Apex

Self-hosted ML platform for small AI teams.
Your GPU. Your team. No cloud tax.

---

Apex is a self-hosted ML platform for teams that own their GPU. One pip install, one command, your browser opens to a full job queue, live GPU monitoring, and browser-native VS Code — all running on a single workstation.

What you get

• Job queue with Docker execution — submit training jobs from the UI or API, scheduler picks them up, runs them inside Docker containers with full GPU access, streams logs back over WebSocket
• Real-time GPU telemetry — GPU util, VRAM, temperature, power draw, CPU, RAM, pushed to the dashboard every 2 seconds via server-sent events
• Browser-native VS Code — launch a code-server dev session in a container with your workspace pre-mounted; one click and you're coding inside the GPU
• Job history + logs — sortable table of every job, filter by status, tail logs from anywhere, cancel running jobs
• Pre-built images — apex/code-server:python and apex/code-server:pytorch (CUDA 12.4 + PyTorch 2.4) bundled as working examples
• No Kubernetes. No cloud bill. No DevOps engineer.

Quickstart

pip install apex-ml
apex start

That's it. Your browser opens to http://localhost:7000.

On first run, Apex prints a one-time login to the terminal:

=== First-run owner account created ===
email:    owner@apex.local
password: <random>
Save this — it will not be shown again.

Log in with those credentials. The password is only shown once.

Requirements

• Python 3.10+
• Docker daemon running (used to execute jobs and dev sessions)
• NVIDIA GPU (optional — Apex degrades gracefully to CPU-only if no GPU is present)
• Linux or macOS (tested on Ubuntu 22.04 + macOS 14)

Build the base images (one-time)

apex build-images                # Python image (~2 min)
apex build-images --pytorch      # + PyTorch/CUDA image (~15 min, ~8 GB)

Submit your first job

curl -X POST http://localhost:7000/api/jobs \
  -H 'Content-Type: application/json' \
  -d '{
    "name": "hello-world",
    "image": "apex/code-server:python",
    "script": "python -c \"print(\\\"hello from apex\\\")\"",
    "gpu_count": 0,
    "priority": "normal"
  }'

Or use the UI — click + Submit job on the dashboard.

Custom images

You can use any Docker image for training jobs:

docker build -t my-team/training-image .
# then specify my-team/training-image when submitting a job

For dev sessions (browser VS Code), the image must have code-server installed. Build on top of the official base image:

FROM apex/code-server:python
RUN pip install torch transformers

Or use codercom/code-server:latest directly.

Train CIFAR-10 in 60 seconds

# Build the PyTorch image (one-time, ~8 GB)
git clone https://github.com/apexhq-dev/apex && cd apex
docker build -t apex/code-server:pytorch -f docker/pytorch.Dockerfile docker/

# Drop the training script into your workspace
cp runway-workspace/cifar10_train.py ~/apex-workspace/

# Submit the job via the UI or curl
# Image:  apex/code-server:pytorch
# Script: python /workspace/cifar10_train.py --epochs 2
# GPU:    1 GPU

The training job hits ~66% test accuracy on an NVIDIA L4 in about 30 seconds.

Screenshots

Overview dashboard — live GPU metrics, stat cards, submit form, job list, active sessions	Live training logs — WebSocket stream of real CIFAR-10 training output
Job history — sortable, filterable, per-row logs + cancel/remove actions	Metrics — large GPU + CPU chart with 8-cell live readings grid
Dev sessions — browser VS Code in one click, workspace pre-mounted	Docker images — reads directly from the host daemon, no registry push

The math

Compare a full month of training compute, 24/7:

Option	Price	Math
RunPod RTX 4090	$316 / mo	$0.44/hr × 720 hrs
Lambda H100	$2,150 / mo	$2.99/hr × 720 hrs
AWS p4d.24xlarge (8× A100)	$23,594 / mo	$32.77/hr × 720 hrs
Your workstation + Apex	$29 / mo	Team tier, 8 seats, unlimited jobs

Pricing

Tier	Price	Gets you
Free	$0 forever	Full feature set, 1 seat, unlimited jobs, community support
Team	$29/mo flat	Everything in Free + 8 seats, multi-user auth, audit log, SSO, priority support
Hosted	$99/mo + GPU	We run Apex for you on a rented GPU, SLA, backups

All tiers run the same codebase. Free and Team are self-hosted on your hardware.

Architecture

Intentionally boring. One Python process, one pip install, one GPU machine.

apex/
├── cli.py                # click CLI: start, stop, status
├── config.py             # loads ~/.apex/config.json
├── docker_mgr.py         # docker SDK wrapper (create, start, clean up)
├── monitor/              # GPU (pynvml) + CPU (psutil) collector thread
├── scheduler/            # SQLite-backed queue + worker thread
├── server/               # FastAPI app + routes
│   ├── app.py            # app factory
│   ├── auth.py           # JWT + bcrypt
│   ├── db.py             # SQLite schema + connection helper
│   └── routes/           # metrics, jobs, sessions, images, users
└── static/               # Vanilla HTML/CSS/JS — no build step

Stack: FastAPI · Uvicorn · SQLite · pynvml · docker-py · SSE-Starlette · code-server · vanilla JS Not in the stack: React Redis Postgres RabbitMQ Kubernetes Helm Node webpack

See SPEC.md for the full technical specification.

CLI

apex start [--host 0.0.0.0] [--port 7000] [--no-browser]
apex status
apex stop              # Ctrl+C works fine, this is a stub
apex logs              # pipe `apex start` output instead
apex build-images      # build base container images (one-time)
apex config show       # print current config
apex config set KEY VALUE   # update a config value (workspace, port, host)
apex --version

Configuration

Apex reads ~/.apex/config.json on startup. Defaults:

{
  "workspace_path": "~/apex-workspace",
  "port": 7000,
  "host": "0.0.0.0",
  "session_port_range": [8080, 8200],
  "jwt_secret": "<auto-generated on first run>"
}

View or change settings with the apex config command:

apex config show                          # print current config
apex config set workspace /mnt/nas/apex   # change workspace path
apex config set port 8000                 # change UI port
apex config set host 127.0.0.1           # bind to localhost only

Or use the APEX_WORKSPACE environment variable (useful for system-level config):

export APEX_WORKSPACE=/mnt/nas/apex
apex start

Shared workspace (NAS / network drives)

Every job and dev session mounts the workspace directory into the container at /workspace. Pointing multiple team members' machines at the same network share means scripts, datasets, and checkpoints are visible to everyone without manual copying.

# 1. Mount the shared drive at the OS level (NFS example)
sudo mount -t nfs 192.168.1.10:/shared /mnt/nas

# 2. Point Apex at it
apex config set workspace /mnt/nas/apex-workspace

# 3. Restart
apex stop && apex start

Samba (SMB) mounts work the same way — mount it at the OS level, then run apex config set workspace <path>. The workspace directory is created automatically if it does not exist.

API

Route	Method	Description
/api/health	GET	{"ok": true}
/api/metrics/stream	GET (SSE)	Live GPU/CPU metrics every 2s
/api/metrics/current	GET	Single snapshot of current metrics
/api/jobs	GET	List jobs (filters: status, limit, offset)
/api/jobs	POST	Submit a new job
/api/jobs/{id}	GET	Job detail
/api/jobs/{id}	DELETE	Cancel running job or remove completed one
/api/jobs/{id}/logs	WS	Stream container logs line-by-line
/api/sessions	GET/POST/DELETE	List, launch, stop dev sessions
/api/images	GET	List Docker images from the host daemon
/api/users/login	POST	JWT login
/api/users/me	GET	Current user
/api/users/invite	POST	Invite a new team member (admin only)

Development

git clone https://github.com/apexhq-dev/apex
cd apex
pip install -e .
apex start --skip-docker-check  # dev mode without Docker

The frontend is plain HTML/CSS/JS in apex/static/ — edit and refresh, no build step.

License

Apache License 2.0 © 2026 Apex contributors

Community

• Website — tryapex.dev
• Discord — discord.gg/RFpDyhdpWJ
• Twitter — @apexhq_dev
• GitHub Issues — for bugs and feature requests
• GitHub Discussions — for questions and ideas