SlackHive

Build your AI-first company on Slack

A Karpathy-inspired team of humans and AI specialists, powered by Claude Code

Documentation · Quick Start · Features · Architecture · Contributing

Why SlackHive?

A workspace where humans and AI agents work side by side beats any team made of just one or the other.

That's the whole bet. Not "AI is going to replace your team." A genuinely hybrid team — your people plus a roster of role-specific AI agents — collaborating in the same Slack channels, on the same threads, with the same @mention your team already uses for each other.

You don't switch to a separate AI app. You don't open a new tab. You @data-analyst in #sales the way you'd ping a real one. The agent reads the full thread, queries your warehouse, and posts the answer 40 seconds later — while another channel is @designer-ing mockups in Figma and your @devops agent is quietly opening a PR for the 500s the engineer just complained about.

Why this beats the alternatives:

• Vs. an all-human team — slower, more context-switching, "let me get back to you" everywhere
• Vs. one mega-AI assistant (single ChatGPT-style entry point) — no domain expertise, no team patterns, lives outside your workflow
• Vs. AI tools with their own siloed UI — separate interface = friction = nobody uses it after week two

Anyone on the team can create an agent. No engineers, no platform team — if you can describe what you need, you can deploy it in minutes.

CEO:        @data-analyst revenue is down 8% this week, can you dig in?
DataBot:    [queries Redshift across 6 dimensions]
            Found it - enterprise churn spiked Tuesday after the pricing change.
            3 accounts, $42k ARR at risk.

Engineer:   @devops the checkout service is throwing 500s
DevOps:     [reads logs, identifies root cause, opens PR]
            Memory leak in the payment processor pool. PR #847 is up with the fix.

PM:         @designer mock up a simpler onboarding flow
Designer:   [creates Figma frames via MCP]
            Done - 3 variants in Figma. Which direction do you want to take?

Tag a specialist directly when you know who to ask. Or tag @boss when you're not sure - Boss finds the right specialist, delegates, and summarizes the result:

You:        @boss can you analyze last week's conversion funnel?
Boss:       That's right up @data-analyst's alley 👇
            @data-analyst - conversion funnel analysis for last week.
            When you're done, please tag @boss.
DataBot:    Conversions up 12% WoW, checkout completion jumped 3×. @boss - done!
Boss:       Conversions are up 12% WoW. The win was checkout - 3× completion rate.
            Want me to pull a channel or cohort breakdown?

Under the hood: small specialist agents instead of one mega-agent (Karpathy: specialists beat generalists), each grounded in a per-folder Karpathy-style wiki (LLMs read structured wikis better than raw file dumps), with a Boss agent that delegates so no single context window has to hold the whole company. The architecture is a means; the team-in-Slack experience is the point.

🚀 Quick Start

Option A: One-command install (recommended)

npm install -g slackhive

slackhive init

The CLI will:

1. Check prerequisites (Git, Node.js ≥ 20, Claude Code)
2. Clone the repository
3. Auto-detect your Claude installation (cross-platform compatible)
4. Walk you through configuration (API key or subscription, admin credentials)
5. Build and start the services natively - no Docker required

Open http://localhost:3001 and create your first agent.

Hit EACCES: permission denied on install?

If you installed Node from nodejs.org on macOS/Linux, /usr/local/lib/node_modules is root-owned and npm install -g fails without sudo. Point npm at a user-owned prefix once - no sudo needed again:

mkdir -p ~/.npm-global
npm config set prefix '~/.npm-global'
echo 'export PATH=~/.npm-global/bin:$PATH' >> ~/.zshrc   # or ~/.bashrc
source ~/.zshrc
npm install -g slackhive

Installing Node via nvm or Homebrew (brew install node) puts it in a user-writable location by default and avoids this entirely. If you'd rather skip global install, npx slackhive init works too - just prefix every command with npx.

CLI Commands

| Command | Description | |---------|-------------| | slackhive init | Clone, configure, and start SlackHive | | slackhive start | Start all services | | slackhive stop | Stop all services (and any orphaned runner processes) | | slackhive status | Show running services + ports | | slackhive logs | Tail runner logs | | slackhive update | Pull latest changes and rebuild |

Option B: Manual setup

Prerequisites: Node.js ≥ 20, Git, Claude Code on your PATH

git clone https://github.com/pelago-labs/slackhive.git
cd slackhive
cp .env.example .env

Edit .env with your credentials:

ANTHROPIC_API_KEY=sk-ant-...        # or use Claude Pro/Max subscription
ADMIN_USERNAME=admin
ADMIN_PASSWORD=<strong-random-password>
AUTH_SECRET=                        # generate: openssl rand -hex 32
ENV_SECRET_KEY=                     # generate: openssl rand -hex 32
DATABASE_TYPE=sqlite                # embedded - no external DB needed

npm install
npm run build

# Start the stack (web on :3001, runner on :3002)
npx slackhive start

Open http://localhost:3001, log in, and create your first agent.

Full setup guide → slackhive.mintlify.app/quickstart

✨ Features

🤖 Real AI Agents

Every agent is a full Claude Code agent - with tools, memory, identity, and instructions. When you @mention one in Slack, you're running a real AI agent that can use tools, take action, and get smarter over time.

| | | |---|---| | 🧠 Persistent Memory | Agents write memories during conversations - feedback, user context, project state. Synced to SQLite, injected on next start. They don't forget. | | 🔌 MCP Tool Integration | Connect any MCP server (Redshift, GitHub, Notion, Figma, custom APIs) - stdio, SSE, or HTTP transports. | | 📝 Inline TypeScript MCPs | Paste TypeScript source directly into the UI - no deployment needed. The runner compiles and executes it. | | 🧵 Full Thread Context | Agents fetch the entire Slack thread on every invocation - zero context lost in handoffs. | | 💾 Session Continuity | Slack thread ↔ Claude session mapping survives restarts. Pick up exactly where you left off. | | 🔐 Encrypted Secret Store | API keys encrypted at rest (AES-256). MCPs reference secrets by name - raw values never touch the API or UI. | | 🔁 Hot Reload | Edit instructions, skills, or tools and the agent picks up changes within seconds. No restart needed. |

👑 Boss + Specialist Hierarchy

| | | |---|---| | 👑 Boss Orchestration | Boss reads your message, finds the right specialist, delegates by @mention in the same thread, and summarizes the result. | | 🏢 Multi-Boss Support | Run multiple Boss agents for different domains (engineering, data, support). Specialists can report to more than one boss. | | 📋 Auto-Generated Registries | Every Boss gets a live team roster auto-regenerated whenever the team changes. No manual maintenance. | | 🛠 Skills | Markdown files deployed as Claude Code slash commands. Give agents SQL rules, writing guidelines, or domain playbooks. | | ⏰ Scheduled Jobs | Cron-based recurring tasks - daily reports, weekly digests, monitoring alerts - posted to any Slack channel or DM. |

⚙️ Platform

| | | |---|---| | 🧙 Onboarding Wizard | 5-step guided setup: identity → Slack app → credentials → tools & skills → review. | | 🕓 Version Control | Every save auto-snapshots the full agent state. Browse history with line-level diffs, restore any version in one click. | | 🔒 Auth & RBAC | 4 roles (superadmin / admin / editor / viewer), HMAC-signed sessions, per-agent access grants (View or Edit). Agents are hidden by default — admins grant access explicitly. No external auth provider needed. | | 🚦 Channel Restrictions | Lock agents to specific Slack channels. Bot auto-leaves uninvited channels with a notice. | | 📊 Live Logs | SSE-streamed log output per agent - with level filters and search. | | 🧠 Memory Viewer | Browse, inspect, and delete agent memories by type - feedback, user, project, reference. | | 📡 Activity Dashboard | Live kanban of every task your agents are working on - Active, Completed, Errors - with drill-down to every tool call. A superadmin-only Usage tab breaks down token consumption per agent and ranks power users by tasks started. Admins see everything; editors see only agents they own or are granted access to; viewers are blocked. | | 🌐 Multi-Workspace | Connect multiple Slack workspaces to a single SlackHive instance. Each workspace gets its own agents and configurations. | | 📚 Knowledge Library | Platform-level catalog of Karpathy-style wiki folders. Owners ingest a repo / file / URL into a folder; assign one folder to many agents. The agent's /wiki skill auto-enables when at least one assigned folder has built content. |

🏗 Architecture

Slack Workspace (@boss, @data-bot, @writer, ...)
        │ Socket Mode (Bolt)
        ▼
┌──────────────────────────────────────────────────┐
│  Local host - two Node.js processes              │
│                                                  │
│  Web (Next.js :3001) ──HTTP──► Runner (:3002)    │
│            │                        │            │
│            └───── SQLite file ──────┘            │
│                   (./data/slackhive.db)          │
└──────────────────────────────────────────────────┘

| Service | Description | |---------|-------------| | Web (Next.js 15) | Dashboard - create agents, edit skills, view logs, manage users | | Runner (Node.js) | Hosts all agent processes and Slack connections | | SQLite | Embedded - stores agents, memories, skills, sessions, users, history. No external DB to install | | Event bus | Web posts to the runner's internal HTTP port to deliver hot-reload events |

How a message flows:

1. User @mentions an agent in Slack
2. Runner receives the event via Bolt Socket Mode
3. Claude Code processes the message with the agent's compiled CLAUDE.md
4. Agent uses MCP tools if needed (Redshift, GitHub, Notion, etc.)
5. Response is formatted as Slack Block Kit and posted to the thread
6. Memory files written during the session are synced to SQLite
7. Next conversation starts with all accumulated knowledge

🔑 Claude Code Authentication

Two options - use whichever fits your setup:

Option A - API Key

ANTHROPIC_API_KEY=sk-ant-api03-...

Billed per token via the Anthropic API. Best for teams and production.

Option B - Claude Pro or Max Subscription

First, install Claude Code on your host machine:

Note: Do not install via npm - the npm package is deprecated and incompatible with SlackHive.

MacOS/Linux (Recommended):

curl -fsSL https://claude.ai/install.sh | bash

Homebrew (macOS/Linux):

brew install --cask claude-code

For more installation options, uninstall steps, and troubleshooting, see the setup documentation.

Then log in:

claude login    # saves credentials to your system keychain / ~/.claude/

Leave ANTHROPIC_API_KEY unset - the runner picks up credentials from the system keychain automatically. Best for individual developers.

Full guide → slackhive.mintlify.app/configuration/env-vars

🔮 Roadmap

• [x] Boss orchestration + auto-generated team registries
• [x] Persistent memory system
• [x] Scheduled jobs
• [x] Version control with diff view
• [x] Encrypted environment variables
• [x] Channel restrictions
• [x] Multi-workspace support
• [ ] Webhook triggers (GitHub, Jira, PagerDuty → agent actions)
• [ ] Agent-to-agent direct messaging
• [ ] Analytics dashboard
• [ ] Custom tool builder (no MCP server needed)
• [ ] Agent templates marketplace
• [ ] RAG integration - connect agents to document stores

Have an idea? Open an issue

🤝 Contributing

git clone https://github.com/pelago-labs/slackhive.git
cd slackhive && npm install

# Configure
cp .env.example .env    # then fill in ADMIN_PASSWORD, AUTH_SECRET, ENV_SECRET_KEY

# Run locally (SQLite - no external services required)
cd apps/web && npm run dev      # http://localhost:3000
cd apps/runner && npm run dev   # http://localhost:3002

Open an issue before submitting large PRs so we can align on the approach.

👥 Contributors

⭐ Star History

🔒 Security

Use GitHub's private vulnerability reporting - click "Report a vulnerability" on the Security tab. Please don't open public issues for security bugs. We respond within 48 hours.

📄 License

MIT © 2026 Pelago Labs