A local model

• AVAILABLEConnecting a local endpointSame tools and same sandbox as any agent, on the loopback only.
• PARTIALChoosing its tools wellNot established for the model tested here: see what was measured.
• PLANNEDPinning a local model from the appNo path in the app pins one on a card; the Terminal does, with --model. Inferred from reading the code, not checked by a run.

#Set it up

1. Install an OpenAI-compatible server on your machine, Ollama, LM Studio or vLLM, listening on http://127.0.0.1:11434/v1.
2. Pull a model. The two exercised here are `qwen3.5:9b-q4_K_M` for generation and `qwen3-embedding:0.6b` for embeddings.
3. Check the endpoint, list what it serves, and ask the bare model, no role, no tool.
4. Run a role on it. --model is required; --offline keeps only the tools that reach no host.

deskworth doctor
deskworth models list
deskworth models ask "What is a 10-K?" --model qwen3.5:9b-q4_K_M
deskworth agents test fundamental --model qwen3.5:9b-q4_K_M --offline

#How it works

• The loopback only: any address other than 127.0.0.1, localhost or `::1` is refused, and no redirection is followed.
• The same tools, served by the local tool server with the card’s token, and the same sandbox as any agent.
• Bounds: 12 turns, 1,800 seconds, 16,384 tokens per turn, reasoning off by default. Without a pinned model, the card says so instead of guessing one.
• No training: Deskworth neither trains nor adapts any model.

#What was measured

• Livia, the fundamental role, on `qwen3.5:9b-q4_K_M`: 4 turns, 214 seconds, Apple’s real segments read; 127.0.0.1 alone across four passes, sampled every 6 seconds.
• A local desk on 2026-09-23: 18 tools served; fundamental read prices instead of filings, execution called no tool in 190 seconds, and macro stopped on a missing FRED key.
• Memory observed on one Mac with 32 GB: 6.6 GB for generation, 5.8 GB for embeddings, 12.4 GB for both, and the system stopped that pass. One instance at a time. These are observations, not a minimum requirement.

#Without a network