dsh-math-input

DeepSeek Harness's composer is a plain-text field. When you need the model to reason about a mathematical formula, the only option is to type LaTeX character by character from memory—slow, error-prone, and jarring. dsh-math-input fixes exactly that: it adds a plus button next to DSH's input row, turning formula entry into a visual, intuitive workflow so you can stop fighting backslashes.

The plugin offers three input methods. You can handwrite a formula on a canvas using a mouse, touch, or stylus, and after a brief pause the CoMER neural network automatically recognises it into LaTeX source. You can paste or upload a screenshot of a formula (PDF captures, lecture slides, etc.) and the engine extracts the LaTeX from the image. Or you can open the built-in LaTeX syntax editor, where Greek-letter buttons and templates for fractions, radicals, summations, and more let you build expressions quickly while a live KaTeX preview renders on the right. All recognition runs locally in the browser via ONNX Runtime Web, backed by WASM or WebGPU. The model is just 7.2 MB and is downloaded once then cached in IndexedDB. The plugin never calls ctx.llm, so token cost is strictly zero and everything works fully offline.

If you regularly submit formula-heavy questions in math, physics, engineering, or statistics courses, or you are a researcher or educator who keeps feeding equations from lecture notes into the model, this plugin turns your workflow from typing LaTeX into sketching or dropping in an image—faster, more accurate, and cost-free.