Ignite

Local AI Runtime

Running local AI today means installing multiple tools, hunting for the right model, guessing whether it fits your hardware, hand-writing config files, and hoping it all works.

Ignite handles the whole thing. Run the install and start scripts, or use docker compose directly if you prefer to manage the stack yourself, and you get a working local AI setup without hand-writing config files or guessing model settings.

It detects your hardware, recommends models that fit, downloads them, configures everything, and gives you a working endpoint. All from one UI.

If you can install Docker, you can run local AI.

Why "Ignite"? Getting local AI running feels like rubbing sticks together. Ignite is the match.

Platform Support

• Linux: supported
• Windows: not supported yet by the current install/start/stop scripts
• macOS: not supported for the current GPU-backed runtime path

Ignite currently targets Linux with Docker and NVIDIA GPU passthrough.

What You Need

• Docker
• Docker Compose
• NVIDIA GPU
• nvidia-smi working on the host
• NVIDIA Container Toolkit configured so this works:

docker run --rm --gpus all --entrypoint sh ghcr.io/ggml-org/llama.cpp:server-cuda -lc 'nvidia-smi -L'

If that command fails, Ignite will not be able to run GPU-backed models.

The install and management scripts in ./scripts are Linux shell scripts. They are not intended for Windows PowerShell or Docker Desktop workflows yet.

Install

git clone <repo-url>
cd ignite
./scripts/install.sh

What install.sh does:

• checks Docker
• prepares ./config and ./models
• checks Docker GPU passthrough
• tells you the next command to run

Start

./scripts/start.sh

Then open:

• http://127.0.0.1:<IGNITE_PORT> (default: 3000)

What start.sh does:

• checks Docker
• prepares ./config and ./models
• runs docker compose up -d --build

If you prefer to avoid helper scripts, you can start Ignite manually:

docker compose up -d --build

You can also override the published ports through .env:

IGNITE_PORT=3000
LLAMA_SWAP_PORT=8090

After changing ports, restart Ignite.

Stop

./scripts/stop.sh

What stop.sh does:

• runs docker compose down

Update

./scripts/update.sh

What update.sh does:

• pulls the latest Git changes
• pulls the latest runtime images
• rebuilds and restarts the stack

Default Folders

• ./models
• ./config

These are the default paths for normal users.

Advanced users can override them:

SWAPDECK_MODELS_DIR=/path/to/models SWAPDECK_CONFIG_DIR=/path/to/config ./scripts/start.sh

With port overrides:

SWAPDECK_MODELS_DIR=/path/to/models
SWAPDECK_CONFIG_DIR=/path/to/config
IGNITE_PORT=3000
LLAMA_SWAP_PORT=8090

First Run

1. Open Setup
2. Check Docker and GPU status
3. Review the recommended model
4. Open Models and download a GGUF
5. Add it to config with a launch preset
6. Open Test and confirm it responds
7. Use Status to copy the API endpoint for other apps

Pages

Page	Purpose
Setup	Guided first-run flow
Status	Runtime health, logs, Docker GPU preflight, API endpoint
Discover	llmfit recommendations for the current machine
Config	Structured and raw YAML editing for llama-swap
Models	Download, inspect, delete, and add GGUFs to config
Test	Send prompts through the running runtime
Settings	Show runtime settings and Docker-managed paths

Stack

• ignite: React + FastAPI app on :3000
• llama-runtime: llama-swap + llama-server
• llmfit: hardware-aware model recommendations

API

Method	Endpoint	Purpose
GET	/api/status	Runtime state, GPU stats, Docker preflight
GET	/api/discover/recommendations	llmfit recommendation proxy
GET	/api/config	Read runtime config
PUT	/api/config	Save runtime config
GET	/api/config/raw	Read raw YAML
PUT	/api/config/raw	Save raw YAML
POST	/api/config/add-model	Generate a model entry from a GGUF
GET	/api/models	List installed GGUF files
POST	/api/models/download	Start model download
POST	/api/test	Send a prompt through llama-swap
GET	/health	App health check

Security

• Ignite is for local or trusted-network use
• there is no built-in auth layer
• do not expose ports directly to the public internet
• use Tailscale or another private overlay if remote access is needed
• the Ignite app container mounts /var/run/docker.sock so the UI can start and stop the runtime container
• that Docker socket mount gives Ignite control over Docker containers on the host

Third-Party Software

Ignite is built around and depends on the following open-source projects:

• llama-swap — MIT License
• llama.cpp — MIT License
• llmfit — MIT License

Ignite does not change the licenses of those projects. Their respective licenses apply to the components and binaries they provide.

License

MIT