Agent Fleet Manager
Fleet management for AI agents on your local network.
Documentation · Issues · Roadmap · Discord
| What | Supported today | Notes |
|---|---|---|
| Control machine OS | Ubuntu, macOS | Tested end-to-end |
| Target host OS | Ubuntu, macOS | macOS hosts must enable Remote Login first (host setup) |
| Agent runtimes | OpenClaw ✅, Hermes ✅, ZeroClaw ✅ | IronClaw planned |
| Inference providers | Anthropic, OpenAI, OpenRouter, Ollama | Claude subscription is NOT supported — API keys only |
| Messaging channels | Discord, Slack | OpenClaw only |
| Container runtime | None required | No Docker, no Kubernetes — SSH + Ansible |
Clawrium uses Ansible under the hood for SSH-based orchestration. You run clawctl from your control machine, which talks to target hosts over SSH. No agents, no containers, no Kubernetes complexity - just processes running on hosts with a unified management layer.
You're running multiple AI agents - coding assistants, internal tools, experiment harnesses - across machines on your network. Without Clawrium, you SSH into each box, manage configs individually, lose track of token spend, and have no unified view of what's running where.
Clawrium gives you kubectl-style fleet control for AI agents:
- One CLI, all hosts. Add machines to your fleet and deploy any agent type to any host.
- Specialized agents. Each agent does one job and does it well. Instead of one overloaded assistant, run a fleet of purpose-built agents - a coding agent, a review agent, a research agent - each with its own context, data, and configuration isolated from the rest.
- Local inference. Use hardware you already have - Mac Minis, NVIDIA DGX Spark, spare servers - as inference providers. Run smaller open models like Gemma, GPT-4o-mini, Kimi, or Llama locally and point multiple agents at them.
- Model experimentation. Swap models across agents to compare performance without touching individual configs.
- Lifecycle management. Upgrades, rollbacks, secrets rotation, backups - handled.
- Token tracking & guardrails. See spend across your fleet. Set limits before someone's experiment burns through your API budget.
A Clawrium agent is a general-purpose AI assistant that runs on a host in your network. Unlike coding-focused assistants (Copilot, Cursor), these agents are designed for broader tasks. Agent implementations:
- OpenClaw ✅ - Open-source general assistant
- Hermes ✅ (Nous Research) - OpenAI-compatible local API
- ZeroClaw ✅ - Lightweight assistant for resource-constrained hosts
- IronClaw (planned) - High-performance assistant for demanding workloads
Clawrium manages the lifecycle of these agents across your fleet - install, configure, start, stop, upgrade, monitor.
Clawrium is for engineers running AI agents in non-trivial setups - home labs, dev teams, research groups. If you have more than one agent running on more than one machine, this tool exists for you.
It is not a hosted platform. There's no dashboard, no SaaS, no account signup. It's a Python CLI that talks to your machines via Ansible. You own everything.
| Requirement | Why |
|---|---|
| Python 3.10+ | Runtime for clawctl CLI |
| uv | Fast Python package installer |
| SSH access to a remote host | Clawrium manages agents over SSH |
| API key (Anthropic, OpenAI, etc.) | Agents need inference providers |
75-second walkthrough of every tab in clawctl gui — Dashboard, Agents, Topology, Providers, Skills, Integrations, Settings
# Install
uv tool install clawrium
# Or run without installing
uvx --from clawrium clawctl --helpFor full installation instructions including how to install uv, see docs/installation.md.
Step 1 — On your control machine: initialize Clawrium
clawctl service initStep 2 — On your control machine: generate the xclm setup commands for the host
clawctl host create 192.168.1.100 --user xclm --alias worker-1This generates a per-host SSH keypair and, because the xclm user doesn't exist on the host yet, prints the exact setup block (Linux and macOS variants) with your fresh public key inlined. Example output (Linux):
## Linux
# Create xclm user
sudo useradd -m -s /bin/bash xclm
# Passwordless sudo
echo "xclm ALL=(ALL) NOPASSWD:ALL" | sudo tee /etc/sudoers.d/xclm
sudo chmod 440 /etc/sudoers.d/xclm
# Authorized key
sudo mkdir -p /home/xclm/.ssh && sudo chmod 700 /home/xclm/.ssh
echo 'ssh-ed25519 AAAA…' | sudo tee /home/xclm/.ssh/authorized_keys
sudo chmod 600 /home/xclm/.ssh/authorized_keys
sudo chown -R xclm:xclm /home/xclm/.sshStep 3 — SSH into the host and paste the printed block
ssh you@192.168.1.100 # log in as any sudo-capable user
# paste the block from Step 2, exit(Full Linux + macOS variants and rationale: docs/host-preparation.md.)
Step 4 — On your control machine: re-run clawctl host create to register the host
clawctl host create 192.168.1.100 --user xclm --alias worker-1
# → host/worker-1 registered (xclm verified)Step 5 — Register a provider, install + start the agent
# Inference provider
clawctl provider registry create anthropic --type anthropic --api-key-stdin
# Install, configure, and start an OpenClaw agent
clawctl agent create my-assistant --type openclaw --host worker-1 --provider anthropic
clawctl agent configure my-assistant
clawctl agent start my-assistant
# Check fleet status
clawctl agent get
# NAME TYPE HOST PROVIDER STATUS AGE
# my-assistant openclaw worker-1 anthropic running 2m
# Chat with your agent
clawctl agent chat my-assistant
# Or open the local web dashboard
clawctl guiYou should see: A running agent in clawctl agent get output with status running.
→ Full setup guide: ric03uec.github.io/clawrium
| Concept | What it is |
|---|---|
| Host | A machine in your network running one or more agents |
| Agent | An installed AI assistant instance managed by Clawrium |
| Agent Type | The implementation/runtime class of an agent |
| Agent Name | The unique identifier for an installed agent instance |
| Registry | Platform-defined agent types with versions, dependencies, and templates |
Ubuntu and macOS are in the test matrix — both as the control machine and as target hosts. On macOS hosts you must enable Remote Login before registering them; see docs/host-preparation.md. Other Linux distributions may work but aren't tested.
OpenClaw is officially supported and tested end-to-end.
Hermes (Nous Research) is supported in 🚧 In Development status — install, configure, and a local OpenAI-compatible HTTP API on 127.0.0.1:8642 are wired end-to-end. clawctl agent chat and external messaging gateways are not yet supported for hermes. See the Hermes Support Matrix for details.
Additional agent types are planned.
No. Clawrium supports API keys only, by design.
Discord and Slack are supported for OpenClaw.
Additional channels are planned.
No. Clawrium does not require Docker or Kubernetes. It manages agent processes over SSH using Ansible.
Yes. You can register multiple hosts and run different agent types on each host from the same clawctl control node.
I'm building Clawrium in my spare time, so I prioritize my own use cases first.
If you want support for a specific agent type or feature, please open an issue and send a PR. See CONTRIBUTING.md for contribution guidelines.
Two reasons:
-
Most AI agent runtimes don't support it. These run as local processes, not containerized services. They expect a home directory, local config files, and direct access to the host. Wrapping them in containers adds friction with no payoff.
-
K8s is overkill for local fleets. You're managing 3-10 machines on a LAN, not orchestrating microservices across cloud regions. Kubernetes brings etcd, control planes, networking overlays, RBAC, and a learning curve that dwarfs the problem. You don't need a container scheduler - you need to SSH into a box and run a process.
Clawrium uses Ansible under the hood instead. Ansible gives you idempotent host management, secrets handling, and multi-machine orchestration without requiring anything on the target machines beyond SSH. Clawrium sits on top of Ansible and adds the agent-specific layer: lifecycle management, token tracking, model swapping, and fleet-wide visibility.
Python · Typer · ansible-runner · uv
git clone https://github.com/ric03uec/clawrium && cd clawrium
make test # Run tests
make lint # Check style
make format # Auto-formatIssues are the source of truth. See CONTRIBUTING.md for the full workflow.
Changes for the current, unreleased version are tracked in CHANGELOG.md. Every shipped release is archived under docs/releases/<version>/, which is the single place to read the full history of what changed in any release.
Apache 2.0