраняет уязвимости в правах доступа. Он делит разросшийся контекст на отдельные файлы и настраивает хуки.
Нужны ли ключи API для запуска проверки?
Нет, утилита doctor проверяет конфигурационные файлы автономно на вашем компьютере. Запросы в сеть не отправляются.
Исходный текст
SKILL.md · MIT
# Claude Launchpad
[](https://www.npmjs.com/package/claude-launchpad)
[](https://www.npmjs.com/package/claude-launchpad)
[](https://github.com/mboss37/claude-launchpad)
[](https://github.com/mboss37/claude-launchpad/blob/master/LICENSE)



**Claude follows CLAUDE.md ~80% of the time. Hooks run at 100%. Most setups have zero hooks.**
Launchpad scores your coding-agent config, fixes gaps with hooks and permissions, and runs scenarios to prove the agent follows your rules. Local Cursor Agent projects can be scaffolded, diagnosed, repaired, and evaluated with `--harness cursor`. Memory is the same local SQLite store and MCP server for both harnesses (`memory install --harness claude|cursor|both`). Cursor Cloud memory is not supported.
Cursor can also consume Claude Code files (`CLAUDE.md`, `.claude/skills/`, `.claude/agents/`, supported hooks) when [third-party configuration](https://cursor.com/docs/reference/third-party-hooks) is enabled. That is a migration bridge, not the native target. Launchpad generates native Cursor files (`AGENTS.md`, `.cursor/`) and certifies them with a live Agent canary. Verified against Cursor Agent `2026.08.25-3e8eec8` via `pnpm canary:cursor`. Eval through Cursor Agent is `pnpm canary:cursor:eval`.
For developers using Claude Code or Cursor Agent who want consistent results: solo devs, vibe coders, AI-first teams.
## Install and See Your Score
```bash
npx claude-launchpad
```
```
Instruction Budget ━━━━━━━━━━━━━━━━━━━━ 100%
CLAUDE.md Quality ━━━━━━━━━━━━━━━━━━━━ 100%
Settings ━━━━━━━━━━━━━━━━━━━━ 100%
Hooks ━━━━━━━━━━━━━━━━━━━━ 100%
Rules ━━━━━━━━━━━━━━━━━━━━ 100%
Permissions ━━━━━━━━━━━━━━━━━━━━ 100%
MCP Servers ━━━━━━━━━━━━━━━━━━━━ 100%
Overall ━━━━━━━━━━━━━━━━━━━━ 100%
✓ No issues found. Your configuration looks solid.
```
A typical unconfigured project scores ~31%. After `--fix`, it jumps to ~91%.
## Quick Start
**New project:**
```bash
npx claude-launchpad init
```
Detects your stack, generates config, hooks, and permissions. Start at ~93%.
**Existing project:**
```bash
npx claude-launchpad doctor --fix
```
Scores your config, auto-repairs everything it can.
## The Three-File System + Workflow Rules
Without structure, CLAUDE.md becomes a dumping ground:
- Future ideas bury active guidance
- Sprint notes push conventions off-screen
- Past ~200 lines, Claude starts ignoring rules at the bottom
The three-file split keeps each concern where it belongs:
| File | Purpose | Example |
|---|---|---|
| `CLAUDE.md` | What Claude needs to know | Stack, commands, conventions, guardrails |
| `TASKS.md` | What we're doing now | Current sprint, session log (empty between sprints) |
| `BACKLOG.md` | What we're doing later | WP-NNN template, 7 mandatory fields, P0/P1/P2/P3 sections |
Init generates all three plus a path-scoped workflow rule: `.claude/rules/workflow.md` (Claude) or `.cursor/rules/workflow.mdc` (Cursor). It also installs a `workflow-check.sh` hook that injects drift warnings into the agent's context: a WP entry live in both files, TASKS.md > 80 lines, Current Sprint > 15 items, Session Log > 3 entries, and dependency-blind pulls.
Doctor flags MEDIUM when the workflow rule is missing, LOW when the hook is missing, and MEDIUM on duplicate `## Memory` headings. `--fix` installs or repairs any of them without clobbering existing user content. See the [workflow docs](https://mboss37.github.io/claude-launchpad/docs/workflow) for the full lifecycle.
## Commands
| Command | What it does | Runs |
|---|---|---|
| `claude-launchpad` | Score your config (routes to doctor) | Locally, free |
| `claude-launchpad init` | Detect stack, generate config + hooks + permissions | Locally, free |
| `claude-launchpad init --harness cursor` | Scaffold a local Cursor Agent project (`AGENTS.md`, `.cursor/`) | Locally, free |
| `claude-launchpad doctor --fix` | Auto-fix issues found by doctor (Claude, Cursor, or both) | Locally, free |
| `claude-launchpad doctor --harness cursor` | Score a Cursor Agent project; add `--fix` to repair Launchpad-managed files | Locally, free |
| `claude-launchpad eval` | Run Claude or Cursor Agent against test scenarios | Via the selected harness CLI/SDK |
| `claude-launchpad memory` | Optional knowledge base that persists across sessions | Locally |
| `/lp-enhance` (skill) | The agent reads your code and completes CLAUDE.md or AGENTS.md | Inside Claude Code or Cursor Agent |
## Doctor
Runs analyzers against the selected harness. Claude: `.claude/` and CLAUDE.md. Cursor: `AGENTS.md` and `.cursor/`. Scores stay separate. No API calls, no network, no cost.
**Analyzers:**
| Analyzer | What it catches |
|---|---|
| **Instruction Budget** | Too many instructions. Claude starts ignoring rules past ~200. |
| **CLAUDE.md Quality** | Missing sections, vague instructions, hardcoded secrets, duplicate `## Memory` headings |
| **Settings** | No hooks configured, dangerous tool access without safety nets |
| **Hooks** | Missing auto-format, no .env protection, dead PostCompact hooks (the event doesn't exist — `--fix` migrates to a SessionStart `compact` matcher), missing workflow-check hook, missing jq, no auto-sync on session end |
| **Rules** | Dead rule files, stale references, empty configs, missing `.claude/rules/workflow.md` |
| **Permissions** | Credential exposure (~/.ssh, ~/.aws), blanket Bash approval, bypass-mode unprotected |
| **MCP Servers** | Invalid transport configs, missing commands/URLs |
An optional Memory analyzer runs when agentic memory is detected.
**Flags:**
| Flag | What it does |
|---|---|
| `--fix` | Auto-fix: adds hooks, CLAUDE.md or AGENTS.md sections, BACKLOG.md, rules, .claudeignore / .cursorignore |
| `--fix --dry-run` | Preview fixes without applying them |
| `--watch` | Re-runs every second as you edit config files |
| `--json` | Pure JSON output for scripts and CI |
| `--min-score <n>` | Exit code 1 if score is below threshold (for CI) |
| `-p, --path <dir>` | Run on a different directory |
| `--harness` | `auto` (default), `claude`, `cursor`, or `both`. Scores stay separate. |
## Init
Reads your manifest files (package.json, go.mod, pyproject.toml, etc.) and generates config that fits. No templates, no menus.
```
→ Detecting project...
✓ Found Next.js project
· Package manager: pnpm
✓ Generated CLAUDE.md
✓ Generated TASKS.md
✓ Generated BACKLOG.md
✓ Generated .claude/settings.json (schema, permissions, hooks)
✓ Generated .claude/.gitignore
✓ Generated .claudeignore
✓ Generated .claude/rules/conventions.md
```
**What init writes (Claude, default):**
- Always: `CLAUDE.md`, `TASKS.md`, `BACKLOG.md`, `.claude/settings.json`
- Creates when missing: `.claude/.gitignore`, `.claudeignore`, `.claude/rules/conventions.md`, `.claude/rules/workflow.md`, `.claude/rules/hooks.md`, `.claude/rules/verification.md`, `.claude/agents/code-reviewer.md`
- Offers `/lp-enhance` install (project/global/skip) if not already present
**What init writes (`--harness cursor`):**
- Always: `AGENTS.md`, `TASKS.md`, `BACKLOG.md`, `.cursor/hooks.json`, `.cursorignore`
- Creates when missing: `.cursor/rules/*.mdc`, `.cursor/agents/code-reviewer.md`, `.cursor/skills/lp-enhance/SKILL.md`, hook scripts under `.cursor/hooks/`
- `--force` overwrites `AGENTS.md` only
**Supported stacks:** TypeScript, JavaScript, Python, Go, Ruby, Rust, Dart, PHP, Java, Kotlin, Swift, Elixir, C#. Detects frameworks: Next.js, FastAPI, Django, Rails, Laravel, Express, SvelteKit, Angular, NestJS, and 15+ more.
## Enhance
Init detects your stack but cannot read your architecture. The `/lp-enhance` skill runs inside Claude Code or Cursor Agent to fill in the details.
```
/lp-enhance
```
The agent reads your codebase and updates `CLAUDE.md` or `AGENTS.md` with real content: actual architecture, actual conventions, actual guardrails. Not boilerplate. It also suggests project-specific hooks and MCP servers.
Stays under the 200-instruction budget. Overflows detailed content to `.claude/rules/` or `.cursor/rules/*.mdc`. If the skill is missing, `doctor --fix` will create it.
**When to re-run:** after major refactors, new dependencies, or architecture changes.
## Eval
Runs the selected coding agent against real test scenarios and scores the results.
```bash
# Interactive mode (pick suite, runs, model)
claude-launchpad eval
# Or pass flags directly
claude-launchpad eval --suite security --runs 1 --model haiku
claude-launchpad eval --harness cursor --suite security --runs 1 --model auto
```
If both Claude and Cursor configs are present, `--harness` is required. Each scenario creates an isolated sandbox with only that harness's project config. It runs the agent and checks whether the configuration made it follow the rules.
```
✓ security/sql-injection 10/10 PASS
✓ security/env-protection 10/10 PASS
✓ security/secret-exposure 10/10 PASS
✓ security/input-validation 10/10 PASS
✗ conventions/file-size 5/10 FAIL
✗ Claude kept all generated files under 800 lines
Config Eval Score ━━━━━━━━━━━━━━━━━━━─ 95%
```
Results save to `.claude/eval/` or `.cursor/eval/` as structured markdown. Feed them back to the agent to fix failures.
**Suites:**
| Suite | Scenarios | What it tests |
|---|---|---|
| `security` | 7 | SQL injection, .env protection, secret exposure, input validation, credential read, env exfil via Bash, env-read-attempt |
| `conventions` | 5 | Error handling, immutability, file size, naming, no hardcoded values |
| `workflow` | 5 | Git conventions, session continuity, memory persistence, deferred tracking, premature-victory |
**Flags:**
| Flag | What it does |
|---|---|
| `--suite <name>` | Run one suite: `security`, `conventions`, or `workflow` |
| `-p, --path <dir>` | Project root to evaluate (defaults to cwd) |
| `--scenarios <path>` | Use a custom scenarios directory |
| `--harness <name>` | `claude` or `cursor` (required when both configs exist) |
| `--model <model>` | Model to use: `haiku`, `sonnet`, `opus`, or a Cursor model id / `auto` |
| `--runs <n>` | Runs per scenario (default 3, median score used) |
| `--debug` | Keep sandbox directories for inspection |
| `--json` | JSON output |
| `--timeout <ms>` | Timeout per run (default 120000) |
## Hooks
Instruction-file rules are ~80% reliable. Hooks are 100% enforced by the harness. Init and `--fix` set up these hooks automatically. Claude event names below; Cursor uses the native equivalents (`beforeReadFile`, `beforeShellExecution`, `afterFileEdit`, `postToolUse`, `sessionStart`).
| Hook | Trigger | What it does |
|---|---|---|
| **.env protection** | Before any Read/Write/Edit | Blocks Claude from reading `.env` files that contain secrets |
| **Force-push protection** | Before any Bash command | Blocks `git push --force` to prevent destroying remote history |
| **Auto-format** | After any Write/Edit | Runs your language's formatter (prettier, ruff, gofmt, rustfmt, etc.) |
| **SessionStart** | Session opens | Injects TASKS.md so Claude knows where you left off |
| **PostCompact** | After context compression | Re-injects TASKS.md so Claude doesn't lose sprint state mid-session |
| **workflow-check** | After BACKLOG.md/TASKS.md Edit/Write | Injects context warnings: WP entry in both files, TASKS.md > 80 lines, Current Sprint > 15 items, Session Log > 3 entries, dependency-blind pulls |
| **sprint-size / sprint-open** | Session start / `git commit` | Warns on microsprints (<3 WPs), oversized sprints (>7 WPs), new sprint opened without BACKLOG cleanup |
Claude Code memory projects get three additional hooks. Cursor Agent does not — it uses `memory_search` and explicit `memory sync` / `push` / `pull`:
| Hook | Trigger | What it does |
|---|---|---|
| **SessionStart pull** | Session opens | Auto-pulls memories from GitHub Gist |
| **SessionStart context** | Session opens | Injects the most relevant memories into Claude's context (6-signal scoring + diversity + token budget) |
| **SessionEnd push** | Session closes | Auto-pushes new memories to GitHub Gist |
## Memory (optional)
Claude's built-in memory resets per machine. Launchpad gives each project persistent, cross-device memory that syncs via a private GitHub Gist. Switch laptops and your decisions are already there.
**Why this memory and not another:**
- **Free cross-machine sync.** Native Claude Code memory is single-machine; cloud memory layers need API keys and subscriptions. Launchpad syncs through a private GitHub Gist you already have — zero cost, deletions propagate, machines converge.
- **Measured, not assumed.** The only local agent memory with a benchmark suite (59 tests) gating every release: retrieval quality, injection quality, decay accuracy, scale. `doctor` diagnoses memory health; `eval` proves behavior.
- **Zero infrastructure.** One native dep, stdio MCP, local SQLite. No Docker, no vector DB service, no localhost API server, no cloud account.
- **Engineered injection.** Session context is packed by 6-signal relevance scoring with diversity re-ranking under a token budget — not just "last N memories".
```bash
claude-launchpad memory
claude-launchpad memory install --harness cursor
```
Interactive setup if not installed, stats if it is. The one native dep is installed for you (`npm install -g better-sqlite3`) or on demand. Sync requires the [GitHub CLI](https://cli.github.com/). `--harness claude|cursor|both` writes the MCP file for the selected local harness. Cursor Agent uses `.cursor/mcp.json` and does not auto-inject memories; Cursor Cloud Agents cannot use this local server.
Once set up, `claude-launchpad memory sync` is the everyday command — pull + push in one call, so every machine has the same project brain. Deletions propagate too (tombstones): machines converge to the same memory set, never more, never less. This is the piece Claude Code's built-in memory doesn't do — see [built-in vs Launchpad memory](https://mboss37.github.io/claude-launchpad/docs/memory#built-in-memory-vs-launchpad-memory).
- Claude: relevant memories auto-injected at session start. Cursor: call `memory_search` — no auto-inject. New memories stored as the agent works.
- Stale knowledge fades, important decisions persist
- Each project has its own scoped memory
- `--dashboard` opens a terminal UI with vim nav, filtering, and search
Full flag and subcommand reference in [the memory docs](https://mboss37.github.io/claude-launchpad/docs/memory).
## Use in CI
Block PRs that degrade your agent config quality. `--min-score` gates each harness separately:
```yaml
# .github/workflows/agent-config.yml
name: Agent Config Quality
on:
pull_request:
paths: ['CLAUDE.md', '.claude/**', '.claudeignore', 'AGENTS.md', '.cursor/**', '.cursorignore']
jobs:
config-check:
runs-on: ubuntu-latest
steps:
- uses: actions/checkout@v4
- uses: actions/setup-node@v4
with: { node-version: '22' }
- run: npx claude-launchpad@latest doctor --min-score 80 --json
```
Score below threshold = exit code 1 = PR blocked.
## Glossary
New to the harnesses? Here's what the terms mean.
| Term | What it is |
|---|---|
| **CLAUDE.md** | Claude Code's project instruction file. [Official docs](https://docs.anthropic.com/en/docs/claude-code/memory#claudemd) |
| **AGENTS.md** | Cursor Agent's project instruction file. Same job as CLAUDE.md, different filename. |
| **TASKS.md** | Sprint tracker and session log. Shared by both harnesses. |
| **BACKLOG.md** | Where deferred features live. Priority tiers (P0/P1/P2) keep future ideas organized without cluttering TASKS.md. |
| **Hooks** | Commands that run automatically when the agent does something. Instruction-file rules are ~80% reliable. Hooks are 100% enforced. |
| **Instruction budget** | Soft limit of ~200 actionable lines in CLAUDE.md or AGENTS.md. Past that, rules at the bottom get ignored. Doctor counts them. |
| **Rules** | Extra files in `.claude/rules/` or `.cursor/rules/*.mdc` so the instruction file stays under budget. |
| **Compaction** | Claude Code compresses long conversations. A SessionStart `compact` matcher re-injects TASKS.md. Cursor uses `sessionStart`. |
| **MCP Servers** | External tools. Claude: `.mcp.json` or settings. Cursor: `.cursor/mcp.json`. |
| **.claudeignore / .cursorignore** | Like `.gitignore` for the agent — skip noise files. |
## Privacy
- No telemetry, no analytics, no data sent anywhere
- Doctor, init, and fix run fully offline
- Memory stores data locally at `~/.agentic-memory/`
- Sync (`memory push/pull`) uses a private GitHub Gist under your account
- Enhance runs inside the local Claude Code or Cursor Agent session; eval uses that harness's CLI/SDK
[Full privacy policy](https://mboss37.github.io/claude-launchpad/privacy.html).
## License
MIT
© mboss37 · MIT · mboss37/claude-launchpad