wmo
World Model Optimizer CLI for routing, distilling, and serving models
TLDR
SYNOPSIS
wmo command [options]
DESCRIPTION
wmo (World Model Optimizer) turns collected agent traces into smaller open models and cost-aware routers. It can distill via external APIs (e.g. Tinker), fit routing policies on held-out traces, and serve a local endpoint that keeps frontier quality while cutting inference cost.The project also ships world-model simulation APIs for closed-loop agent testing and optional E2B sandbox backends for harness optimization. Local state lives under `.wmo/`.
PARAMETERS
providers set
Register model providers and candidates in `.wmo/pool.toml`build --file traces --name name
Build a world model / routing setup from OTel or agent tracesoptimize
Subcommands for route sweep/fit/report, distill, harness, and pinserve --name name
Serve an endpoint that routes between frontier and smaller modelslogin
Authenticate against the Experiential Labs hosted platformrun id
Run a hosted world model or agent harnessconfig telemetry {enable|disable|status}
Control anonymous usage telemetryeval
Evaluate harnesses/tasks (optional E2B backend)
CAVEATS
Requires Python packaging (`pip install world-model-optimizer`) and provider API keys for most optimize/serve workflows. Hosted features need `wmo login`. Telemetry is on by default; disable with `wmo config telemetry disable` or `DONOTTRACK=1` / `WMO_TELEMETRY=0`.
HISTORY
Open-source CLI from Experiential Labs for continual improvement of agent models from production traces.
