Vibebin – code and host with incus containers on your VPS/server

Vibebin, a new Incus/LXC-based platform for self-hosting persistent AI coding agent sandboxes, has been released, enabling users to create and host vibe-coded apps on a single VPS or server. The platform, which is 99.9% vibe-coded on the exe.dev platform using Shelley Web AI Coding Agent and Claude Opus 4.5, includes Caddy reverse proxy, direct SSH routing, and pre-installed AI coding tools such as opencode, nanocode, claude-code, and shelley. It supports Ubuntu 24.04 LTS and Debian 13 containers and is designed for efficient operation on KVM, VMware, Xen, Hyper-V, and other hypervisors.

An Incus/LXC-based platform for self-hosting persistent AI coding agent sandboxes with Caddy reverse proxy and direct SSH routing to containers suitable for VS Code remote ssh . Create and host your vibe-coded apps on a single VPS/server. Vibebin TUI Manager - Container detail view with quick actions: AI Tools Admin Web App - Toggle AI coding tools on/off: AI Tools Admin Web App - One-click update for all tools: This project is 99.9% vibe-coded on the exe.dev https://exe.dev/i/rlWM5HS3J26ONAE platform using their Shelley Web AI Coding Agent and Claude Opus 4.5. Take that as you will. Note that the above exe.dev URL is a referral link for a free 30 day trial if you register. If you convert to a paid plan, both you and I will receive a one-time $50 in extra Shelley credits. With that said, I am a huge proponent of the exe.dev platform, and if you can, you should definitely try it out and use their service. The love and care for that project/service is extremely evident... AND it is incredibly awesome and I think it's in its infancy stages, so should only get better . This is a very new project. Feel free to test and experiment but it's likely to have bugs and may not be ready for production. Use at your own risk. This project provides the infrastructure to self-host your own AI coding environment on virtually any Linux server -- a VPS, cloud VM EC2, GCP, Azure , or dedicated hardware. Because it uses Incus/LXC container-based virtualization rather than nested VMs , it runs efficiently on KVM, VMware, Xen, Hyper-V, and most other hypervisors. Each container is a fully persistent Linux sandbox running Ubuntu 24.04 LTS Noble or Debian 13 Trixie , with: , opencode https://github.com/anomalyco/opencode , nanocode https://github.com/nanogpt-community/nanocode , and Claude Code https://github.com/anthropics/claude-code - AI coding agents with terminal and web UI interfaces Shelley https://github.com/boldsoftware/shelley AI coding web UI accessible via HTTPS at code.yourdomain.com Basic Auth protected AI Tools Admin web app at admin.code.yourdomain.com for managing AI coding tools Your app/site accessible via HTTPS at yourdomain.com SSH access for direct terminal access to your sandbox VS Code Remote SSH compatible Persistent filesystem that survives container restarts Pre-installed development tools : Docker, Go, Node.js, Bun, Deno, uv AI-assisted development : Use opencode/nanocode/claude-code/shelley as your AI pair programmer with full system access Vibe coding : Spin up isolated sandboxes for experimental projects App/site hosting : Deploy and iterate on web applications Learning environments : Safe, isolated Linux environments for experimentation CI/CD sandboxes : Temporary or persistent build environments | Component | Purpose | |---|---| Incus LXC | Container runtime - lightweight, persistent Linux containers | Caddy | Reverse proxy with automatic HTTPS Let's Encrypt | SSHPiper | SSH routing - access any container via ssh -p 2222 container-name@host | SQLite | Local database for container metadata /var/lib/vibebin/containers.db | Ubuntu/Debian | Native Incus images user choice during creation | opencode | Open source AI coding agent with terminal and web UI | nanocode | NanoGPT-powered AI coding agent fork of opencode | shelley | AI coding web agent from Bold Software | claude-code | Anthropic's CLI coding agent | This project installs opencode , nanocode , claude-code , and shelley in each container: OpenCode https://github.com/anomalyco/opencode is a 100% open source AI coding agent that works with multiple LLM providers: - Anthropic Claude - OpenAI GPT - Google Gemini - OpenCode Zen - Local models via custom endpoints NanoCode https://github.com/nanogpt-community/nanocode is a fork of OpenCode configured to work with NanoGPT https://nano-gpt.com/subscription/hGfwRWc9 as the default provider. It includes: - Automatic model loading from NanoGPT API - Built-in NanoGPT MCP server - Interleaved thinking for reasoning models NanoGPT subscription https://nano-gpt.com/subscription/hGfwRWc9 is $8/month for 60,000 requests across all open models - an excellent value that works great with NanoCode. Note: The nanocode web UI requires LLM configuration before it will work. SSH into the container and run nanocode CLI mode first to configure your LLM provider/API keys. Disclosure: The NanoGPT link is an affiliate link. Claude Code https://github.com/anthropics/claude-code is Anthropic's official CLI coding agent: - CLI only no web UI - Requires Anthropic API key ANTHROPIC API KEY environment variable - Supports Claude models with agentic coding capabilities - Run with claude command in terminal Note: Claude Code is CLI-only and is not managed via the admin.code web UI. Use it directly in the terminal. Shelley https://github.com/boldsoftware/shelley is a powerful AI web agent from Bold Software: - Web-based UI only no CLI mode - Support for multiple LLM providers via environment variables Anthropic, OpenAI, Gemini, Fireworks - Custom model configuration available within the web UI - Ctrl/Cmd-K to start a new conversation or add custom models - Runs on port 9999 like the other AI tools Note: Before starting Shelley, add your API keys to ~/.shelley env . Custom models can be configured in Shelley's web UI, but doing so switches to "custom model mode" and the env var models will no longer be shown. All AI Coding tools support terminal/web UI modes with the exception of Shelley web UI only and Claude Code CLI only . OpenCode, NanoCode, and Shelley support configuring LLM providers from within their web UIs. Through the webUIs on OpenCode/NanoCode, adding NanoGPT as the provider is seamless With Shelley, you just need to add a custom model and set it up with the same concept as shown here https://docs.nano-gpt.com/integrations/openhands . See the NanoGPT integrations https://docs.nano-gpt.com/integrations/claude-code page for setting up NanoGPT with various coding agents. vibebin - Interactive TUI for container management vibebin sync daemon - Background service for config synchronization auto-heals Caddy routes on restart containers.db - SQLite database storing container metadata at /var/lib/vibebin/containers.db Fresh/minimal Linux installation : Ubuntu 22.04+ or Debian 12+ amd64 or arm64 VPS or VM : Works on most virtualization platforms KVM, VMware, Xen, EC2, GCP, Azure, etc. Memory : Minimum 4GB RAM; 8GB+ recommended for best performance Disk : Minimum 70GB; 100GB+ recommended for multiple containers and project data Go 1.21+ : Required to build the tools see Quick Start for installation A domain name with DNS you control A regular user with sudo access avoid running as root Any KVM-based VPS will work well with vibebin. Here are some recommended options: HostBRR - Budget-friendly with excellent performance: HostBRR Epyc Turin VPS https://my.hostbrr.com/order/main/packages/BF2025/?group id=74&a=MTA4OTY= - AMD Epyc Turin processors, great price/performance HostBRR Threadripper Performance VPS https://my.hostbrr.com/order/main/packages/BF2025/?group id=72&a=MTA4OTY= - High-performance Threadripper line Netcup - German provider with reliable infrastructure: Netcup VPS https://www.netcup.com/en/?ref=75646 - Recommend RS1000 8GB RAM or RS2000 16GB RAM for vibebin workloads OVHcloud - Large European provider with competitive pricing: OVHcloud VPS https://www.ovhcloud.com/en/vps/vps-8gb/ - Recommend 8GB or 12GB VPS; larger options 24GB, 48GB available at excellent prices if needed Oracle Cloud Infrastructure OCI - Free tier available: OCI Always Free Resources https://docs.oracle.com/en-us/iaas/Content/FreeTier/freetier topic-Always Free Resources.htm - Includes free ARM-based VMs up to 4 OCPUs, 24GB RAM that work well with vibebin Disclosure: The HostBRR and Netcup links above are affiliate links. Complete these security steps before installing vibebin. Ensure your host SSH is properly secured in /etc/ssh/sshd config : PermitRootLogin no PasswordAuthentication no Restart SSH after changes: sudo systemctl restart sshd If you want to use UFW firewall, install and configure it with the required rules: Install UFW sudo apt update && sudo apt install -y ufw Allow required ports SSH must be allowed before enabling sudo ufw allow 22/tcp Host SSH sudo ufw allow 80/tcp HTTP Caddy sudo ufw allow 443/tcp HTTPS Caddy sudo ufw allow 2222/tcp SSHPiper container SSH Enable UFW sudo ufw enable sudo systemctl enable ufw After the first vibebin run which installs Incus , add the bridge rules for container networking: Allow Incus bridge traffic run after vibebin installs Incus sudo ufw allow in on incusbr0 sudo ufw route allow in on incusbr0 sudo ufw route allow out on incusbr0 Note:UFW is enabled at boot automatically after ufw enable . Verify status with sudo ufw status verbose . For additional protection against brute-force attacks, you can install fail2ban. By default, fail2ban protects the host SSH daemon sshd . The docs/fail2ban.md /jgbrwn/vibebin/blob/main/docs/fail2ban.md guide extends this with Caddy and SSHPiper protection. Note:Install fail2banafterthe initial vibebin run so that Caddy and SSHPiper are already in place. All administrative tasks should be performed as a regular user with sudo privileges, not as root directly. Install essential tools sudo apt update && sudo apt install -y wget curl git jq htop tmux Ubuntu/Debian - install from official Go downloads wget https://go.dev/dl/go1.23.5.linux-amd64.tar.gz sudo rm -rf /usr/local/go && sudo tar -C /usr/local -xzf go1.23.5.linux-amd64.tar.gz echo 'export PATH=$PATH:/usr/local/go/bin' ~/.bashrc source ~/.bashrc Verify installation go version Note: For ARM64 systems, use go1.23.5.linux-arm64.tar.gz instead. Check https://go.dev/dl/ for the latest version. Use the install script for a one-liner install or upgrade: curl -fsSL https://raw.githubusercontent.com/jgbrwn/vibebin/main/install-upgrade.sh | bash The script automatically: - Clones and builds the project - Installs binaries to /usr/local/bin/ - Detects fresh install vs upgrade - Handles the vibebin-sync service accordingly If you prefer to build manually: Clone and build git clone https://github.com/jgbrwn/vibebin.git cd vibebin go build -o vibebin vibebin.go go build -o vibebin sync daemon vibebin sync daemon.go Install binaries sudo cp vibebin vibebin sync daemon /usr/local/bin/ This auto-installs Incus, Caddy, and SSHPiper sudo vibebin See the SSHPiper Manual Setup %EF%B8%8F-required-sshpiper-manual-setup-after-first-run section below. sudo vibebin The creation wizard will guide you through: - Enter domain name - Select base image Ubuntu or Debian - Configure DNS optional auto-creation via Cloudflare/deSEC - Set app port - Provide SSH public key - Set basic auth credentials for web UI protection The first run automatically installs: Incus 6.20+ from Zabbly stable repository Caddy web server with automatic HTTPS SSHPiper SSH routing proxy During container creation, the following is automatically installed: Development Tools: Docker via official get.docker.com script Go latest version, architecture auto-detected Node.js latest LTS via NodeSource Bun JavaScript runtime and package manager Deno JavaScript/TypeScript runtime uv Python package manager from Astral GitHub CLI gh for GitHub operations Build tools build-essential, make, etc. AI Coding Agents: opencode open source AI coding agent nanocode NanoGPT-powered AI coding agent claude-code Anthropic's CLI coding agent shelley AI web agent from Bold Software System Utilities: Editors : neovim, mc Midnight Commander Terminal multiplexers : screen, tmux, byobu Monitoring : htop, atop, btop, iotop, ncdu Networking : dnsutils, net-tools, socat, lftp, ncftp Search : ripgrep, sqlite3 Media : imagemagick, ffmpeg Fonts : emoji support fonts-noto-color-emoji, fonts-symbola - And many more development and debugging utilities Configuration: Project directory ~/projects for AI coding tool workspaces Custom MOTD shows container info, URLs, and tool versions on SSH login Passwordless sudo for the container user Ping without sudo CAP NET RAW capability Before creating containers , verify SSHPiper is running: Check SSHPiper status should be active sudo systemctl status sshpiperd If not running, start it sudo systemctl enable --now sshpiperd SSHPiper listens on port 2222 for container SSH access. Host SSH remains on port 22. sudo vibebin List View: n - Create new container Enter - View container details d - Delete container u - Show untracked containers import existing D - Manage DNS API tokens Cloudflare/deSEC i - View Incus logs l - View sync daemon logs q - Quit Detail View: s - Start/Stop container r - Restart container p - Change app port a - Change auth credentials S - Snapshot management u - Update AI coding tools opencode/nanocode/claude-code/shelley Esc - Back to list Snapshot View: n - Create new snapshot Enter/r - Restore selected snapshot d - Delete selected snapshot ↑/↓ or j/k - Navigate snapshots Esc - Back to container details Native Incus Images : Choose Ubuntu 24.04 LTS or Debian 13 Persistent Sandboxes : Full filesystem persistence across restarts Boot Behavior : Containers respect their last state on host reboot Resource Monitoring : Live CPU and memory usage in TUI Untracked Import : Detect and adopt existing Incus containers Snapshots : Create, restore, and delete container snapshots Automatic HTTPS : Caddy handles Let's Encrypt certificates Reverse Proxy : Each container gets https://domain.com → Container's app port 8000, configurable SSH Routing : SSHPiper on port 2222 enables ssh -p 2222 container-name@host access Auto DNS : Cloudflare and deSEC API integration tokens saved securely for reuse opencode : Open source, supports multiple LLM providers nanocode : NanoGPT-optimized fork with built-in features claude-code : Anthropic's official CLI agent CLI only shelley : AI web agent with full coding capabilities Easy Configuration : All tools prompt for API keys on first run Web UI Access : Start opencode, nanocode, or shelley in serve mode on port 9999 After creating a container, SSH in and run any of the tools. A project directory is pre-created at ~/projects for your coding workspaces. SSH to your container ssh -p 2222 container-name@host.example.com cd to projects directory first cd ~/projects Run opencode opencode Or run nanocode nanocode Or run claude-code claude All tools will prompt you to configure your LLM provider and API key on first run. Note that Shelley is web-only no CLI mode and Claude Code is CLI-only no web UI . To access the AI coding agent via web browser: SSH to your container ssh -p 2222 container-name@host.example.com opencode cd to project directory first : cd ~/projects screen -S code opencode serve --port 9999 --hostname 0.0.0.0 nanocode cd to project directory first : cd ~/projects screen -S code nanocode serve --port 9999 --hostname 0.0.0.0 Note:nanocode web UI requires LLM configuration first. Run nanocode CLI to configure your provider/API keys before starting the web UI. shelley web UI only - no CLI mode : cd ~/projects screen -S code start-shelley.sh 2 &1 | tee -a ~/.shelley.log Note:Before starting Shelley, add your API keys to ~/.shelley env . Custom models can be configured in Shelley's web UI, but doing so switches to "custom model mode" and the env var models will no longer be shown. Then access via https://code.yourdomain.com protected by Basic Auth credentials you set during container creation . Press Ctrl+A, D to detach from screen. Reattach with screen -x code . Note: Only one web UI can run on port 9999 at a time. Stop the current one before starting another. Each container includes an AI Tools Admin web app accessible at https://admin.code.yourdomain.com . This provides: MANAGE View: - Toggle AI coding tools OpenCode, NanoCode, Shelley on/off - Only one tool can run at a time on port 9999 - View real-time output logs - Links to App URL and Code UI - DNS health check indicators UPDATE View: - One-click update for all three AI tools - Automatically stops running processes before updating - Shows live progress of updates The admin app runs as a systemd service admin-app and is protected by the same Basic Auth credentials as the Code UI. Note: For Shelley, ensure your API keys are configured in ~/.shelley env before starting. From the container detail view in the TUI, press u to update opencode, nanocode, claude-code, and shelley to their latest versions. ssh -p 2222 container-name@host.example.com You'll be logged in as 'ubuntu' or 'debian' with sudo access ssh user@host.example.com For HTTPS to work, DNS must point to the host server: domain.com → Host IP code.domain.com → Host IP for AI coding web UI admin.code.domain.com → Host IP for AI tools admin app Caddy will automatically obtain Let's Encrypt certificates for both domains. Internet │ ┌────────────┴────────────┐ │ │ ▼ ▼ HTTPS :443 SSH :2222 via SSHPiper │ │ ┌───────────────────▼─────────────────────▼─────────────────┐ │ Host System │ ├─────────────────────────────────────────────────────────────┤ │ │ │ ┌─────────────┐ ┌─────────────┐ │ │ │ Caddy │ │ SSHPiper │ │ │ │ reverse │ │ SSH │ │ │ │ proxy │ │ router │ │ │ └──────┬──────┘ └──────┬──────┘ │ │ │ │ │ │ │ Routes by domain │ Routes by username │ │ ▼ ▼ │ │ ┌─────────────────────────────────────────────────────┐ │ │ │ Incus LXC │ │ │ │ │ │ │ │ ┌─────────────────┐ ┌─────────────────┐ │ │ │ │ │ Container │ │ Container │ ... │ │ │ │ │ Ubuntu/Debian │ │ Ubuntu/Debian │ │ │ │ │ │ │ │ │ │ │ │ │ │ ┌───────────┐ │ │ ┌───────────┐ │ │ │ │ │ │ │ AI Coding │ │ │ │ AI Coding │ │ │ │ │ │ │ │ Tool UI │ │ │ │ Tool UI │ │ │ │ │ │ │ │ :9999 │ │ │ │ :9999 │ │ │ │ │ │ │ └───────────┘ │ │ └───────────┘ │ │ │ │ │ │ ┌───────────┐ │ │ ┌───────────┐ │ │ │ │ │ │ │admin.code │ │ │ │admin.code │ │ │ │ │ │ │ │ :8099 │ │ │ │ :8099 │ │ │ │ │ │ │ └───────────┘ │ │ └───────────┘ │ │ │ │ │ │ ┌───────────┐ │ │ ┌───────────┐ │ │ │ │ │ │ │ Your App │ │ │ │ Your App │ │ │ │ │ │ │ │ :8000 │ │ │ │ :8000 │ │ │ │ │ │ │ └───────────┘ │ │ └───────────┘ │ │ │ │ │ └─────────────────┘ └─────────────────┘ │ │ │ └─────────────────────────────────────────────────────┘ │ │ │ │ ┌───────────────────┐ │ │ │ vibebin │ TUI for container management │ │ │ this tool │ - Create/delete containers │ │ │ │ - Configure domains & auth │ │ └───────────────────┘ │ │ │ └─────────────────────────────────────────────────────────────┘ HTTPS requests to myapp.example.com → Caddy → Container's app port 8000 HTTPS requests to code.myapp.example.com → Caddy with Basic Auth → AI coding web UI port 9999 SSH connections to port 2222 as myapp-example-com@host → SSHPiper → Container's SSH as ubuntu / debian Routes are managed via Caddy's Admin API localhost:2019 , not config files: - Routes use @id for identification e.g., container-name-app , container-name-code - Changes are atomic and immediate no reload required Self-healing : The vibebin-sync service monitors Caddy routes and automatically restores them if the server restarts, Caddy restarts, or routes are otherwise modified/removed- Query current routes: curl http://localhost:2019/config/apps/http/servers/srv0/routes Containers use Incus's default "last-state" behavior by not setting boot.autostart : - Running containers will restart when the host reboots - Stopped containers will stay stopped Incus automatically tracks each container's power state and restores it when the daemon starts. Snapshots allow you to save and restore the complete state of a container. From the container detail view, press S to access snapshot management, then n to create. Before risky changes : Snapshot before major updates or experiments Known-good states : Save working configurations you can restore to Quick rollback : Instantly revert if something breaks Note: Snapshots are stored by Incus and consume disk space. Delete old snapshots to free space. Container won't start: journalctl -u incus -f incus info container-name Caddy certificate errors: - Ensure DNS is configured and pointing to host IP before creating container - Check Caddy logs: journalctl -u caddy -f - Check current routes: curl -s http://localhost:2019/config/apps/http/servers/srv0/routes | jq . SSH to containers not working: - Verify SSHPiper is running: systemctl status sshpiperd - Ensure you're using port 2222: ssh -p 2222 container-name@host - Verify upstream config: cat /var/lib/sshpiper/container-name/sshpiper upstream AI coding tools not working: - SSH to the container and run opencode , nanocode , or claude interactively, or start-shelley.sh for Shelley web UI - All tools will prompt for API key configuration on first run - Check tool versions: opencode --version , nanocode --version , claude --version , shelley version Sync daemon issues: journalctl -u vibebin-sync -f You can use subdomains for your containers: app.example.com - works correctly staging.app.example.com - works correctly my-app.example.com - works correctly Domains with two-part TLDs like .co.uk , .com.au are not fully supported for automatic DNS creation. The zone detection assumes a single-part TLD. For example: - ✅ app.example.com → zone: example.com correct - ❌ app.example.co.uk → zone: co.uk incorrect, should be example.co.uk Workaround : For two-part TLDs, select "No" for auto DNS creation and configure DNS records manually. Each container created by vibebin is fully isolated by default: - Each container has its own complete, persistent filesystem Ubuntu 24.04 or Debian 13 - Containers cannot access each other's filesystems - Uses Incus/LXC which provides OS-level virtualization with separate root filesystems - Data persists across container restarts - Each container gets its own private IP on an internal bridge network typically 10.x.x.x - Containers can communicate with each other by default as they are on the same bridge network on the host - All external access is routed through: Caddy HTTPS reverse proxy for web traffic SSHPiper port 2222 for SSH access - Each container only exposes what's explicitly configured by default this is: app port/8000 configurable , code UI port/9999, admin.code/8099 - Containers use Linux namespaces PID, network, mount, user, etc. - Processes in one container cannot see or interact with processes in another - Each container has its own init system and process tree security.nesting=true is enabled to allow Docker-in-container useful for development . This is standard for development environments but worth noting for security-sensitive deployments. If you run 3 containers app1.example.com, app2.example.com, app3.example.com : - Each has its own isolated Linux environment - Each has its own AI tools installation opencode, nanocode, claude-code, shelley - An agent in app1 cannot access files or processes in app2 or app3 - They share the host's resources CPU, RAM but are otherwise independent It's essentially like having separate lightweight VMs, but with much lower overhead than full virtualization. Currently, this implementation uses the Incus DIR storage driver for proof of concept. The DIR driver uses basic filesystem-level storage and is: - Simple to set up no additional dependencies - Compatible with any filesystem Slow for snapshots full copy-on-write not available Planned : During the dependencies/installation phase, users will be able to choose between: | Driver | Pros | Cons | |---|---|---| DIR | Simple, works everywhere | Slow snapshots, no CoW | Btrfs | Fast snapshots, CoW, compression | Requires Btrfs filesystem | ZFS | Fast snapshots, CoW, excellent features | Requires significant RAM 1GB+ per TB of storage | Btrfs and ZFS provide instant snapshots via copy-on-write, making them much more suitable for production use. This project is licensed under the MIT License. See LICENSE /jgbrwn/vibebin/blob/main/LICENSE for details. opencode : anomalyco/opencode https://github.com/anomalyco/opencode - MIT License nanocode : nanogpt-community/nanocode https://github.com/nanogpt-community/nanocode - Fork of opencode claude-code : anthropics/claude-code https://github.com/anthropics/claude-code - Apache License 2.0 shelley : boldsoftware/shelley https://github.com/boldsoftware/shelley - Apache License 2.0 Incus : linuxcontainers/incus https://github.com/lxc/incus - Apache 2.0 License Caddy : caddyserver/caddy https://github.com/caddyserver/caddy - Apache 2.0 License SSHPiper : tg123/sshpiper https://github.com/tg123/sshpiper - MIT License