Guides
Planning that never leaves your machine.
Run the Conductor on LM Studio, MLX, or Ollama. The model that turns your goal into a mission lives on your own hardware, costs nothing per token, and needs no key.
Three ways in
Pick the runtime you already use.
LM Studio
Load a model, start LM Studio's local server, and paste the endpoint into Loom Settings. The server speaks the OpenAI API, which is all the Conductor needs.
MLX
Apple Silicon's native ML stack. Serve an MLX model locally and Loom treats it like any other OpenAI-compatible endpoint, with unified memory carrying the weights.
Ollama
Pull a model, start the server, point Loom at localhost. The shortest path from a bare machine to a working local Conductor.
Setup
Two minutes to a local Conductor.
Start the server, then open Settings, choose LM Studio, MLX, or Ollama, and paste the endpoint. The key field stays empty because there is nothing to protect.
# pull a model and start the server
$ ollama pull qwen3
$ ollama serve
# Conductor endpoint: http://localhost:11434/v1What local actually means here
A local Conductor keeps the planning loop on your hardware. Splitting your goal into tasks, dispatching them, reading terminal output, deciding what counts as done: all of that reasoning streams from localhost. The fleet itself is still six real Claude Code CLI sessions running Claude through your existing login, because Claude is a hosted model. Local planning, hosted hands.
Two tips from running it this way:
- The Conductor's job is judgment, not typing. It splits goals, watches six terminals, answers permission prompts, and verifies work before it counts as done. Favor the strongest reasoning model your hardware can hold over the fastest one, because plan quality decides mission quality.
- Loom ships no telemetry and stores nothing in the cloud, so with a local Conductor your goals and plans simply never leave the machine. If you later want a hosted brain for a harder mission, swapping one in is a Settings change. See Models for the full picture of what runs where.
Hand it the work.
Walk away.
macOS, Linux, and Windows. Around 13 MB. Free and open source.