The real llama.cpp server, run from a browser tab.
Pick a model, see if it fits your GPU, load it, chat. Every llama-server flag is still there when you want it.

irm https://raw.githubusercontent.com/dadwritestech/LlamaForge/master/install.ps1 | iex   # Windows, no admin

curl -fsSL https://raw.githubusercontent.com/dadwritestech/LlamaForge/master/install.sh | sh   # Linux / macOS

Then Install llama.cpp (the official build for your GPU, no compiler) → Discover → Chat. Early preview: Windows + NVIDIA is the most-tested path. Linux and macOS pass CI but have had little real-hardware use.

LlamaForge runs no models itself. It installs and drives llama.cpp's own llama-server router and edits its models.ini for you. Not affiliated with ggml-org. If you'd rather have polish than control, use LM Studio, Ollama or Jan.

Why LlamaForge

New model architectures land in llama.cpp first. Desktop apps pass them on when they next update their bundled engine. LlamaForge runs the official llama.cpp release itself (or your own build or fork), so a new model is one Update click away, and it puts a UI over every server flag instead of a curated subset.

_{Discover opens on what llama.cpp just learned to run, checked against the engine you have.}

_{As of October 2026, to the best of our knowledge. Spot something wrong? A PR to fix this table is very welcome.}

What's in it

• Models: every model on your machine in one list, with live VRAM/util/temp per GPU. Load, unload, or open Chat from the row.
  • Expand a model to edit every llama-server flag, grouped and searchable, next to a GGUF metadata card (architecture, quant, trained context, layers).
  • Save reloads the model in place. A failed load shows the last error from the router log with a best-guess hint.
  • A loading model shows real progress (weights, draft, vision), straight from llama.cpp's load callback. One-click switches for Thinking (Auto / On / Off plus a token budget, with a note on whether the model's chat template can really turn it off) and Vision (off frees the projector's VRAM; the file stays on disk).
  • Presets, launch profiles (model + preset + pinned llama.cpp build in one click), side-by-side compare, and copy-paste client snippets.
  • Game mode unloads your models when a game starts and loads them again when it closes; pick any to keep loaded.
  • Elastic VRAM (on master, coming in the next release; NVIDIA): when another app needs the GPU, a loaded model moves layers to the CPU instead of unloading, keeps its conversation, and moves back once the memory is free. Slower while it's shrunk.
  • Turn on Multi-model in Setup to keep a main model and workers loaded at once. A planner places each one on the GPUs that fit it, using footprints measured on your machine. A model can also be pinned to its own llama.cpp (or ik_llama.cpp) build.
• Chat: llama.cpp's own chat UI (markdown, reasoning, images) inside the dashboard, on its own port, with the API key added for you.
• Voice: text to speech on your GPU through llama.cpp's own llama-tts: Qwen3-TTS in 10 languages, or the tiny English Pocket TTS that runs faster than real time even on a CPU, optionally in a voice you record or upload (only clone voices you may use).
• Embers: small local agents that keep watch on a topic, keep their own wiki and write you a brief on a schedule. A wiki item only counts if it quotes its source verbatim. Forge builds one by interviewing you; Model Scout needs no setup. Embers have no browser or tools, and nothing leaves the machine unless you turn on push notifications.
• Discover: Hugging Face GGUF search that opens on what's new this week. Every quant gets a rough fit rating for your VRAM before you download (FITS / TIGHT / CPU OFFLOAD). Downloads resume after interruption, register themselves, and end in Load & Chat.
• Will it run?: pick a repo and quant, get the fit and a rough speed estimate.
• Build / Update: one-click official llama.cpp builds with rollback, or build from source with flags detected for your GPU. Also drives ik_llama.cpp and, on Windows, vLLM in WSL2.
• Stats: per-model tokens, speed and run counts from the router's own metrics. Per-client stats aren't possible, because clients talk to the router directly.
• Context: Markdown context docs composed into profiles, injected into requests or written into CLAUDE.md / AGENTS.md.
• Recipes: share a profile as readable JSON; others import it in one paste and LlamaForge downloads the model if missing. There's a community gallery.

A first-run wizard and a Lite / Advanced toggle keep the deep knobs out of the way until you want them. The default look, Stowage, draws each GPU as a bay plan ruled in 1 GiB cells; Hearth and Classic are one click away, and each comes in light, dark and colorblind-safe.

Use it from other apps

Anything that speaks the OpenAI API (Open WebUI, SillyTavern, Continue, Cline, Aider, the OpenAI SDKs) works against http://127.0.0.1:8080/v1. The router always runs with an API key, and Client Config on any model gives you the base URL, key and model id ready to paste.

• Anthropic-compatible POST /v1/messages on the panel, with streaming and tool use.
• OpenAI speech-compatible POST /v1/audio/speech on the panel (WAV or PCM), backed by llama-tts.
• Connect an agent writes the config for Claude Code, Codex and pi.dev (any file it touches is backed up first).
• Load/unload endpoints let an agent swap models on demand.
• MCP server (stdio, backend/mcp_server.py): Claude Code, Codex or any MCP client can see what's loaded, load and unload models, check what fits, pull GGUFs from Hugging Face, and hand a whole task to pi running on a loaded local model (pi_run). One-line setup under Setup -> MCP server, e.g. claude mcp add --scope user llamaforge -- python <LlamaForge>/backend/mcp_server.py.

Screenshots

Models: each GPU drawn to scale, with the open model booked in	Stats: memory and live speed per model

Discover: new in llama.cpp, then trending GGUFs	Build / Update

Install

The one-liners above need no git, admin or compiler. Re-run them any time to update.

The installer finds Python 3.10+ (on Windows it drops a private, SHA-256-pinned copy of python.org's embeddable Python if you have none), downloads the latest release, adds a Start menu / app-menu entry (~/Applications/LlamaForge.app on macOS, plus a llamaforge command on Linux/macOS) and opens the dashboard. Updating keeps your settings, models and engines. Uninstall from Apps & Features on Windows, or llamaforge uninstall; it asks before touching your settings or models.

git clone https://github.com/dadwritestech/LlamaForge
cd LlamaForge
powershell -ExecutionPolicy Bypass -File bootstrap.ps1   # Windows
./bootstrap.sh                                           # Linux / macOS

Daily use: open LlamaForge from the Start menu / app menu, or run llamaforge. It starts the router and the dashboard and opens your browser.

• Dashboard: http://127.0.0.1:8090
• Chat: http://127.0.0.1:8091 (also the dashboard's Chat tab)
• API for your other apps: http://127.0.0.1:8080/v1

llamaforge stop (or stop.ps1 / stop.sh) shuts down the dashboard, the router and the models it spawned. It stops only processes LlamaForge started; any other llama-server you run is left alone.

Requirements: Windows 10/11, Linux, or macOS on Apple Silicon. Python 3.10+ (stdlib only; the Windows installer brings its own). NVIDIA (CUDA), AMD/Intel (Vulkan) or Apple (Metal) GPUs run models when present; CPU-only works. The VRAM preview, fit ratings and GPU stats read NVIDIA and Apple memory today; AMD/Intel cards don't show VRAM in the panel yet. Building from source also needs Git, CMake, Ninja and a C++ compiler, which the Setup tab can install where a package manager allows.

How it works

LlamaForge contains no llama.cpp code. A pure-stdlib Python backend drives llama.cpp's own router API, edits models.ini, and fetches official builds (or shells out to git / cmake). The knob list is parsed live from llama-server --help, so it tracks whatever build you run.

LlamaForge doesn't pin context size, GPU layers or the multi-GPU split. llama.cpp's --fit (on by default) sizes all three to your free VRAM at load, and moves MoE experts to CPU when needed. Pinning any of them turns fit off, so they stay unset unless you set them yourself.

Security: the dashboard only listens on 127.0.0.1. The router is keyed even when local, so a web page you visit can't drive it. LAN access is opt-in from Setup and always needs a key. Details and how to report a vulnerability privately: SECURITY.md.

Docs

Everything is in the in-app Help tab and at dadwritestech.github.io/LlamaForge: configuration, keyboard shortcuts, themes & colorblind-safe mode, vLLM, troubleshooting, and what's new. ROADMAP.md has what's shipped and planned; it's an early preview, so priorities follow feedback.

Credits & license

LlamaForge is MIT-licensed (LICENSE). It builds and drives llama.cpp - MIT, (c) The ggml authors - see NOTICE and LICENSE.llama.cpp.txt. The hard part is theirs; please star and support the upstream project.

pi_run drives pi, Mario Zechner's open-source coding agent (MIT). LlamaForge doesn't ship it: Setup → Install pi fetches the published npm package into LlamaForge's own agents/ folder using your Node.js (or install it yourself with npm install -g @earendil-works/pi-coding-agent).