Run Ollama where Docker cannot.

---

Overview

A lightweight Java wrapper that bundles the Ollama runtime directly inside a .jar file. On startup it extracts the bundled binary and libraries, then exposes a reverse proxy with a small admin API for runtime configuration — no Docker, no root, no package manager required.

How It Works

The proxy sits between the client and the Ollama process:

Client → Proxy (server-port) → Ollama (server-port + 1)

All standard Ollama API requests are forwarded transparently via streaming. The two /admin/* endpoints allow changing environment variables and restarting the Ollama process without restarting the JVM.

Dependencies

Library	Version	Bundled
Ollama	v0.16.3	yes (tar inside jar)
Gson	2.13.1	yes (jar inside jar)

No other external libraries. Uses only Java standard library (com.sun.net.httpserver).

Resource Structure

src/main/resources/
  ollama-linux-amd64.tar    # ollama binary + cpu libs, gpu libs removed
  gson-2.13.1.jar

Quick Start

java -jar server.jar

Configuration

server.properties next to the jar:

server-port=11434

Ollama will run internally on server-port + 1. The proxy listens on server-port.

API

Proxy

All standard Ollama endpoints are available at the proxy port.

# List models
curl http://host:11434/api/tags

# Generate
curl http://host:11434/api/generate \
  -d '{"model":"llama3","prompt":"hello"}'

# Chat
curl http://host:11434/api/chat \
  -d '{"model":"llama3","messages":[{"role":"user","content":"hello"}]}'

Admin

PATCH /admin/settings — update Ollama environment variables. Takes effect after restart.

curl -X PATCH http://host:11434/admin/settings \
  -H "Content-Type: application/json" \
  -d '{
    "OLLAMA_NUM_PARALLEL": "4",
    "OLLAMA_CONTEXT_LENGTH": "8192",
    "OLLAMA_KEEP_ALIVE": "10m",
    "OLLAMA_MAX_LOADED_MODELS": "2"
  }'

Response:

{
  "applied": {
    "OLLAMA_NUM_PARALLEL": "4",
    "OLLAMA_CONTEXT_LENGTH": "8192"
  },
  "rejected": {}
}

POST /admin/restart — restart the Ollama process with current settings.

curl -X POST http://host:11434/admin/restart

Response:

{ "status": "restarting" }

Supported Environment Variables

OLLAMA_HOST             OLLAMA_ORIGINS          OLLAMA_REMOTES
OLLAMA_NO_CLOUD         OLLAMA_CONTEXT_LENGTH   OLLAMA_NUM_PARALLEL
OLLAMA_NUM_THREADS      OLLAMA_MAX_QUEUE        OLLAMA_MAX_LOADED_MODELS
OLLAMA_SCHED_SPREAD     OLLAMA_MODELS           OLLAMA_MULTIUSER_CACHE
OLLAMA_NOPRUNE          OLLAMA_NOHISTORY        OLLAMA_LOAD_TIMEOUT
OLLAMA_KEEP_ALIVE       OLLAMA_NEW_ENGINE       OLLAMA_LLM_LIBRARY
OLLAMA_FLASH_ATTENTION  OLLAMA_KV_CACHE_TYPE    OLLAMA_VULKAN
OLLAMA_GPU_OVERHEAD     CUDA_VISIBLE_DEVICES    GPU_DEVICE_ORDINAL
HIP_VISIBLE_DEVICES     HSA_OVERRIDE_GFX_VERSION ROCR_VISIBLE_DEVICES
GGML_VK_VISIBLE_DEVICES HTTPS_PROXY             HTTP_PROXY
NO_PROXY                LD_LIBRARY_PATH         HOME
PATH

OLLAMA_NUM_THREADS defaults to the number of CPU threads available on the host.

Fun Fact

This project uses server.properties for configuration:

— the same filename Minecraft servers use. That is not a coincidence. Most Minecraft game hosting panels (Pterodactyl, Pelican, etc.) are designed around Java .jar files and expose a server-port variable through server.properties out of the box. This means you can deploy Ollama on virtually any Minecraft-compatible hosting plan without touching the panel configuration

— just upload the jar and start it.

Credits

	ollama/ollama