Skip to main content

/local-models can now wire your project, not just recommend

· 2 min read
Carefree Investments LLC
Anchor maintainers

/local-models has always probed your machine and recommended a fit local model. Until now it stopped there — even when Ollama and a fit model were already running, the operator still had to hand-wire endpoints, .env, and Aider config themselves. local_models_autowire.py closes that gap.

Confirmed, not just reachable​

"The API answered" isn't the same as "the model I want is actually being served there." local_models_autowire.py --detect checks both: a guest-reachable OpenAI-compatible API, and that API's /models actually lists the Default-for-wire model id. Either gate failing means not confirmed — no guessing, no writing a stanza for a model that isn't there.

python scripts/local_models_autowire.py --detect
CONFIRMED: base_url=http://172.26.0.1:11434/v1 model=qwen3-8b tag=qwen3:8b served=True aider=True

One confirm, one write set​

When confirmed, /local-models now asks "Wire this project?". A yes writes, all idempotent and machine-local only:

  • a var/endpoints.local.yaml overlay (the memory-safer Default-for-wire model, never the largest fit)
  • the gitignored .env with OLLAMA_API_BASE (reusing write_aider_env.py's existing contract — bare aider stops Connection-refusing)
  • an idempotent conventions-file note, only if the project already has one
  • .aider.conf.yml and a copied launcher script, but only when aider is actually on $PATH
python scripts/local_models_autowire.py --project . --write

Nothing installs Ollama or downloads weights here — this is "use what's already running," not a converge/install pipeline. Declining, or an unconfirmed stack, falls through to the existing draft-offer path unchanged.