Build persistent AI agents that run 24/7. Local LLM for routine tasks, cloud API for complex ones. Monitor your projects, track deadlines, write code, deliver standup summaries — all from a Docker container.

Omniclaw is an open-source framework for building persistent AI agents — ones that don't disappear when you close a tab. Your agent runs in Docker, remembers what happened yesterday, checks your Jira boards on a schedule, and messages you on Slack when something needs attention.

It's built on OpenClaw, an open-source AI agent daemon that handles the hard parts: sessions, channels, model routing, and tool execution.

Who it's for: Developers who want an AI assistant that works autonomously — checking boards, monitoring repos, tracking deadlines — without babysitting it.

# Clone the repo
git clone https://github.com/engindearing-projects/omniclaw.git
cd omniclaw

# Run the setup wizard (needs Node.js)
node scripts/wizard.mjs

# Start everything
docker compose up -d

The wizard walks you through configuring your agent name, API keys, and integrations. After setup, your agent is live at http://localhost:18789.

No Node.js? Run ./scripts/setup.sh instead — it copies example configs for you to edit manually.

• Local-first AI — Ollama (llama3.1/3.2) handles routine tasks for free. Claude Sonnet kicks in for complex work.
• Jira tracking — Monitor boards, surface overdue tickets, track critical paths
• GitHub monitoring — Watch PRs, CI status, merge conflicts across repos
• Deadline alerts — Proactive warnings when timelines slip
• Daily standups — Morning briefs delivered to Slack with overnight activity + priorities
• Autonomous code work — Create branches, write code, run tests, open PRs (with guardrails)
• Cron scheduling — Run checks on exact schedules (8am standup, 2pm follow-up)
• Heartbeat system — Periodic background checks every 30 minutes
• Multi-channel — Slack, Telegram, CLI
• Memory system — Daily logs + long-term memory that persists across sessions
• MCP bridge — Expose your agent as MCP tools for use in other AI workflows

Omniclaw runs as two Docker containers:

┌─────────────────────────────────┐
│  omniclaw container             │
│  ┌───────────────────────────┐  │
│  │  OpenClaw daemon          │  │
│  │  ├─ Gateway (port 18789)  │  │
│  │  ├─ Agents & Sessions     │  │
│  │  ├─ Cron scheduler        │  │
│  │  └─ Channel connectors    │  │
│  └───────────────────────────┘  │
│  Volumes:                       │
│    config/ → system config      │
│    workspace/ → agent personality│
│    cron/ → scheduled jobs       │
│    memory/ → persistent memory  │
└─────────────────────────────────┘
         │
         ▼
┌─────────────────────────────────┐
│  ollama container               │
│  llama3.1 / llama3.2           │
│  (port 11434)                  │
└─────────────────────────────────┘

See ARCHITECTURE.md for the full system design.

Skills are modular capabilities defined in workspace/skills/. Each skill has a SKILL.md that documents its workflow, API usage, and guardrails.

1. Create workspace/skills/your-skill/SKILL.md
2. Add YAML frontmatter with name, description, and requirements
3. Document the workflow, commands, and response format
4. Enable it in config/openclaw.json under skills.entries

See existing skills for the pattern.

Omniclaw uses a layered personality system:

The .example files are templates. The setup wizard generates real configs from them.

The MCP bridge lets you use your Omniclaw agent as a set of MCP tools from other AI workflows (like Claude Code). It connects via WebSocket to the OpenClaw gateway.

# Install bridge dependencies
cd mcp-bridge && npm install

# Add to your MCP config (.mcp.json)
{
  "mcpServers": {
    "omniclaw": {
      "command": "node",
      "args": ["./mcp-bridge/index.mjs"]
    }
  }
}

# From your local machine
./scripts/deploy-to-ubuntu.sh user@your-server

# On the server
cd ~/omniclaw
./scripts/setup.sh

For GPU-accelerated Ollama on Ubuntu, uncomment the deploy section in docker-compose.yml.

See CONTRIBUTING.md for development setup, skill creation guide, and PR process.

MIT — see LICENSE for details.