Open Source Tool That Saved me from Paying Huge LLM API Bills

Description and use

NoRefund is a local desktop app that accurately tokenizes documents with real model tokenizers, compares costs across many LLMs, and estimates VRAM fit for self‑hosted models, keeping all data on the user’s machine.

Problem

Existing online token counters fail for large documents, support few models, give inaccurate counts, and expose sensitive data to external servers.

Outcome

NoRefund enabled accurate, local token counting, cost comparison across many models, and VRAM fit estimation, eliminating the risk of unexpected bills and data exposure.

Workflow

• 1. Install NoRefund from the GitHub releases page (no Python required).
• 2. Run the app; no admin permissions needed on most systems.
• 3. Drag and drop files or folders into the Parser tab to see token counts instantly.
• 4. Open the Compare tab to test multiple models side by side and view cost comparisons.
• 5. Use the Fit Check tab to estimate if an open‑weight model’s weights and KV cache fit your GPU’s VRAM.
• 6. Browse the Registry to view supported models, pricing, and specs.
• 7. Manage downloaded tokenizers in the Resources tab; the app works fully offline after initial downloads.

Evidence

NoRefund counts tokens in local files against real tokenizers. No documents ever leave your machine.

Source