FAQ

FAQ — the honest part

Does this work with every model?

No, by design. Only models that passed the reading benchmark receive images (fail-closed model gate). Today that's Claude Fable 5 at production density; Opus 4.8 only in an opt-in safe mode at 4× glyph size. GPT-5.5 (0/60 reads) and Gemini 2.5-flash (confabulates) are blocked with receipts. Everything unproven passes through as normal text.

Can the model misread a page and give me a wrong answer?

The gate exists to prevent exactly that. Approved models read at 100% on the production density, and byte-exact values (hashes, IDs, numbers) ride alongside as text rather than being trusted to a visual round-trip. When a model can't read something it abstains instead of guessing — zero silent confabulations across ~300 probes.

Will it slow my requests down?

Rendering is cheap and cached (LRU render cache for deterministic pages). The proxy only converts blocks where the token math wins, so the trade is fewer tokens for a small, local render cost.

Does anything leave my machine?

No more than already would. OmniGlyph is a local proxy; it forwards to the same provider you already use and adds no third party. Responses stream back untouched. The dashboard binds to 127.0.0.1 by default.

What if a model refuses the image?

The refusal sniffer replays the original text request automatically (retryRefusalWithOriginal), so a refusal degrades to normal behavior rather than a failure.

Why not use the high-resolution image tier for even more text per page?

Because it's a measured billing trap: the high-res page is billed at full size, but the vision encoder never receives the extra resolution. Standard pages win on every probe, so that's what ships.

Where do the numbers come from?

Every number has a JSONL receipt in benchmarks/*/results/, produced by the two harnesses (see Benchmarks). Measured decisions are not re-litigated without new data.

Is it really MIT licensed?

Yes. MIT, with the upstream copyright line preserved in LICENSE as required.