Modal Personal DevBox Launcher

April 6, 2026 · View on GitHub

Table of Contents

Backstory

Most of us aren't privileged enough to buy those high end PCs or pay for cloud VPS subscriptions from AWS, GCP, and the rest. We mostly stick to their free tiers which, let's be real, are super limited. But either way, we are forever grateful for the generosity from these companies 🙏.

The AI Exploration Rabbit Hole

About a while back last year, I was actively exploring the AI landscape. And when I say explore, I mean explore — I was freaking looking into a lot of new technologies as they came to the space. New models, new prompting techniques, image and video generation, embeddings — I mean virtually everything! Heck, at one point I had wanted to train my own AI model 😅, but I told myself calm down, you no fit sabi everything (although, not totally crossed off my checklist yet 😉).

But anyone who's been in the AI space already knows you can't run AI models on potato hardware. Even at that, you can't get decent results from some models under 4B parameters, and from 4B above you're looking at purchasing expensive GPUs to get something going. Ayoo! Doesn't mean I didn't try running some models on my potato PC 😂, but damn was it slow. I'll leave this at that.

My Hardware Reality

u0_a191 @localhost
-----------------
OS: Android 7.1.1 armv7l
Host: samsung SM-G600S
Kernel: 3.10.49-13209801
CPU: Qualcomm MSM8916 (4) @ 998MHz
Memory: 1484MiB / 1889MiB

Yeah not exactly a powerhouse this phone. Most stuff I did on it was through Termux and trying to do any kind of serious work on it was a constant struggle. The thing would get hot enough to be uncomfortable, lag at the worst possible moments 😭, and just couldn't handle the tools I wanted to experiment with. It was frustrating but I loved the little guy either ways 😆

Potato PC
---------
CPU: Intel Celeron N3150 @ 2.08GHz (max)
RAM: 8GB

I did also have a laptop with entry level specs but bandwidth data where I'm from is a huge problem and really expensive. Downloading large packages was a straight up NO GO zone!

At about the same time, I was also scanning the VPS landscape, constantly looking for an affordable but quality solution for my use case. Although the search didn't quite come out successfully because of my budget at the time.

The Rabbit Hole Journey

I am one of those guys who has watched nearly every YouTube video promising "free VPS servers". Man that was a wild and fun rabbit hole. It exposed me to a ton of scripts clever workarounds, which ultimately shaped my knowledge in networking and Linux as a whole. Still got a long way to go sha, but it was a good run.

Discovering Modal and the Spark

But as time moved on, I casually stumbled across this platform — Modal. Look man, I think this has got to be one of the most insane coincidences ever. The AI Platform For Developers was the first message I was greeted with on their website. Then I started looking around, read their docs, looked at their example codes, and went through their guide on how to get started.

I was experimenting, trying out their examples, noting down key points, ran some AI models I always wanted to try out, saw that their systems use Linux underneath the hood, saw in the docs the ability to expose ports as tunnels, saw storage functionality, and the mindmap just kept expanding — linking pieces together, running some awesome stuff from their docs. And boom, an idea emerged 💡.

My Conviction and Closing Note

One thing has always been crystal clear to me though, true innovation comes from the most trying and difficult times in a person's life. You look at every available option and alternative, dig deeper than anyone else into a specific problem, and just when everything feels pointless BAM! you hit a breakthrough. One that doesn't just solve your problem but ends up helping others too.

My brothers and sisters the world is EVIL. But do not conform to the pattern of this world. Love one another, help others in need and be the reason someone smiles today 😊

Which brings me, ladies and gentlemen, to the main focus of my creation log and overview — DevBox.

I hope you enjoy using DevBox as much as I am!

⚠️ Warnings & Open Letter

I plead with all users of this project to refrain from any illegal activities on the Modal platform. This includes but is not limited to:

  • Cryptocurrency mining — Do not use DevBox for mining any cryptocurrency
  • DDoS attacks — Do not use this project to launch distributed denial-of-service attacks
  • Hosting illegal content — Do not store or distribute illegal material
  • Spam & botnets — Do not run spam operations or botnet infrastructure
  • Password cracking — Do not use DevBox for unauthorized credential attacks
  • Any activity that violates Modal's Terms of Service

It's common for people to try and abuse free or generous cloud platforms, which ultimately hurts legitimate experimenters, researchers, and developers who are just trying to learn, build, and create awesome stuff. Don't be that person.

An open letter to the team at Modal:

I plead with you to ban only the accounts of people who actively abuse your platform, and leave legitimate users be. The vast majority of us are here to experiment, learn, and build real things. Please don't let the actions of a few bad actors restrict the many who use your platform responsibly.

Thank you all for your understanding. Enjoy DevBox!

Quick Start

Already got Python and a Modal account? You're 3 commands away from your first DevBox.

Step 1: Install the Modal CLI:

pip install modal
modal setup

Step 2: Set up your SSH key (DevBox needs this to let you in):

modal secret create ssh-public-key PUBKEY="$(cat ~/.ssh/id_rsa.pub)"

Step 3: Launch it:

modal run devbox.py

After a bit (first run takes 30-45 mins to build the image), you'll see something like this:

🚀 Your DevBox is ready!
Paste this command into your terminal:
ssh root@<host> -p <port>

Paste that SSH command into a new terminal and you're in. That's it.

Heads up: Only the /data directory is persistent. Save your work there. Everything else resets on each launch.

For the full setup guide (all 3 secrets, GPU options, RDP, WebUI), check the sections below.

Overview

This repository contains a Modal script (devbox.py) for launching a secure, personal, and on-demand remote development environment (DevBox). It provides a flexible and cost-effective way to get a powerful, cloud-based development environment with persistent storage, accessible via SSH.

The script launches a general-purpose DevBox, a Debian environment with essential development tools, and allows for the dynamic installation of extra packages at launch time.

Features

  • Standard DevBox: A general-purpose Debian environment with essential development tools (git, vim, curl, build-essential, etc.). Supports optional GPU acceleration (T4, L4, A10G). Allows dynamic installation of extra apt packages at launch time.
  • Persistent Storage: Uses a modal.Volume to provide a persistent /data directory, ensuring your work is saved between sessions.
  • Secure SSH Access: Automatically injects your public SSH key for secure, passwordless access.
  • Auto-Shutdown: Includes an idle timer that automatically shuts down the container after 5 minutes of inactivity to save costs.
  • Resource Allocation: The environment is configured with appropriate CPU and memory resources for its intended task.
  • Interactive Launcher: A simple command-line prompt to choose from 7 specialized DevBox configurations.

Launcher Options

When you run modal run devbox.py, you'll be presented with 7 DevBox configurations. Here's what each one does:

1. 🛠️ Standard DevBox

A general-purpose Debian environment with essential dev tools — git, nano, neovim, curl, build-essential, clang, cmake, htop, and more. You can optionally attach a GPU (T4, L4, or A10G) at launch. Install extra apt packages on the fly.

Best for: Everyday coding, scripting, compiling, general Linux work.

Resources: 0.5 vCPU / 1GB RAM (CPU) · 1.0 vCPU / 2GB RAM (GPU)

2. 📄 Document Processing Box

Comes with Pandoc and a full TeX Live distribution pre-installed. Convert between document formats (Markdown → PDF, LaTeX → HTML, etc.) with everything you need for professional document workflows.

Best for: LaTeX/PDF generation, academic papers, technical documentation, format conversion.

Resources: 1.0 vCPU / 2GB RAM

3. 🤖 AI Assistants Box

Pre-installed with OpenCode and Gemini CLI for AI-assisted coding sessions right out of the box.

Best for: AI-powered pair programming, code review, and development assistance.

Resources: 0.5 vCPU / 1GB RAM

4. 🧪 LLM Playroom

Comes with Ollama pre-installed. Pull and run any open-source model with a single command — no local hardware needed.

Best for: Running and testing open-source language models, AI experimentation.

Resources: 1.0 vCPU / 2GB RAM

5. 🖥️ RDP Desktop Box

A full XFCE graphical desktop accessible via RDP. Comes with xrdp, xfce4, and a suite of desktop goodies. Username is root, password is devbox123. Optional GPU acceleration available.

Best for: When you need a full desktop environment — GUI apps, browsers, visual workflows.

Resources: 1.0 vCPU / 2GB RAM (CPU) · 1.5 vCPU / 4GB RAM (GPU)

6. 🔬 llama.cpp Research Center

The most feature-rich box. Comes with a curated catalog of 7 pre-configured GGUF models plus the ability to load any custom HuggingFace GGUF. Each model has individually tuned temperature, top_p, and top_k settings. Includes a built-in WebUI (SvelteKit-based chat interface on port 8080) and EXA Search integration for function-calling web search.

Models available:

  • Qwen3.5-9B (General reasoning, 256K context)
  • Dolphin3.0-3B (Fast, abliterated)
  • DeepSeek-R1-7B-Uncensored (Math/science specialist)
  • Qwen2.5-Coder-7B (Coding specialist)
  • SmolVLM-500M (Image captioning)
  • Gemma3-4B (Multimodal)
  • Custom (Any HuggingFace GGUF URL)

Note: Model selection happens after you connect via SSH — you'll be prompted interactively.

Resources: 2.0 vCPU / 8GB RAM · 1 hour idle timeout

7. 🔍 Forensics Analysis Box

Comes with Volatility3 and symbol files for Windows, Linux, and macOS pre-installed. Memory forensics is a branch of digital forensics that involves extracting evidence and artifacts from a RAM image or dump of a running computer. This lets you analyze running processes, environment variables, active network connections, loaded kernel modules, and much more.

Best for: CTF challenges, incident response, malware analysis, digital forensics investigations.

Resources: 0.5 vCPU / 1GB RAM

GPU Options

Several DevBox types support optional GPU acceleration. Here's a quick comparison:

GPUBest ForVRAM
T4Cost-effective inference, light workloads16 GB
L4Newer architecture, better performance than T424 GB
A10GHeavy workloads, more VRAM, highest performance24 GB

⚠️ GPU usage incurs charges on your Modal account. See modal.com/pricing for current rates.

For reference, Modal's free tier covers approximately 270 hours/month of CPU-only usage. GPU hours are billed separately.

Prerequisites

Before you can use this script, you need to have the following:

  1. Python: Python 3.10 or newer installed on your local machine.
  2. Modal Account: A free Modal account. You can sign up at modal.com.
  3. Modal CLI: The modal Python library installed and configured. 
    pip install modal
modal setup
  4. SSH Key Pair: An SSH key pair generated on your local machine. If you don't have one, you can create it with: 
    ssh-keygen -t rsa -b 4096

This will work on both Windows, Mac & Linux Machines

Setup

DevBox uses Modal Secrets to securely store your credentials. Here's what you need:

Required: SSH Key

This is needed for all DevBox types. Without it, the container will still start but you won't be able to connect via SSH.

  1. Find your public key: Your public key is usually at ~/.ssh/id_rsa.pub (or ~/.ssh/id_ed25519.pub if you use ed25519). Copy the entire content.

  2. Create the secret:

modal secret create ssh-public-key PUBKEY="$(cat ~/.ssh/id_rsa.pub)"

Optional: Gemini API Key

Only needed if you plan to use the AI Assistants Box (OpenCode + Gemini CLI). If missing, Gemini CLI will prompt you for the key on first use — no hard crash.

Get your key from Google AI Studio, then:

modal secret create gemini-api-key GEMINI_API_KEY="your-api-key-here"

Optional: EXA API Key

Only needed if you plan to use the llama.cpp Research Center with web search. If missing, EXA Search is disabled and the box prints a warning, but everything else works fine.

Get your key from EXA Platform, then:

modal secret create exa-api-key EXA_API_KEY="your-api-key-here"

Connecting

After launching your DevBox, Modal will print the connection info once the container is ready. How you connect depends on which box type you chose:

SSH (All Boxes)

Every DevBox supports SSH access. Once ready, you'll see:

🚀 Your DevBox is ready!
Paste this command into your terminal:
ssh root@<host> -p <port>

Just copy and paste that command into a new terminal window and you're in.

RDP (RDP Desktop Box)

The RDP Desktop Box gives you a full XFCE graphical desktop. Once ready, you'll see:

============================================================
🖥️ Your RDP Desktop is ready!
============================================================

📡 RDP Address: <host>:<port>
👤 Username: root
🔑 Password: devbox123

============================================================

Use those credentials with an RDP client:

PlatformRecommended Client
Android/iOSHaven SSH Client, aRDP, or Windows App
WindowsBuilt-in Remote Desktop (mstsc.exe)
macOSMicrosoft Remote Desktop (App Store)
LinuxRemmina

WebUI (llama.cpp Research Center)

The llama.cpp Research Center includes a built-in WebUI — a browser-based chat interface powered by llama-server's SvelteKit frontend. Once ready, you'll see:

🌐 WebUI: http://<host>:<port>
   (Open in browser for chat interface)

Just paste that URL into any modern browser (Chrome, Firefox, Safari, Edge) and you'll get a full chat interface with your selected model. No additional setup needed.

Alias for Quick Launch

To make launching your DevBox even faster, you can create an alias for the modal run devbox.py command in your shell configuration file (e.g., .bashrc, .zshrc, or PowerShell profile).

For Bash/Zsh:

Add the following line to your ~/.bashrc or ~/.zshrc file:

alias devbox="modal run -d /path/to/your/DevBox/devbox.py"

Here i added the "-d" flag for the container to run in "detached" mode, when network hiccups occur, the instances are usually terminated when run with the normal "modal run devbox.py" command, but by using "modal run -d devbox.py" it runs without direct connection to your device, but uhm... Just don't forget to terminate when you're done 🙂

Remember to replace /path/to/your/DevBox/ with the actual path to your devbox.py file. After saving, reload your shell configuration:

source ~/.bashrc  # or source ~/.zshrc

Now you can simply type devbox in your terminal to launch your DevBox.

For PowerShell:

Add the following lines to your PowerShell profile (you can open it by typing $profile and then notepad $profile in PowerShell):

function Start-DevBox { modal run C:\path\to\your\DevBox\devbox.py }
Set-Alias -Name devbox -Value Start-DevBox

Remember to replace C:\path\to\your\DevBox\devbox.py with the actual path to your devbox.py file. After saving, restart PowerShell or run . $profile.

Important Notes

Storage

  • /data is persistent: This is the only directory that survives between sessions. Save all your important work here.
  • /root is ephemeral: Your home directory resets on every container start. Dotfiles, installed packages, and configs in /root will be wiped.

Idle Timeouts

  • Standard boxes: Auto-shutdown after 5 minutes of no active SSH connection.
  • llama.cpp Research Center: Extended timeout of 1 hour to accommodate long-running inference sessions.
  • A countdown warning will appear in the terminal where you ran modal run before shutdown.

First Build Time

The first run will take 30-45 minutes to build the container image. Subsequent runs are instant — Modal caches the image so you're up and running in seconds.

Cost Estimate

Running a Standard DevBox for 1 hour costs approximately $0.20. GPU usage incurs additional charges — see modal.com/pricing for current rates.

Modal containers run in a sandboxed environment. Here are some things to be aware of:

  • Init system: Containers use dumb-init, not systemd. You won't have systemd services available.

    root@modal:/data# ps -p 1 -o comm=
dumb-init
root@modal:/data#

  • No virtualization: Containers don't have access to /dev/kvm, so you can't run VMs or nested virtualization.

  • No Docker/containerd: Running Docker or containerd inside a Modal container is not supported.

  • No Unix sockets: Modal doesn't support Unix sockets. This means services like MariaDB that rely on them won't work out of the box.

  • No Homebrew: brew does not work inside Modal containers.

  • No sudo by default: You're already running as root, so sudo isn't needed (or installed by default).

  • htop readings: Modal containers run on massive shared instances. htop shows host-level resource usage, not your container's allocated resources. Don't trust the CPU/RAM percentages you see there.

Secrets — How They Work

DevBox uses Modal Secrets to securely pass credentials into containers at runtime. Here's the flow:

  1. You create secrets locally with modal secret create <name> KEY="value"
  2. Modal stores them encrypted and injects them as environment variables inside the container
  3. DevBox code reads them via os.environ.get() — they never appear in your code
SecretEnv VarRequired ForIf Missing
ssh-public-keyPUBKEYAll boxesContainer starts, but you can't SSH in
gemini-api-keyGEMINI_API_KEYAI Assistants BoxGemini CLI prompts for key on first use
exa-api-keyEXA_API_KEYllama.cpp Research CenterEXA Search disabled, everything else works

Understanding GGUF Model Names

If you're using the llama.cpp Research Center, you'll see model filenames like gemma-3-4b-it-Q4_K_S.gguf. Here's what each part means:

PartMeaningExample
Model familyThe base modelgemma-3, qwen2.5-coder
VersionModel generation3, 2.5
Parameter countModel size4b (4 billion), 7b (7 billion)
TuningTraining typeit (instruction-tuned), blank (base)
QuantizationCompression levelQ4_K_S (4-bit, small), Q5_K_M (5-bit, medium)

Lower quantization = smaller file, faster inference, slightly lower quality. Higher quantization = better quality but needs more RAM.

Troubleshooting

"Secret not found" errors

Make sure you've created all the required secrets. Run modal secret list to check what's stored. If a secret is missing, recreate it with the modal secret create commands from the Setup section.

SSH key not working or stale keys

If you've regenerated your SSH key or it's just not connecting, old keys might be lingering on Modal's side. Delete the stale secret and reinsert it:

modal secret delete ssh-public-key
modal secret create ssh-public-key PUBKEY="$(cat ~/.ssh/id_rsa.pub)"

First build taking forever

Yep, that's normal. The first run takes 30-45 minutes because Modal is building the container image from scratch. Grab a coffee, stretch your legs — subsequent runs will be instant.

Container terminated unexpectedly

Usually caused by network hiccups on your end. If you're running in attached mode (modal run devbox.py), a dropped connection can kill the container. Try running in detached mode instead:

modal run -d devbox.py

Just don't forget to terminate it when you're done 🙂

RDP not connecting

Double-check the address, username, and password printed when the container started. Make sure your RDP client is connecting to the right port. If you're on mobile, try a different client — some can be finicky.

llama.cpp model download fails

Check your internet connection and make sure the HuggingFace URL is valid. Some models require you to be logged in to HuggingFace — if it's a gated model, you'll need to set up HF token authentication. Try a different model from the catalog first to isolate the issue.

Architecture

The codebase is currently flat — all files live in the root directory. A structure rehaul is planned for the future, but for now here's what each file does:

FilePurpose
devbox.pyMain launcher — interactive menu, Modal function definitions, entry point
config.pyResource allocations (CPU, memory, timeouts), package lists, GPU configs
images.pyContainer image definitions for all 7 DevBox types
shared_runtime.pyShared runtime logic — SSH/RDP idle monitoring, auto-shutdown, connection output
utils.pySSH key injection utility
gpu_utils.pyGPU configuration helpers and base GPU launch args
exa_helper.pyEXA Search integration for llama.cpp Research Center
quotes_loader.pyRandom motivational quotes displayed at launcher startup
quotes.jsonQuote catalog loaded by quotes_loader.py
ui_utils.pyUI helper functions — create_box(), show_spinner()
backup_utils.pyBackup registration and data persistence helpers
persistence_utils.pySymlink management for dotfiles persistence across sessions

Roadmap

See ROADMAP.md for planned improvements and future work.

Contributing

The project is open to contributions and optimizations. No formal guidelines yet — just make sure your changes work and don't break existing functionality. If you want to add a new DevBox type, look at how the existing ones are defined in devbox.py and images.py — they follow a consistent pattern.

Thanks & Inspirations

The Project Is Open To Your Contributions And Or Optimizations 🙏