Claude follows your rules.With proof.

Claude follows CLAUDE.md ~80% of the time. Hooks run at 100%. Most setups ship with zero hooks. Launchpad scores Claude Code and local Cursor Agent config, fixes the gaps with hooks and permissions, and runs scenarios to prove the agent follows your rules. Cursor Cloud memory is not supported.

$npx claude-launchpad

Open quickstart View npm package

New

Official support

Cursor Agent

Scaffold, diagnose, repair, and evaluate with native Cursor config. Memory is the same local SQLite + MCP — Cloud memory is not supported.

Explore support

new project

init first

$claude-launchpad init

Generate Claude Code or Cursor Agent config (`--harness cursor`).

#Build your project

Write code, add dependencies, shape the architecture.

>/lp-enhance

The agent reads your code and completes CLAUDE.md or AGENTS.md.

$claude-launchpad eval

Run scenarios to verify the agent actually follows your rules.

existing setup

doctor first

$claude-launchpad

Runs doctor automatically when config already exists.

$claude-launchpad doctor --fix

Auto-repair hooks, permissions, and missing sections.

>/lp-enhance

The agent reads your code and completes CLAUDE.md or AGENTS.md.

$claude-launchpad eval

Run scenarios to verify the agent actually follows your rules.

Measured improvement

What changes after doctor --fix

One command finds misconfigurations. One flag fixes them. Your score goes from failing to passing in seconds.

before vs after

31 -> 91

Before31/100

✗Claude forgets context between sessions

✗Credentials (SSH, AWS) readable

✗No .claudeignore, reads noise files

✗Rules followed ~80% of the time

✗BACKLOG drifts, same WP in both files

✗No guardrails or conventions enforced

After --fix91/100

✓ TASKS.md auto-injected every session

✓ Credentials blocked system-wide

✓ .claudeignore filters noise

✓ Hooks enforce rules at 100%

✓ workflow-check warns on BACKLOG/TASKS drift

✓ Path-scoped rules load on the right files

See what doctor checks

Verify behavior

Eval: prove the agent follows your rules

A doctor score of 95% means your config looks good. An eval FAIL means the agent ignores those rules in practice. Eval closes the gap between configuration and actual behavior — Claude Code or Cursor Agent.

17 built-in scenarios across 3 suites: security, conventions, and workflow.

Each run creates an isolated sandbox with that harness's config copied in (.claude/ or .cursor/). Your code is never touched.

Write your first scenario

eval output

sandbox

$ claude-launchpad eval

PASS respects-claudeignore (12s)

PASS follows-off-limits (8s)

PASS uses-conventional-commits (6s)

FAIL blocks-credential-leak (9s)

Score: 3/4 (75%), 1 scenario needs attention

Optional add-on

Your decisions follow you everywhere

Claude's built-in memory resets per machine. Launchpad uses one local SQLite store for Claude Code and Cursor Agent, synced via a private GitHub Gist. Cursor Cloud Agents cannot use this local MCP server.

Device A14 memories

push

›

Private Gistauto-syncs each session

pull

›

Device B14 memories

claude-launchpad memory

One command to set up. Same store for Claude Code and Cursor Agent.

✓ One `memory sync` carries the project brain across devices (private GitHub Gist)

✓ Claude injects memories at session start; Cursor uses memory_search

✓ Stale knowledge fades, important decisions persist

✓ Each project gets its own scoped memory

✓ Retrieval quality is benchmark-gated: 59 tests, mutation-verified — measured, not assumed

Set up memory

memory --dashboard

cockpit [project:my-app | sort:importance][/]=search [p]=projects [1-5]=type

── Memories ────────────────────────────────

 ▸ Never rush irreversible…  SEMA 90% acc:6

   Push back hard on bad…    SEMA 80% acc:2

   CHANGELOG is CLI-only     SEMA 80% acc:4

   Command order - canoni…   SEMA 80% acc:8

   No em dashes in docs      SEMA 70% acc:0

   Architecture decisions    SEMA 60% acc:2

   Eval engine design        SEMA 60% acc:0

   Agent permissions         SEMA 70% acc:4

   Read before judging       SEMA 70% acc:0

── Projects ────────────

  All projects    28 mem

> my-app          14 mem

  api-server       8 mem

  shared-lib       6 mem

── Detail ──────────────

Never rush irreversible

Type: semantic

Health:  ████████░░ 100%

Import:  █████████░ 0.90

28 memories | 4 relations | 1.7MBhealthy:24 fading:3 stale:1

The full picture

Each tool has one job

init scaffolds. doctor maintains. /lp-enhance makes it smart. eval proves it works. memory remembers.

Responsibility	init	doctor --fix	/lp-enhance	eval	memory
Detect stack and framework	✓	-	-	-	-
Generate CLAUDE.md / AGENTS.md, TASKS.md, BACKLOG.md	✓	-	-	-	-
Generate settings.json or hooks.json	✓	✓	-	-	-
Generate .claudeignore / .cursorignore	✓	✓	-	-	-
Generate workflow rules (Claude paths or Cursor globs)	✓	✓	-	-	-
Install workflow-check hook (staleness warnings)	✓	✓	-	-	-
Score config 0-100	-	✓	-	-	-
Fix missing sections, rules, permissions	-	✓	-	-	-
Collapse duplicate ## Memory headings	-	✓	-	-	-
Add stop-and-swarm rule	✓	✓	-	-	-
Rewrite CLAUDE.md / AGENTS.md with real project content	-	-	✓	-	-
Suggest hooks, MCP servers, path-scoped rules	-	-	✓	-	-
Overflow conventions to harness rule files	-	-	✓	-	-
Prove the agent follows your rules	-	-	-	✓	-
Run sandbox test scenarios	-	-	-	✓	-
Persistent cross-session memory	-	-	-	-	✓
Dashboard, sync, migration	-	-	-	-	✓

Read the full docs

get started

docs-first

Zero config to start. Five minutes to a scored setup.

The quickstart walks you through init, doctor, enhance, and eval. Flags, edge cases, and CI integration are in the full docs.

$npx claude-launchpad

Open docs npm package