Numkit: In-process numerical computing and MATLAB-compatible runtime for LLM agents (WASM & C++)

[hotredman]

Overview

👉 Repo https://github.com/hotredman/numkit

Numkit is an open-source (0BSD) numerical computing engine and MATLAB-compatible runtime written in C++17. It runs either as a native binary (with Highway dynamic SIMD) or in-process via WebAssembly (Node.js and browser environments).

We built it to address a practical bottleneck in LLM agent tool-calling pipelines: running a full Python / Jupyter environment inside Docker containers often introduces noticeable cold-start latency (300–800 ms), substantial memory footprint, and token overhead for basic numerical tasks.

Development with GLM

The project is under active development. An interesting aspect of our workflow is that we originally started developing the codebase using Claude Opus, but have since transitioned to GLM as our primary coding agent. GLM currently drives our day-to-day engineering tasks — including C++ toolbox implementations, cross-platform build pipelines (MSVC on Windows, Ninja/GCC on Linux/WSL), stack-safety architecture, and automated parity audits against MATLAB.

Try It in the Browser (Live Demo)

You can try the engine immediately in the browser without installing anything (runs 100% client-side via WASM):
👉 Web IDE Demo: https://hotredman.github.io/numkit-demo/

The demo includes a built-in Examples explorer with 180+ ready-to-run scripts across all domains (DSP, control systems, audio, linear algebra, ODE solvers, and image processing) — all runnable in one click with interactive visualization.

You can also run custom code:

% Matrix solve
A = [1 2; 3 4];
b = [5; 11];
x = A \ b;
disp(x);

% Signal processing & interactive plotting
t = linspace(0, 1, 500);
sig = sin(2*pi*5*t) + 0.3*randn(size(t));
[b, a] = butter(4, 0.1);
filtered = filtfilt(b, a, sig);

plot(t, sig, 'Color', [0.7 0.7 0.7]);
hold on;
plot(t, filtered, 'b-', 'LineWidth', 2);
title('Butterworth Filtering in NumKit');
legend('Noisy Signal', 'Filtered');
grid on;