Concrete recipes — each one lists the mode, the model type, and the moves. Every one of them works with the Wi-Fi off.
Load the PDF into the RAG Tool and ask questions like "What are the termination clauses?" or "List every obligation with a deadline." The document is indexed on your disk; the analysis happens on your GPU. Client confidentiality survives contact with AI.
Scanned records, internal dashboards, evidence photos — upload the image in Generate mode with a vision model and get descriptions, transcriptions, or answers. Then click Generate > Chat to keep discussing it conversationally.
Connect the SSH Bot to the box and describe the symptom: "The web app is slow — find out why." It checks load, memory, disk, logs — reading each output and deciding the next command itself — then reports its findings. You watch the terminal live the entire time.
Ask for a PowerShell script to, say, rename a season of media files. Review the generated code in the response — then execute it directly from the app. Python, PowerShell, or Batch, no copy-paste detour through an editor.
Brainstorm and outline with a fast small model — instant responses, zero cost. When the draft is ready, switch the dropdown to your biggest model mid-conversation. It inherits the whole thread and does the final pass. Two-tier quality, one chat, no tokens billed.
Hit the screenshot button, capture the cryptic dialog, and ask. No retyping stack traces, no cropping personal info before uploading to a cloud — because there is no uploading.
Record directly in the app — or drop in an .mp3 — and have an audio-capable model turn rambling into structure: summary, decisions, action items. Recordings of your own meetings finally have somewhere private to go.
Ask for a derivation and get rendered formulas, not ASCII approximations. Combine with file reading — import a problem set as PDF and work through it — for a tutor that's always available and never metered.
Paste invoices, emails, or logs and enforce a JSON response: {"vendor": …, "amount": …, "due": …}. Guaranteed parseable output makes the model a reliable pipeline stage — feed the results to whatever comes next.
With tools enabled, the model can call local Python when it needs real computation — crunch the numbers in your CSV instead of hallucinating them. Function calling, running on the same machine as everything else.
Generate concept images and placeholders locally with x/z-image-turbo or x/flux2-klein. Iterate as many times as you like — the only rate limit is your GPU fan.
Deterministic extraction wants temperature near zero; brainstorming wants it high. Set system prompts and parameters per session and keep presets for the configurations you reuse — the dials the cloud never let you touch.