Skip to content

nociza/clawie

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 

History

56 Commits
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 

Repository files navigation

clawie

clawie

A local control plane for provisioning, orchestrating, and monitoring a fleet of AI agents from one CLI.

Why clawie

You have multiple agents across providers. Each needs its own config, credentials, channels, and runtime. clawie gives you one place to manage all of it — create agents, delegate tasks between them, isolate their environments, and monitor everything from a terminal dashboard.

Core capabilities

Agent orchestration — Delegate tasks across agents in recursive trees with automatic tier-based routing. Fast agents handle lookups, power agents handle analysis, balanced agents handle everything else.

Multi-provider fleet — Run agents on openclaw, picoclaw, or zeroclaw. Switch providers with a single command. Authorize once, share credentials across agents.

Linux isolation — Each agent gets its own Linux user, home directory, and credential scope. No agent sees another's secrets.

Terminal dashboard — Real-time TUI showing agent status, delegation trees, channels, and health across your entire fleet.

Requirements

  • Linux (Debian/Ubuntu recommended). Uses useradd, systemd, Unix domain sockets, and /tmp — no macOS or Windows support.
  • Python 3.10+
  • No external Python dependencies — stdlib only.
  • Root/sudo required for runtime isolation (runtime create, credentials sync, provider set, auth apply). Agent creation and the dashboard work without root.
  • Provider runtimes (optional): Homebrew for zeroclaw/picoclaw, pnpm or npm for openclaw.
  • Terminal: UTF-8 with color support for the dashboard.

Install

uv tool install clawie        # from PyPI
uv tool install -e .          # from source

Quick start

clawie config set --provider picoclaw --subscription pro
clawie agent create alice --template baseline
clawie dashboard

Agent orchestration

Agents delegate work to each other through a recursive task system with three model tiers:

Tier Budget Use for
fast 4K tokens Status checks, lookups, validation
balanced 16K tokens Most tasks (default)
power 64K tokens Architecture, deep analysis, refactoring
# Delegate with a tier
clawie delegation submit --parent planner --child worker --tier fast --payload '{"task":"check"}'

# Spawn session sub-agents on the fly
clawie delegation spawn-session --parent planner --child researcher --tier power

# See the delegation tree
clawie delegation tree --agent-id planner

Tiers include context budgets that track token usage and trigger compaction warnings to prevent context rot in deep delegation chains.

Key commands

# Agents
clawie agent create alice --model-tier balanced
clawie agent clone alice bob --channel-strategy migrate
clawie agent list
clawie agent show alice

# Delegation
clawie delegation submit --parent p --child c --tier fast --payload '{}'
clawie delegation spawn-session --parent p --child c --tier power
clawie delegation tree --agent-id p
clawie delegation status

# Providers & runtime
clawie config set --provider picoclaw
sudo clawie runtime create alice --user alice
clawie runtime detect

# Dashboard
clawie dashboard

Dashboard

Launch with clawie dashboard. Press v to cycle views:

  • Agents — status, provider, auth, service health per agent
  • Channels — all channels across agents, assign/move with keyboard
  • Delegation — live delegation trees with tier icons, active sockets, task history

Navigate with arrow keys, Enter to drill into an agent, Tab to switch sections, q to quit.

Limitations

  • Linux only — no macOS or Windows. Relies on Linux users, systemd, and Unix sockets.
  • Single machine — all agent communication is over localhost Unix sockets. No network/multi-host delegation.
  • User-level isolation, not container-level — agents get separate Linux users and home directories, but share the same kernel, /tmp, and localhost. No Docker/VM boundary.
  • Delegation depth capped at 10, max 50 children per agent, 5-minute default timeout.
  • SQLite storage — single-writer, not designed for concurrent multi-process access to the same state directory.
  • Token estimation is approximate — uses a chars/4 heuristic, not a real tokenizer.

See docs/requirements.md for full details.

Documentation

Full documentation is in docs/, deployable to GitHub Pages:

Development

uv run clawie --help
uv run --with pytest pytest -q

License

MIT

About

Command line tool to automatically spin up and manage openclaw / zeroclaw / picoclaw agents on linux machines as users.

Topics

Resources

Stars

Watchers

Forks

Releases

No releases published

Packages

 
 
 

Contributors