Cloud agents are great at judgment and tool use — and expensive when you ask them to classify the same kind of short text a hundred times. I wanted a local co-pilot on my Arch box: a small open model that does the boring label/first-draft work over HTTP, so the cloud path only sees what still needs a human (or a heavier agent).
Stack on the laptop: Ollama on loopback (127.0.0.1:11434), models qwen2.5:0.5b then qwen2.5:1.5b. No agent harness yet — raw /api/generate calls. That’s deliberate: prove the API, measure quality, then wrap a script.
What I measured
- Connectivity — tags + generate from localhost in well under a few seconds.
- Constrained labels — with a strict “one token class on line 1” system prompt,
0.5bwas fast and often wrong;1.5bwas good enough to act on for simple triage and usage categorisation. - Free-form TLDR — usable as a draft, not as truth. Local model proposes; I dispose.
Where this is going
Next step is a tiny harness: JSONL in → classify → JSONL out → accuracy against a hand-graded eval set. Until that exists, I treat local output as a filter, not a decision.
Privacy and binding rules (loopback-only, no standing public tunnels, redacted eval data) live in my private notes — not on the public site:
grokbot/jarvis/local-ollama-privacy-and-harness.md (private vault; you’ll need access).