Feature deep-dive

Eight tabs that answer "which model should I run?"

The app is a native Windows program with a dark UI and eight numbered tabs across the top. Here is what each one does.

1 — Welcome

Lands you with the Dojo Badge speed chart and the day's quick actions. Resume an interrupted benchmark. Read what shipped this week. Jump straight to the planner-bridge popup.

Welcome is also where the planner-bridge story starts: paste the architecture plan from your Claude / Gemini / ChatGPT session, hit a button, the app installs every model on the plan and benchmarks each one against your real silicon. The output is a single report you can paste back into the planner.

2 — System Specs

A complete read of what's actually in your computer — not what marketing claims, what the system reports.

CPU model, core count, RAM totals, all GPUs and their VRAM, OS build, free disk, AVX support, NPU presence. The "Re-run scan" button reads the same data the app uses to filter the model catalog, so what you see here is exactly what gates which models you can install.

3 — Efficiency

Eight checkboxes. Every change is reversible. Nothing happens without your consent.

Ultimate Performance power plan · sleep / hibernate disable · Windows Defender exclusion for the models folder · CPU core parking off · GPU power-state lock · background-app shutdown · disk-write-cache verification · paging file sized to RAM. Each row has a "Revert all" button at the top of the tab so you can undo every change in one click.

4 — Models

The full open-weight catalog, filtered to what your machine can actually run.

13 curated models grouped by tier (tiny, small, medium, large, huge). Each row shows VRAM required, estimated tok/s on GPU, model size on disk, category (generalist, coder, reasoner, vision). Greyed-out rows are too big for your hardware. New LLMs added to the catalog every week — no reinstalls, no manual updates.

5 — Actions

Where the work happens. One-click model downloads, benchmark runs, leaderboard submissions.

Tick the models you want, click Download. The app pulls them through Ollama with progress bars per model. Click Run benchmark — five prompts per model (speed, JSON, code, reasoning, long-form). Results land in a SQLite database in your AppData folder. Click Submit to publish your scores to the public leaderboard under a display name you choose.

6 — Guidance

Plain-language recommendations based on what your benchmark just measured.

"For your hardware, the best chat model is X. The best coder is Y. For long-context tasks, only Z holds 32k+ at usable speed." This tab is what most non-technical buyers spend the most time on.

7 — License

Your license key, the machines it's bound to, and a one-click portal to your Stripe billing.

Add or remove a machine from your license. See last validation timestamp. Open the Stripe Customer Portal in your browser to update card or change tier. Revoke a stolen key (we email a magic link to your billing address).

8 — Contact

Four mailto buttons routed to the right address: support, suggestions, billing, security.

Plus an inline form that posts to our API. Auto-bot triages license / bug / billing / feature / other in seconds. Human reply within 48 hours.

First-run wizard

From install to first benchmark in under 8 minutes.

  1. EULA + privacy notice — single screen, single click to accept
  2. License key entry — paste the key from your email
  3. Hardware scan — runs automatically, takes ~10 seconds
  4. Recommended models for your hardware — picked from the leaderboard
  5. Optional Ollama install if not already present
  6. First benchmark on a tiny model so you see results immediately
What it won't do

The app does nothing without an explicit checkbox.