🦞 Corellis

We run a 28-lobster AI team that handles ops, marketing, releases, and weekly reports. This is the system behind it.

Turn one OpenClaw assistant into a self-managing AI workforce — lobsters that coordinate tasks, learn from their mistakes, and get better every week.

Production-tested since February 2026 · 28 lobsters · 50,000+ Slack messages indexed · 500+ self-corrections · single server

A Real Example

Last month we tested fleet-wide goal coordination for the first time. The founder typed one message:

"Launch a user acquisition campaign for the new product line."

Here's what happened over the next 2 hours — with zero human intervention after that single message:

1. The controller decomposed the goal into 6 sub-goals and created a milestone + 7 tracking cards on the task board
2. 6 lobsters were assigned — marketing, engineering, payments, ops, partnerships, frontend — each in a dedicated thread
3. All 6 confirmed within minutes, following the goal-participant protocol: analyze scope → break into subtasks → define acceptance criteria
4. Spontaneous cross-team coordination emerged — the payments lobster and the partnerships lobster started aligning on API design in their thread (4 rounds of back-and-forth). The marketing lobster and the ops lobster resolved scope boundaries on their own
5. One lobster independently created 5 subtask cards with goal IDs, owners, dependencies, and deadlines — without being asked

Result: 36 task cards on the board — 1 milestone, 7 sub-goals, 28 self-created subtasks. Two cross-team interfaces aligned. All from one sentence.

How It Works

Each team member gets their own AI assistant (a "lobster") running in a Docker container — with private memory and conversations. Behind the scenes, they share company knowledge and a searchable team memory.

┌─────────────────────────────────────────────┐
│  Your Server                                │
│                                             │
│  🎛️ Controller (your OpenClaw)              │
│  "Spawn a lobster for alice"                │
│                                             │
│  ┌──────────┐ ┌──────────┐ ┌──────────┐   │
│  │ 🦞 alice │ │ 🦞 bob   │ │ 🦞 carol │   │
│  │ Marketing│ │ Ops      │ │ Finance  │   │
│  └────┬─────┘ └────┬─────┘ └────┬─────┘   │
│       └─────────────┴─────────────┘         │
│         Shared knowledge & team memory      │
└─────────────────────────────────────────────┘

Key Capabilities

🧠 Teamind — Collective Team Memory

Every Slack conversation is indexed with embeddings. Any lobster can search "what did we decide about the pricing model?" and get accurate, sourced answers — across all channels, all history.

🧬 Self-Improving Lobsters

Lobsters detect when they're corrected, record the lesson, and permanently improve. Validated patterns promote fleet-wide — one lobster's mistake becomes every lobster's knowledge.

🎯 GoalOps — Goals, Not Instructions

Give your controller a high-level goal. It decomposes into sub-goals, assigns to lobsters, and they self-coordinate — with P2P handoffs, dependency tracking, and human-in-the-loop approval for sensitive actions.

📋 Task Management

Unified task board with sprint planning, breakdown, and tracking. Backend-agnostic: works with Notion, Linear, GitHub Projects, or plain markdown.

🐣 30-Second Spawning

Tell your controller "spawn a lobster for alice" — it creates the Slack app, handles OAuth, and launches the container. You click Allow and paste one token.

📦 17+ Built-in Skills

Deep research, SEO monitoring, landing page optimization, weekly reports, structured decision alignment, approval workflows, Excalidraw diagrams, data dashboards, and more. See templates/skills/.

🤖 Coding Agent Workflow

Confidence-based routing for ACP coding agents (Claude Code, Codex, Cursor). High confidence → auto-execute. Medium → structured prompt + review. Low → ask human first. Every change goes through automated tests + manual review before shipping.

🔍 Proactive Task Discovery

Lobsters don't just wait for assignments — a daily cron triggers them to scan task boards for unassigned work, score items by capability match, and propose actionable items to their owner. Self-driving by default.

🔄 Fleet Operations

Rolling upgrades with canary + auto-rollback. Config broadcasting. Health checks. Credential management. Gateway watchdogs. 24 operational scripts — all battle-tested.

Get Started

Prerequisites: OpenClaw on your host machine + Docker

git clone https://github.com/CorellisOrg/corellis.git
cd corellis
cp .env.example .env           # add your LLM API key
docker compose up -d           # launch your first lobster

Note: Each lobster needs a Slack bot app. Run ./scripts/create-slack-app.sh <name> to create one automatically, or see Slack Bot Setup for manual setup.

→ Full walkthrough: Tutorial: Set Up a 3-Person Team in 30 Minutes

Production Evidence

This isn't a weekend project. Corellis has been running continuously since February 2026:

The 24 operational scripts and Teamind modules were built iteratively from real production needs — not designed in a vacuum.

Architecture

┌──────────────────────────────────────────────────┐
│  Host Machine                                     │
│                                                   │
│  🎛️ Controller (OpenClaw on host)                 │
│  ├── Manages Docker containers                    │
│  ├── Writes company-skills/, company-memory/      │
│  └── Runs spawn, broadcast, sync, health-check    │
│                                                   │
│  ┌──────────┐ ┌──────────┐ ┌──────────┐         │
│  │ 🦞 alice │ │ 🦞 bob   │ │ 🦞 carol │         │
│  │ (Docker) │ │ (Docker) │ │ (Docker) │         │
│  └────┬─────┘ └────┬─────┘ └────┬─────┘         │
│       │             │             │               │
│  ┌────┴─────────────┴─────────────┴────┐         │
│  │  Shared Volumes (bind mount, ro)    │         │
│  │  company-memory/ company-skills/    │         │
│  │  company-config/ shared-knowledge   │         │
│  └─────────────────────────────────────┘         │
│                                                   │
│  Each lobster also has private rw storage:        │
│  configs/<name>/workspace/ → ~/.openclaw/         │
└──────────────────────────────────────────────────┘

Developer Reference

Looking for the full technical details? Everything from the README is documented in depth:

Alternatives

Requirements

• OpenClaw on the host machine
• Docker + Docker Compose
• Slack workspace with bot apps (one per lobster)
• 2GB RAM per lobster (3GB with Claude Code)

Links

• 🌐 corellis.ai
• 📖 OpenClaw Documentation
• 💬 OpenClaw Community
• 🐛 Issues & Feature Requests

License

MIT