One file. Zero commands. Makes Claude smart about your codebase.

Drop one folder into any project — PHP, JavaScript, Python, Go, Rust, Ruby, Java, anything — and Claude becomes a senior developer who knows your entire codebase. It reads your code, builds a registry, detects duplicates, learns your conventions, sets up safety guards, and writes a lean CLAUDE.md that persists across every session.

# Clone Claude Boost
git clone https://github.com/Mixvoip/claude-boost.git

# Copy the init folder into your project
cp -r claude-boost/.claude/init your-project/.claude/init

# Let Claude learn your codebase
cd your-project
claude "Read .claude/init/learn.md and execute every task in it"

That's it. Claude handles everything interactively.

composer require mixvoip/claude-boost
php artisan claude:init
claude "Read .claude/init/learn.md and execute every task in it"

When Claude reads learn.md, it runs a 12-step interactive setup:

0. Checks for previous progress — resumes if interrupted, or enters Refresh Mode if already set up
1. Discovers your stack — languages, frameworks, databases, testing tools
2. Asks you questions — permission level, features to enable
3. Drafts CLAUDE.md early — safety net in case of interruption
4. Deep scans your codebase — language-aware scanning (not just file listing)
5. Builds a registry — every class, function, route, model cataloged in grouped markdown
6. Detects duplicates — synonym-aware comparison (30+ synonym groups)
7. Learns your conventions — from your actual code, not imposed rules
8. Installs safety hook — blocks destructive commands via shell hook
9. Maps dependencies — traces who depends on what
10. Creates skills, architecture docs, git standards — module docs, architecture map, branch/commit rules
11. Finalizes lean CLAUDE.md — project essentials with pointers to deep docs
12. Final summary — reports what was created and next steps

If interrupted at any point, just say "continue" or re-paste learn.md. Claude reads learn-progress.json and picks up exactly where it left off.

Already set up? Just run the same command again. Claude enters Refresh Mode — no questions, no setup, just re-scans and updates.

your-project/
├── CLAUDE.md                        <- Lean essentials (~20 lines), loaded every session
├── .claude/
│   ├── .gitignore                   <- Ignores logs/, settings.local.json, learn-progress.json
│   ├── settings.json                <- Permissions, hooks, permission_level
│   ├── registry.md                  <- Every class, service, function cataloged
│   ├── architecture.md              <- Module map, data flow (read on-demand)
│   ├── guidelines.md                <- Conventions learned from your code
│   ├── learn-progress.json          <- Resume tracker (gitignored)
│   ├── init/                        <- The learning prompts & agents
│   │   ├── learn.md                <- Codebase learning (the core product)
│   │   ├── unload.md               <- Clean uninstall guide
│   │   ├── guard-rules.md
│   │   ├── plan.md                 <- Ticket planner agent
│   │   ├── develop.md              <- Autonomous developer agent
│   │   ├── review.md               <- Autonomous reviewer agent
│   │   ├── AGENTS.md               <- Agent pipeline guide
│   │   └── templates/
│   │       └── skill.md
│   ├── guidelines/                  <- Git standards and other guides
│   ├── skills/                      <- Module documentation (read on-demand)
│   ├── plans/                       <- Implementation plans
│   ├── hooks/
│   │   └── preToolUse.sh            <- Safety guard (always active)
│   └── logs/                        <- Guard logs (gitignored)

When your codebase changes significantly:

claude "Read .claude/init/learn.md and execute every task in it"

Claude detects your existing setup and enters Refresh Mode automatically — no questions, no confirmation. It skips setup, conventions, safety hooks, and dependencies (already done), and only re-runs what matters: scan → registry → duplicates → skills/architecture → CLAUDE.md. Fast and silent.

composer update mixvoip/claude-boost
php artisan claude:update

This refreshes learn.md, hooks, and templates to the latest version. Then run the learn command above — Claude will use Refresh Mode automatically.

To cleanly remove Claude Boost from your project:

claude "Read .claude/init/unload.md and execute every task in it"

Claude will:

1. Inventory all Claude Boost files and classify them (ours vs yours)
2. Ask what you'd like to keep (registry, architecture, skills, etc.)
3. Back up everything before making changes
4. Restore your pre-boost CLAUDE.md and settings.json from git history
5. Remove only Claude Boost artifacts
6. Show you the full diff to review before you commit

For Laravel projects, also run composer remove mixvoip/claude-boost after.

The Composer package provides three commands:

Everything else is handled by Claude reading learn.md.

Guard hook (preToolUse.sh) blocks destructive commands in real-time:

• DROP DATABASE, TRUNCATE TABLE, DELETE without WHERE
• rm -rf on critical directories
• git push --force to protected branches
• chmod 777, direct .env manipulation
• Production migrations, curl | bash

The hook is pure bash — no PHP runtime needed. Works for any language.

Claude Boost includes 3 autonomous agents that turn your ticket workflow into a CI/CD-like pipeline:

You + Planner ──> Developer ──> Reviewer ──> You merge

# Tab 1 — Plan tickets with Claude
claude "Read .claude/init/plan.md and execute it"

# Tab 2 — Claude develops autonomously
claude "Read .claude/init/develop.md and execute it"

# Tab 3 — Claude reviews autonomously
claude "Read .claude/init/review.md and execute it"

The agents need a CLI tool to create tickets and manage PRs/MRs:

The agents auto-detect your platform from .claude/settings.json or directory structure (.github/ vs .gitlab-ci.yml).

1. You + Planner discuss the feature → Planner creates a structured ticket with file paths, acceptance criteria, and test plan
2. Developer picks up the ticket → breaks it into sub-tasks → launches parallel agents → opens PR/MR
3. Reviewer picks up the PR/MR → dispatches 3 specialist reviewers → approves or returns with specific fixes
4. You merge — the boring part is automated

Returned tickets go back to the Developer (max 2 returns, then blocked for human intervention). See .claude/init/AGENTS.md for full documentation.

CLAUDE.md instructs Claude to keep everything updated:

• New code created -> Claude updates registry.md
• Module changed -> Claude updates the skill file
• Major feature lands -> Claude updates CLAUDE.md

You don't maintain these files manually. Claude does it during normal development.

Once Claude Boost has learned your project, you can publish the generated .claude folder so your entire team benefits — every developer gets the same registry, conventions, safety hooks, and project brain from the first session.

The simplest approach — just commit the .claude folder to your project repo:

git add .claude/
git commit -m "Add Claude Boost project context"
git push

Sensitive files like learn-progress.json, settings.local.json, and logs/ are already gitignored. Everything else — registry, guidelines, skills, hooks, and CLAUDE.md — is safe and meant to be shared.

1. One developer runs Claude Boost to learn the project
2. Commit the .claude folder to the repo
3. Every team member gets the full context on git pull
4. Claude reads CLAUDE.md automatically — no setup needed for new developers
5. When the codebase evolves, re-run the learning step and commit the updates

This turns Claude from a generic assistant into a team-wide senior developer who knows your entire codebase — and stays in sync.

Most approaches to enhancing Claude Code rely on plugins, background services, or additional AI calls to give Claude context about your codebase. This introduces overhead that works against you:

• Extra token consumption — AI-powered compression and summarization tools make additional API calls on every session, tool use, or prompt. Those tokens add up fast, especially on large projects.
• Runtime dependencies — background daemons, vector databases, and additional runtimes add infrastructure overhead for what should be a zero-friction experience.
• Plugin system lock-in — if the plugin API changes or your environment doesn't support it, the tool breaks. Your project context shouldn't depend on a third-party lifecycle.
• Lossy context — AI-generated summaries lose detail. A compressed memory of your codebase is never as useful as a structured, complete registry.

Claude already knows how to read files — it does it extremely well. Instead of building middleware that summarizes your code for Claude, Claude Boost lets Claude read structured context directly:

A registry, guidelines, and skills — all in plain files that cost zero extra tokens, survive across every session, and work with any language.

The best tools work with the system, not around it.

Any project: Claude CLI installed. That's it.

Laravel package: PHP 8.1+, Laravel 10+, jq, git

Built in partnership with Mixvoip. Thanks for supporting open-source development.

Found a bug or have an idea? Open an issue or submit a pull request.

MIT License

One file. Zero commands. Makes Claude smart about your codebase.

GitHub | Packagist | Mixvoip