🦙 Ollama Forge

A lightweight, zero-dependency web UI and CLI for managing Ollama:

• start/stop the Ollama server
• browse, inspect, pull, create, push and remove models
• chat with any installed model to test it
• sign in/out to ollama.com
• install, upgrade, version-check and uninstall the Ollama binary itself

The backend is plain Node.js (standard library only) and the frontend is vanilla HTML/CSS/JS — no frameworks and no runtime dependencies. Everything can also be compiled into a single self-contained executable (Node's SEA), with the UI embedded inside the binary.

Requirements

• Node.js 18 or newer to run from source (uses the built-in global fetch)
• Node.js 24 or newer to compile the single-file binary (uses the Single Executable Application tooling)
• Ollama installed locally (optional — Ollama Forge can install it for you)

Quick start

npm start

Then open http://127.0.0.1:11435.

To listen on a different address/port:

FORGE_HOST=0.0.0.0 FORGE_PORT=8080 npm start

CLI

The CLI mirrors the web UI's Manage tab. Run it via the unified entry (main.js), the dedicated cli.js, the npm script, or the built binary:

node main.js install    # Install Ollama
node main.js upgrade    # Upgrade Ollama to the latest release
node main.js update     # Check current vs. latest version
node main.js uninstall  # Remove Ollama from the system

node cli.js update       # equivalent, CLI-only entry
npm run cli -- update     # equivalent, via npm
./dist/ollama-forge update   # equivalent, from the compiled binary

With no arguments, main.js (and the compiled binary) starts the web UI instead. An unknown command prints help and exits non-zero.

Single-file binary

Build a self-contained executable with no runtime dependencies — Node, the backend, and the web UI all bundled into one file:

npm install          # installs the two build-time dev deps (esbuild, postject)
npm run build        # -> dist/ollama-forge   (dist/ollama-forge.exe on Windows)

The binary is a drop-in replacement for both entry points:

./dist/ollama-forge          # start the web UI
./dist/ollama-forge update   # run a CLI command

npm run build:bundle produces only the bundled dist/ollama-forge.cjs (useful for inspecting the bundle without injecting a native binary).

How it works

1. build.mjs embeds everything in public/ and bundles main.js (plus lib/) into a single CommonJS file with esbuild.
2. Node's --experimental-sea-config turns that bundle into a SEA blob.
3. postject injects the blob into a copy of the node runtime, producing the final binary.

esbuild and postject are build-time only; the produced binary has no dependencies.

Prebuilt binaries in CI

.github/workflows/build.yml builds and uploads a binary for every supported platform on each push/PR, and attaches them to a GitHub Release when a v* tag is pushed:

Target	Runner
linux-x64	ubuntu-latest
linux-arm64	ubuntu-24.04-arm
macos-arm64	macos-latest
macos-x64	macos-15-intel
windows-x64	windows-latest
windows-arm64	windows-11-arm

Features

Models tab

• Server control — start/stop ollama serve with live status polling.
• Pull — with autocomplete suggestions for popular models; streamed progress.
• Create — build a model from a Modelfile (type it inline or browse for a file).
• Push / Remove — publish to or delete from your local library.
• Details — base model, format, family, quantization, size and digest.
• Auth — sign in (opens your browser) and sign out of ollama.com.

Manage tab

• Binary path — auto-detect, browse, or set the ollama executable path.
• Version — shows installed vs. latest release.
• Install / Upgrade / Uninstall — with streamed output and cancellation.

Chat tab

• Run a model — pick any installed model from the dropdown to test it.
• Streaming replies — tokens arrive live; reasoning/thinking output (for models like Qwen3 and DeepSeek-R1) is shown in a collapsible block.
• Multi-turn memory — the whole conversation is sent back each turn.
• Stop / New Chat — cancel a generation mid-stream or reset the transcript.
• Stats — token count and tokens/sec for each reply.

This tab drives the model through Ollama's /api/chat streaming endpoint — the same engine ollama run uses under the hood — so what you test here matches the CLI, with clean streamed output instead of a terminal REPL.

Configuration

Settings live in ~/.ollama_gui/settings.json:

{
  "ollama_host": "http://localhost:11434",
  "ollama_bin": "/usr/local/bin/ollama"
}

• ollama_host — base URL of the Ollama API. Falls back to the OLLAMA_HOST environment variable, then http://localhost:11434.
• ollama_bin — path to the ollama binary. Auto-detected from PATH (and common install locations) when unset.

HTTP API

All endpoints are JSON. POST /api/run returns a stream of newline-delimited JSON (application/x-ndjson) for long-running tasks.

Method	Endpoint	Description
GET	/api/state	Server status, auth, binary path, host
GET	/api/models	List installed models
GET	/api/show?name=<model>	Model details
GET	/api/manage-version	Installed and latest Ollama versions
GET	/api/detect	Auto-detected ollama binary path
GET	/api/fs?path=<dir>	List a directory (file browser)
GET	/api/read?path=<file>	Read a text file
POST	/api/settings	Persist { "ollama_bin": "..." }
POST	/api/run	Run a task (see below)
POST	/api/cancel	Cancel the active task

Tasks for POST /api/run (body { "task": "..." }):

pull · push · create · chat · remove · serve-start · serve-stop · signin · signout · manage

manage takes an extra "action": update | install | upgrade | uninstall, and an optional "password" for sudo on Linux/macOS.

chat takes "model" and a "messages" array ([{ "role": "user", "content": "…" }]) and streams NDJSON frames: { "stream": "chat", "type": "thinking" | "content" | "stats", … }. Cancel an in-flight chat with POST /api/cancel.

Platform notes

• Linux / macOS — installing, upgrading and uninstalling uses the official install.sh and requires sudo. The web UI prompts for your password when needed; the CLI uses sudo interactively.
• Windows — installs per-user via the official OllamaSetup.exe (no admin rights required) and uninstalls through the registered uninstaller.

Security

The server binds to 127.0.0.1 by default. /api/fs and /api/read expose local filesystem reads to power the in-app file browser, so only change FORGE_HOST to a non-loopback address if you understand and accept that risk.

Project structure

.
├── main.js           # unified entry: web UI by default, CLI with a command
├── server.js         # zero-dependency HTTP backend + static host
├── cli.js            # command-line manager
├── build.mjs         # single-file binary build (bundle + SEA + inject)
├── lib/
│   ├── helpers.js    # config, binary detection, auth parsing, formatting
│   ├── exec.js       # child_process wrappers + streaming
│   ├── ollama.js     # Ollama HTTP API client
│   ├── manage.js     # install/upgrade/uninstall/version logic
│   └── assets.js     # front-end assets (disk or embedded in the binary)
├── public/
│   ├── index.html    # web UI markup
│   ├── style.css     # web UI styles
│   └── app.js        # web UI logic (vanilla JS)
└── .github/
    └── workflows/
        └── build.yml # multi-OS binary build + release

Build output is written to dist/ (git-ignored).

License

MIT