LlamaForge: a browser panel for upstream llama.cpp with local agents that must quote their sources

Sleepybear2611 · reddit · 2026-10-14

The author released LlamaForge, a free MIT-licensed browser control panel for llama.cpp — not a fork, it drives the real upstream llama-server so new models are available same-day. Nearly all code was written by Claude Code (Opus) while the author made decisions and tested on real hardware.

Features

Shows, to scale, where a model will land on each GPU before loading, using sizes measured on your machine.
Runs several models at once; the main one goes on the fastest GPU that fits.
v0.16 adds Embers: small local agents that read sources overnight and write briefs where every claim must quote its source verbatim, verified by plain code against saved snapshots — unverified claims are dropped.
Voice tab powered by locally run Qwen3-TTS.

Claude Code workflow (directly reusable)

Spec first, then task-by-task plans executed by fresh subagents, followed by separate spec-compliance and code-quality reviews that catch real bugs.
Probe llama.cpp behavior with throwaway sandbox scripts before building on assumptions.
Python stdlib backend, vanilla JS frontend, no build step — a deliberate constraint to keep the codebase reasonable for both human and model.
A CLAUDE.md of past mistakes; e.g. the browser pane silently answers "cancel" to confirm(), so the panel uses custom dialogs plus a regression test.
Lesson: verify the test harness before the code — Claude once chased a "bug" caused by its own test script feeding a fake earlier answer.

Install scripts for Windows/Linux/macOS; author's box is NVIDIA 5080 + 5060 Ti.

Original post →

LlamaForge: a browser panel for upstream llama.cpp with local agents that must quote their sources

More from coding & agent