Your own persistent AI coding workspace on AWS. SSH in from any device, reconnect where you left off, and run Claude Code or Codex inside the same environment.

~$14/mo on AWS. One slash command to set up.

Don't want to manage infrastructure? Always-On Claude Hosted gives you a fully managed workspace — no AWS account, no Docker, no setup. Pay, get an SSH key, connect.

• Claude Code CLI installed (npm install -g @anthropic-ai/claude-code)
• Claude auth for provisioning slash commands — either a subscription (Pro/Max) or an API key (ANTHROPIC_API_KEY)
• AWS account with CLI configured (brew install awscli && aws configure)

The runtime workspace itself supports both Claude Code and Codex. Set DEFAULT_CODE_AGENT=codex before install/provision (or export it on the workspace host) if you want the SSH workspace picker to launch Codex by default. On provisioned hosts, first-run onboarding now guides Codex users through codex --login; on remote SSH hosts that means completing the device-code browser step Codex shows to sign the workspace into ChatGPT for subscription-based access. Codex auth can also use OPENAI_API_KEY. Provisioned hosts also set Codex defaults to approval_policy = "never" and sandbox_mode = "danger-full-access" in ~/.codex/config.toml, and re-materialize repo-managed Codex home state like ~/.codex/AGENTS.md, custom skills, MCP wrappers, and selected repo .codex/ templates from this repo on install/update.

git clone https://github.com/verkyyi/always-on-claude.git
cd always-on-claude
claude

Inside the Claude Code session:

/provision

Claude walks you through the entire AWS setup — SSH keys, security groups, instance launch — and connects you in ~40 seconds.

ssh claude-dev

On first login, the workspace walks you through setup — git config, GitHub auth, preferred assistant auth, and cloning your first repo. After that, you'll see a workspace picker to launch Claude Code or Codex in any repo. Press t in the picker to toggle which assistant is the default for future launches.

Back in your local Claude Code session:

/destroy

All lifecycle operations run from inside a Claude Code session in this repo:

The lifecycle control plane is still Claude-based for now. Codex is supported inside the provisioned workspace for repo work, onboarding, and SSH-launched coding sessions.

Your Mac / Phone / Tablet
    │
    └── SSH
         └── Ubuntu 24.04 (EC2 t4g.small, 20GB)
              ├── Docker container (claude-dev)
              │    ├── Claude Code + Codex
              │    ├── Node.js 22, Bun, npm
              │    ├── Git, GitHub CLI, AWS CLI
              │    └── Your project repos
              ├── tmux (session persistence)
              ├── host scheduling bridge (container → atd)
              └── Workspace picker on SSH connect

Everything persists — auth, settings, repos, tmux sessions, Claude/Codex state — all survive container restarts and reconnects.

Stop the instance when not in use to save money. No additional fees — you bring your own Claude or OpenAI credentials.

Destroy has a script fallback for automation:

bash <(curl -fsSL https://raw.githubusercontent.com/verkyyi/always-on-claude/main/scripts/deploy/destroy.sh)

For details, see Deployment.

• Architecture — Docker, networking, sessions, resource limits, update pipeline
• Deployment — provisioning, install, config, AMI, portable mode, Tailscale
• Operations — onboarding, sessions, worktrees, updates, mobile, slash commands
• CI/CD — all GitHub Actions workflows

MIT