OpenKrew

AI agents on separate machines, working as a team.

OpenKrew runs named AI agents on different physical machines and lets them work together on projects. Each agent has its own identity, role, and permissions. They coordinate through Slack and a central Hub API.

Most multi-agent frameworks run everything in one process. OpenKrew doesn't. Your agents are actually distributed, each on its own box, talking to each other over real channels. Think less "threads sharing memory" and more "remote coworkers in a Slack workspace."

Project Vision

How it works

Primary Machine (Hub) -- Always On
├── Agent Registry API (:4000)
├── Shared Memory DB (PostgreSQL + pgvector)
├── Task Queue (Redis / BullMQ)
├── Provisioning Service (SSH)
└── Agent: "Quinn" -- Dev Lead

Spoke 1 (e.g., MacBook Mini)
└── Agent: "Morgan" -- Frontend Developer

Spoke 2 (e.g., Linux Server)
└── Agent: "Reese" -- Infrastructure / DevOps

The Hub runs the registry, shared memory, and task queue. Spokes are remote machines, each running one agent. Agents talk through Slack for conversation and the Hub API for structured handoffs.

What's in the box

Agents run on separate machines, not threads. Each one has a name, role, and personality defined in YAML. They share a three tier memory system: personal (local files), team (PostgreSQL + pgvector), and project scoped. You set per-agent permissions for shell access, filesystem, git, packages, and secrets.

Slack is the communication layer. Works with Claude, GPT, DeepSeek, Ollama, LM Studio, and others. SSH provisioning handles remote setup. A privilege broker lets agents request elevated access without holding root credentials directly.

Agent config

Each agent gets a YAML file:

agent:
  name: Quinn
  role: dev-lead
  machine: mac-mini-lan
  model: claude-opus-4-6
  teammates:
    - name: Morgan
      role: frontend
      channel: "#frontend"
    - name: Reese
      role: infra
      channel: "#infrastructure"
  tools: [git, gh, dotnet-cli, docker]
  mcps: [context7, sequential-thinking]

Permissions

You control what each agent can do. Morgan here can push code and open PRs but can't merge, can't sudo, and can't touch anything outside the UI and docs folders:

agent: Morgan
role: frontend

shell:
  enabled: true
  sudo: false
  blocked_commands: [rm -rf /, shutdown, reboot]

filesystem:
  mode: scoped
  allowed_paths:
    - ~/projects/my-app/UI/**
    - ~/projects/my-app/docs/**

git:
  push: true
  create_pr: true
  merge_pr: false # Only lead can merge
  force_push: false

packages:
  install: true
  managers: [npm, npx]

secrets:
  access: false

Memory

Tech stack

Building from source

Needs Node 22.16+ (24 recommended).

git clone https://github.com/OpenKrew/openkrew.git
cd openkrew

pnpm install
pnpm build

Roadmap

We're starting with the basics: get two agents talking on one machine, nail the communication and memory protocols. After that, SSH provisioning for actual multi machine setups, then the interactive setup wizard, then the permission system. Agent to agent coordination (handoffs, conflict resolution) comes last. Then we open source it properly.

Origin

OpenKrew started as a fork of OpenClaw (MIT). We kept the Gateway, Pi agent framework, LLM adapters, and Slack integration. Everything else -- multi machine distribution, shared memory, role based permissions, team coordination -- is new.

See THIRD_PARTY_LICENSES.md for attribution.

License

Business Source License 1.1 -- free for non-commercial and internal use. Converts to Apache 2.0 after 3 years per version. Commercial licensing available at licensing@openkrew.com.