clem

Continuously Looping Engineering Machines.

docker-compose for Claude Code agents.

Run a team of Claude Code agents 24/7 on any Linux host. Each agent is a real OS user in a tmux session under systemd. Configure once, walk away.

Get started GitHub

install.sh

# grab the latest release binary
curl -fsSL https://github.com/jahwag/clem/releases/latest/download/clem_linux_amd64 \
  -o /usr/local/bin/clem && chmod +x /usr/local/bin/clem

# provision your first team
mkdir my-team && cd my-team
clem init
clem vault init
sudo clem provision
sudo clem up

Not a framework.

Agents are real Linux users, running under real systemd. No Kubernetes. No in-process orchestration. Just the pieces your OS already has.

OS identity

One agent, one Linux user

Each agent gets its own home dir, git identity, Discord bot, GitHub PAT. Crash boundaries are real.

Coordination

Discord or Slack

Swappable backend via one config knob. Agents coordinate over a channel you already use.

Runtime

claude-code or opencode

Mix Anthropic, Bedrock, Vertex, Ollama, or any OpenAI-compatible backend. One surface.

Secrets

age + sops

Per-agent .env decrypted at provision time. Nothing leaves the host.

Self-healing

systemd + watchdog

Dead or stalled agents restart automatically. Alerts only fire after repeated failures.

Anywhere

Laptop. Pi. VPS.

Runs on anything with systemd. No cloud services required.

How it works

flowchart TB
  subgraph host["Linux host — laptop / server / VPS"]
    direction TB
    subgraph lead["OS user: myteam-lead"]
      direction TB
      tmux1["tmux session"] --> loop1["claude loop"]
    end
    subgraph worker["OS user: myteam-worker"]
      direction TB
      tmux2["tmux session"] --> loop2["claude loop"]
    end
    loop1 <-. "MCP stdio" .-> loop2
    wd["watchdog timer
(every 5 min)"]
    wd -. restarts .-> lead
    wd -. restarts .-> worker
  end
  host ==> chat(["Discord or Slack
#general · #tasks · #lessons"])
  classDef orange fill:#141417,stroke:#f97316,stroke-width:1.5px,color:#e5e7eb
  classDef chat fill:#1a1a1d,stroke:#fb923c,stroke-width:2px,color:#fff
  class tmux1,tmux2,loop1,loop2,wd orange
  class chat chat

Each loop launches claude, injects a prompt via tmux send-keys, waits for the session to exit, sleeps the configured iteration, repeats. Secrets live encrypted in secrets.sops.yaml; clem provision decrypts them into per-agent .env files.

Ship a team.

Free and open source. MIT licensed.

Star on GitHub Join Discord