SuperAPI AI Directory
Partner with us中文AI newsFavorites
All toolsAI developer platformscorellis

corellis

Corellis is an open-source system that extends a single OpenClaw AI assistant into a collaborative AI team (whose members are called "lobster"), used for task orchestration, shared knowledge, and collective memory. Each member runs in an independent Docker container, with private memory and sessions, while sharing the company knowledge base and searchable team memory. Core capabilities include: Teamind collective memory (builds an embedding index of Slack conversations, enabling retrieval of the basis for decisions across channels and history); a self-improvement mechanism (detects corrected content, records lessons, and propagates them to the whole team); GoalOps goal-driven operation (the controller breaks high-level goals into subgoals, assigns them, and lets members self-coordinate, including dependency tracking and human approval); a unified task board (supports Notion, Linear, GitHub Projects, or plain Markdown); generate a new member in 30 seconds; 17+ built-in skills; a confidence-routing workflow for coding agents; proactive task discovery; and fleet operations (rolling upgrades, configuration broadcast, health checks, and 24 other scripts). Prerequisites are that the host has OpenClaw installed and Docker available, and each member needs a Slack bot app. The repository claims continuous production operation since February 2026, with 28 members, 50,000+ indexed Slack messages, and 500+ self-corrections.

Open website

About this tool

Purpose Corellis aims to extend a single OpenClaw AI assistant into a collaborative AI team, allowing multiple AI members (commonly called lobsters) to share knowledge, collective memory, and distributed goals, thereby taking on operations, marketing, launches, weekly reports, and other work.

Main capabilities

  • Teamind collective team memory: every Slack conversation is embedded and indexed, so any member can retrieve questions like "what decision did we make about the pricing model" and give accurate, sourced answers covering all channels and history.
  • Self-improving members: when a member detects a correction, it records the lesson and improves permanently; validated patterns are propagated to the entire team, so one member's mistake can become knowledge for all members.
  • GoalOps goal-driven operation: give the controller a high-level goal, and it breaks it into subgoals and assigns them to members; members coordinate among themselves, with support for peer-to-peer handoffs, dependency tracking, and human-in-the-loop approval for sensitive operations. In the official example, the founder enters a single sentence, "launch a user acquisition campaign for the new product line," and the controller breaks the goal into 6 subgoals, creates milestones and 7 tracking cards, assigns and confirms 6 members, after which spontaneous cross-team coordination emerges, ultimately producing 36 task cards.
  • Task management: a unified task board that supports sprint planning, breakdown, and tracking, backend-agnostic, and can connect to Notion, Linear, GitHub Projects, or plain Markdown.
  • Generate a member in 30 seconds: tell the controller "spawn a lobster for alice," and it creates the Slack app, handles OAuth, and starts the container; the user only needs to click Allow and paste a token.
  • 17+ built-in skills: deep research, SEO monitoring, landing page optimization, weekly reports, structured decision alignment, approval workflows, Excalidraw diagrams, data dashboards, etc., see templates/skills/ for details.
  • Coding agent workflow: confidence-based routing for ACP coding agents (Claude Code, Codex, Cursor): high confidence executes automatically; medium confidence uses structured prompting plus review; low confidence asks a human first. Every change goes through automated testing and human review before release.
  • Proactive task discovery: a daily cron triggers members to scan the task board for unassigned work, score matches by capability, and propose actionable items to the owner, self-driven by default.
  • Fleet operations: supports canary rolling upgrades with automatic rollback, configure broadcasting, health checks, credential management, and gateway watchdog, 24 ops scripts in total.

How to use Prerequisite: OpenClaw is installed on the host, and Docker and Docker Compose are available. A Slack workspace needs a bot app prepared for each member. You can use ./scripts/create-slack-app.sh <name> to create one automatically, or configure it manually according to Slack Bot Setup. The controller runs on the host rather than inside Docker, because it needs to execute docker compose, manage host files, and run fleet scripts; each member runs in an isolated Docker container and has read-only access to shared knowledge. Configuration is placed in the host's company-config/ directory and bind-mounted read-only into each member to achieve consistent behavior across the whole team. Each member has two secret files: secrets.json (read-only, shared API keys injected at generation time) and personal-secrets.json (read-write, each member's private credentials); OpenClaw's SecretRef system ({"$ref": "secrets://KEY"}) keeps secrets out of configuration files.

Applicable scenarios Suitable for teams that already use a single OpenClaw assistant in production and want to scale to multi-person AI team collaboration, especially scenarios requiring shared memory, goal decomposition, task boards, cross-team coordination, and operational governance.

Limitations Only supports environments with OpenClaw installed on the host plus Docker; each member needs a Slack bot app; in terms of resources, each member requires about 2GB of memory (3GB when including Claude Code). The repository states that this capability has been running continuously in production since February 2026 (28 members, indexing 50,000+ Slack messages, 500+ self-corrections, single server). Regarding specific fees, prices, exact performance data, etc., the materials do not provide them.

Related tools

Browse category