Shellack 🦞

A Claude-powered Slack bot for multi-project development automation

Shellack connects your Slack workspace to Claude AI, giving each project channel its own intelligent agent. Ask questions, start full coding sessions, manage plugins, and track usage — all from Slack.

---

✨ What it does

Capability	Description
💬 Quick chat	@Shellack <question> — project-aware answers with thread context
🖥️ Full sessions	@Shellack run: <task> — interactive Claude Code session in a thread
🎨 Canvas output	Code blocks route to a Slack Canvas instead of flooding the thread
🤖 Project agents	Each channel gets a dedicated agent loaded with CLAUDE.md context
🧠 Token Cart	Haiku-powered context compaction — replaces full-history replay, ~57% token savings
🛡️ Agent teams	Infosec + architect consultants review responses when triggered
✅ Gut check	Sanity check against project registry before posting responses
📋 Project registry	Auto-maintained index of reusable components, patterns, and rules
🔄 Cross-thread memory	Handoffs persist across threads via .shellack/thread-memory/
🐛 GitHub issues	Bugs and crashes auto-open issues with correct labels
🔍 Peer review	Staged Quality / Security / Performance review in #code-review
📔 Journal	Narrative journal entries via GitHub Discussions (weekly threads)
📱 App Store Connect	Auto-monitors reviews and TestFlight feedback
🔌 Plugin management	Install Claude plugins, MCP servers, and bot extensions from Slack
⚙️ Live config	Switch AI mode/model/features without restarting the bot
💰 Cost tracking	Per-turn token spend displayed in the Churned block
🌉 Terminal bridge	claude-slack — respond to Claude Code prompts via Slack buttons on any device

---

🚀 Quick start

git clone https://github.com/your-org/Shellack.git
cd Shellack
python3 -m venv venv && source venv/bin/activate
pip install -r requirements.txt
cp .env.example .env
# Edit .env with your credentials
python bot_unified.py

---

⚙️ Configuration

Slack app

1. Go to api.slack.com/apps → Create New App → From a manifest
2. Paste slack-app-manifest.yml — this sets all scopes and event subscriptions automatically
3. Install to workspace and copy SLACK_BOT_TOKEN + SLACK_APP_TOKEN

Environment variables

# .env
SLACK_BOT_TOKEN=xoxb-...
SLACK_APP_TOKEN=xapp-...
ANTHROPIC_API_KEY=sk-ant-...

# GitHub integration — enables issue auto-creation
GITHUB_TOKEN=ghp_...
GITHUB_ORG=your-org            # Your GitHub username or org

# Slack user ID for operator escalations (Profile → ··· → Copy member ID)
OWNER_SLACK_USER_ID=U0XXXXXXX

# Peer review channel
CODE_REVIEW_CHANNEL_ID=C0XXXXXXX

# AI mode — "api" (Anthropic API) or "max" (Claude Max subscription, zero API cost)
SESSION_BACKEND=api
SESSION_MODEL=claude-sonnet-4-6

# Project paths (defaults to ~/Repos/<Name>)
# MY_PROJECT_PROJECT_PATH=~/Repos/MyProject
# ... see .env.example for all variables

# App Store Connect (optional)
APP_STORE_CONNECT_KEY_ID=...
APP_STORE_CONNECT_ISSUER_ID=...
APP_STORE_CONNECT_PRIVATE_KEY_PATH=~/.appstoreconnect/AuthKey_XXXXX.p8

Project & channel routing

Copy projects.example.yaml to projects.yaml and fill in your projects:

cp projects.example.yaml projects.yaml
# Edit projects.yaml with your channels, paths, and repos

# projects.yaml (gitignored — your config, not committed)
github_org: your-org
projects:
  myapp:
    name: MyApp
    primary_channel: myapp-dev
    language: swift
    platform: ios
    github_repo: your-org/MyApp
    path: ~/Repos/MyApp
    # Optional: Token Cart feature flags
    features:
      token-cart: true        # context compaction (default: on)
      gut-check: true         # sanity check (default: on)
      consultants: true       # infosec + architect (default: on)
      registry: true          # pattern index (default: on)
      agent-manager: false    # model selection (default: off)

See projects.example.yaml for the full schema with all options.

---

💡 Usage

Quick chat — @Shellack <question>

Ask anything about the project in its channel. The bot uses the project's CLAUDE.md as context and maintains conversation history per thread.

@Shellack what does the NetworkClient class do?
@Shellack why might login fail on iOS 17?
@Shellack show me recent commits touching the auth flow

The :claude: emoji appears on your message while the bot is thinking, and disappears when it replies.

---

Full session — @Shellack run: <task>

Starts an interactive Claude Code session in the thread. Claude can read files, write code, run tests, and commit — just like the CLI. Output is streamed back to Slack in real time; code blocks go to a session Canvas.

@Shellack run: fix the crash in LoginView
@Shellack run: add unit tests for the cart total calculation
@Shellack run: refactor NetworkClient to use async/await

Continue the session by replying in the thread (no @mention needed). Stop it with stop or cancel.

Sessions time out after 30 minutes of inactivity (warned at 15 and 25 minutes).

---

Config commands

These work in any channel, any context:

Command	Effect
@Shellack set mode max	Switch to Claude Max subscription (zero API cost)
@Shellack set mode api	Switch to Anthropic API
@Shellack set model opus	Use claude-opus-4-6
@Shellack set model sonnet	Use claude-sonnet-4-6 (default)
@Shellack set model haiku	Use claude-haiku-4-5
@Shellack config	Show current mode, model, and onboarding status
@Shellack config show	Show all Token Cart feature flags for this channel
@Shellack config <feature> on/off	Toggle a feature at runtime (e.g., config gut-check off)
@Shellack usage	Show this month's session and mention counts

All changes take effect immediately — no restart required.

Configurable features: token-cart, gut-check, registry, consultants, external-handoff, cost-observability, code-review, agent-manager

---

Plugin management

Command	Effect
@Shellack plugins	List all installed Claude plugins, MCP servers, and bot extensions
@Shellack add plugin <name>	Install a Claude Code plugin
@Shellack remove plugin <name>	Uninstall a Claude Code plugin
@Shellack add mcp <name> <command>	Register an MCP server
@Shellack remove mcp <name>	Remove an MCP server
@Shellack add bot-plugin <git-url>	Install and hot-reload a bot extension
@Shellack remove bot-plugin <name>	Uninstall a bot extension

---

Peer review

Triggers automatically when an agent edits or creates files, or opens a GitHub issue. Posts to #code-review:

1. Stage 1 — Quality, Security, and Performance agents review in parallel
2. Stage 2 — Maestro selects ≤2 peer agents from projects sharing the same platform/language

Blocking findings tag the operator in the thread.

---

GitHub issues

When a crash or bug is detected, the agent automatically:

• Opens an issue in the correct repo with appropriate labels (crash, bug, p0, p1)
• Posts the issue link in the Slack thread and top-level in the project channel
• Closes the issue when the fix is confirmed

---

🌉 Terminal bridge (claude-slack)

Run claude-slack instead of claude from any project repo. Claude Code posts Block Kit button messages to the project's Slack channel when it needs input — click a button on your phone, tablet, or any device to feed the answer back.

# From any project repo
claude-slack

# Claude Code runs as normal, but prompts appear in Slack

Sessions are UUID-scoped so concurrent sessions never cross-contaminate. Stale button clicks show an ephemeral error only to you.

---

📊 AI cost modes

Mode	Set with	Monthly cost	Notes
API	set mode api	Pay-per-token	Anthropic API, full control
Max	set mode max	$0 extra	Uses your Claude Max subscription via CLI

Both modes route through the same backend — switching affects quick replies and run: sessions alike.

---

🗂️ Project structure

Shellack/
├── bot_unified.py          # Main bot — entry point
├── orchestrator_config.py  # Project registry & channel routing
├── orchestrator.py         # Cross-project operations
├── peer_review.py          # Staged peer review system
├── agents/
│   ├── project_agent.py    # Per-project agent with CLAUDE.md context
│   └── sub_agents.py       # Crash / Testing / Review / Docs sub-agents
├── tools/
│   ├── token_cart.py       # Haiku-powered context compaction + gut check
│   ├── registry.py         # .shellack/registry.md management
│   ├── thread_memory.py    # Cross-thread persistence
│   ├── cost_tracker.py     # Per-turn/thread cost tracking
│   ├── consultants.py      # Infosec + architect consultants
│   ├── agent_manager.py    # Complexity-based model selection
│   ├── github_journal.py   # GitHub Discussions journal posting
│   ├── journal_polisher.py # Sonnet journal polish
│   ├── session_backend.py  # APIBackend + MaxBackend + quick_reply()
│   ├── slack_session.py    # run: session lifecycle, canvas routing
│   ├── lifecycle.py        # Structured Slack status posts
│   ├── github_client.py    # Issue creation & management
│   ├── plugin_manager.py   # Claude plugins, MCP, bot extensions
│   ├── config_writer.py    # Live .env updates
│   ├── usage_tracker.py    # Monthly usage counters
│   └── journal_writer.py   # JOURNAL.md entries
├── app_store_connect.py    # App Store Connect monitoring
├── claude-slack            # Terminal bridge wrapper script
├── .env.example            # All supported env vars with docs
├── slack-app-manifest.yml  # Slack app manifest (scopes + events)
└── SETUP_GUIDE.md          # Detailed setup walkthrough

---

🧠 Token Cart — Multi-Agent Architecture

Shellack uses a three-tier model hierarchy to minimize costs while maximizing quality:

Model	Role	Cost
Haiku 4.5	Token Cart — context compaction, gut checks, correction detection	$0.25/MTok
Sonnet 4.6	Consultants (infosec, architect), journal polish, output editing	$3/MTok
Opus 4.6	Primary reasoning — main agent work	$15/MTok

Instead of replaying full conversation history (quadratic token cost), the Token Cart compacts each turn into a structured handoff. By turn 10, this saves ~57% on tokens.

How it works per turn:

1. Haiku enriches context from prior handoff + project registry
2. Reasoning model (Opus/Sonnet) handles the task
3. Gut check verifies response against registry
4. Consultants review if security/architecture triggers detected
5. Haiku compacts the turn into a new handoff (async, non-blocking)
6. Corrections auto-update the project registry

All features are independently toggleable via @Shellack config <feature> on/off.

---

🔐 Security

• Tokens stored in .env (gitignored)
• Pre-commit hook scans for 16 secret patterns (Slack tokens, API keys, AWS, GitHub, private keys)
• Owner-only gates on plugin/MCP/config commands (fail-closed when OWNER_SLACK_USER_ID is unset)
• Each channel is isolated to its configured project directory
• Bot extensions require a full Git URL — no implicit registry installs
• All destructive operations (file writes, issue creation) require an active agent task
• Error messages sanitized — no exception details leak to Slack
• Triage separates system prompts from user input (prompt injection prevention)

---

📮 Support

• 📚 Setup guide
• 🐛 Report issues
• 💬 Discussions

---

📝 License

MIT — see LICENSE for details.