Version 0.1.0 (unreleased prototype). See versioning and the changelog.
wm manages Docker workloads across hosts from one declarative configuration. You state what each workload should be: running or stopped, which image or model it serves, what "ready" means and how much recovery is allowed. wm plans the difference, brings the workload there and confirms it works before reporting success. One CLI shows status for every workload and runs plan, apply, start, stop and restart.
This is an early prototype: it adopts existing containers rather than creating new ones. Provisioning, image or model replacement and multi-host orchestration are design goals.
My containers sprawl across Mac and Linux machines: apps on one host, model servers on GPU hosts over SSH, each started by hand or a one-off Compose file. Knowing what should run, what does run and whether it works meant logging into each host. wm gives each workload one declared intent, one status view and one reviewed way to change it, without replacing the Docker setups that already work.
OpenTofu holds the declaration, plan and state. During apply, a bounded agent works out how to reach the approved intent using a few scoped tools (observe, start, stop, wait, logs, probe, restart). Deterministic code checks every action against identity, policy and budgets, and only an independent readiness probe can mark the work done. Plans never call the model.
flowchart LR
subgraph CTL["Controller (Mac or Linux)"]
CLI["wm CLI + OpenTofu"] --> PR["wm provider"]
PR --> AG["Agent loop"]
AG --> BR["Broker<br/>identity, policy, budgets"]
end
BR -->|local| LD["Docker daemon<br/>local apps"]
BR -->|SSH| GPU["Linux GPU host<br/>Docker: model servers"]
BR -.->|SSH| LX["Linux host<br/>Docker: apps"]
BR -.->|local or SSH| LA["macOS host<br/>launchd services"]
BR -.->|local or SSH| SD["Linux host<br/>systemd services"]
classDef planned stroke-dasharray: 5 5,opacity:0.7
class LX,LA,SD planned
Solid lines are implemented; dashed lines are planned. See declarative intent, agentic reconciliation for the design, safeguards, tradeoffs and alternatives considered.
Requires Go (see go.mod), Node >=22.19, npm, Python 3, OpenTofu and SSH, plus Docker CLI access on each workload host.
npm ci
make build BIN_DIR=.local/release-check
npm run typecheck && npm test && make test-go
python3 -m unittest discover -s agent/remote -p 'test_*.py'
python3 -m unittest discover -s scripts -p 'test_*.py'Plain make build installs into .local/bin/, replacing the binaries workload
launchers use; make build-mac targets Apple Silicon. macOS and Linux are
supported; Windows is not. See build targets.
Copy examples/workloads to a directory outside this checkout and replace its synthetic identities with the exact containers you want to adopt. Then, from that directory:
./wm
./wm plan models/d3
./wm apply models/d3 -vApply does not ask for confirmation, so review the saved plan first. Destroy stops an adopted container and keeps the container and its data.
- Design: reconciliation approach, safeguards, tradeoffs and alternatives
- Configuration: HCL, companion bindings and agent model profiles
- Operations: status, STATE SYNC, plan/apply and lifecycle commands
- Development: code layout, build, tests and versioning
- Proposals: target design beyond the prototype
- Contributor instructions