Ollama (local)

Ollama listens on http://127.0.0.1:11434. The Aarmos daemon proxies loopback traffic and still writes AVAR receipts for every tool invocation — LLM calls themselves are opaque and out-of-scope for policy.

1 · Environment

eval "$(aarmos env)"   # authenticated HTTPS_PROXY + HTTP_PROXY
export OLLAMA_HOST=http://127.0.0.1:11434

2 · Policy

# LLM inference is loopback; policy governs the tools the agent calls.
[[allow]]
verb = "read"
host = "api.github.com"
ceiling.per_minute = 30

[[allow]]
verb = "communicate"
host = "127.0.0.1:11434"          # ollama itself, if you want a receipt trail

3 · Run

aarmos run -- python agent.py

Notes

  • Local model path: no cloud LLM call, no telemetry off-device, receipts still verifiable with `avar verify`.
  • If you skip the 127.0.0.1:11434 allow rule, Ollama calls bypass receipts (they're loopback); explicit allow makes them auditable.
  • Same recipe applies to LM Studio, llamafile, vLLM — swap OLLAMA_HOST.

See all providers → recipes index.