Local Agent models (free)
JobJitsu’s Agent runs on your machine via Ollama. Résumé text and job details stay local — Ollama is contacted only at 127.0.0.1.
1. Install Ollama (free)
| OS | Where |
|---|---|
| macOS / Windows / Linux | Download the app from https://ollama.com/download |
After install, confirm in a terminal:
ollama --version
Keep the Ollama app running while you use JobJitsu.
2. Pull a free model
These are free to download from Ollama’s library (no API key). Pick by RAM:
| Model (save this name in Preferences) | Approx. size | Good when |
|---|---|---|
qwen2.5:3b (recommended start) | ~2 GB | 8 GB+ RAM — solid résumé/cover-letter drafts |
llama3.2:3b | ~2 GB | 8 GB+ RAM — general local craft |
phi3:mini | ~2.3 GB | 8 GB+ RAM — compact Microsoft Phi |
qwen2.5:7b | ~4.7 GB | 16 GB+ RAM — stronger writing quality |
llama3.1:8b | ~4.7 GB | 16 GB+ RAM — stronger general quality |
Pull one (example — recommended):
ollama pull qwen2.5:3b
List what you have:
ollama list
Smoke-test in the terminal (optional):
ollama run qwen2.5:3b "Write one calm résumé summary sentence for a staff engineer."
3. Point JobJitsu at the model
- Open Preferences
- Under On-device Agent model, choose an installed model from the list (Refresh if you just pulled one)
- Save — chrome should move toward Agent · On-device when Ollama is up
4. Privacy notes
- JobJitsu talks to Ollama only on localhost (
127.0.0.1:11434) - No JobJitsu cloud holds your résumé
- Agent prepares drafts; it does not send applications for you
- Optional remote providers are Experimental and must stay opt-in (not this guide)
Troubleshooting
| Symptom | What to try |
|---|---|
| Agent · Unavailable | Start the Ollama app; run ollama list |
| Model not found | ollama pull <name>, then Refresh list and choose it in Preferences |
| Slow first reply | First load pulls weights into memory — wait, then retry |
| Want better quality | Pull qwen2.5:7b or llama3.1:8b if you have RAM |
Craft chat (PE28-S03)
On Craft, after drafts exist, use Refine with Agent for a focused edit. If the résumé/JD is thin or the request asks to invent experience, Agent asks clarifying questions instead of fabricating facts. Nothing is sent from Craft.
Related
- Epic: PE28 Local Craft Studio (#101)
- Story: PE05-S06 real on-device Agent (#102)
- Story: PE05-S07 list local Ollama models (#112)
- Story: PE28-S03 craft chat refine (#105)
- Architecture: AI_ARCHITECTURE.md
