Welcome back to the show. We've got a really meaty listener question today, and I want to frame it carefully because the prompt is doing two jobs at once. The listener is asking, basically, how do you take an agentic CLI like Claude Code, Codex, whatever your poison is, and put it on a server in a way that you can actually use it from anywhere, including from your phone, and not just by VNC-ing into a desktop and squinting at a tiny screen.
Right, and there are two distinct use cases buried in there that I think we should separate out before we even start naming tools. Use case one is sysadmin. The listener mentions running Claude Code locally for system administration, which means it's bound to that specific workstation because that's where the stuff it's administering lives. Use case two is the more conventional thing, which is using an agentic CLI to work on code projects, GitHub repos, web apps, whatever, where the actual work could happen anywhere with a checkout and a network connection.
Yes. And those two have very different answers, so let's not pretend they have the same one.
Exactly. The sysadmin case is fundamentally a remote access problem dressed up in agent clothing. If Claude Code is administering your home workstation, then the agent has to live on that workstation. You can't move it. So the question becomes, how do you talk to that workstation from your phone in a Tel Aviv cafe. And the listener already named the answer in passing, which is a Cloudflare tunnel, or its cousins.
Or Tailscale. Which I'd argue is the better default for this.
Agreed. Tailscale gives you a tailnet, your devices identify each other by Tailscale identity, you don't open any ports on your home router, and you get end-to-end WireGuard between your phone and your desktop. Then on top of that you SSH in, and on top of SSH you run tmux, and inside tmux you run Claude Code. That stack, Tailscale plus SSH plus tmux, is genuinely the boring correct answer for case one.
Add mosh to that if you're on flaky mobile data. Mosh is a small thing that solves a huge problem, which is that regular SSH falls over when your phone switches from wifi to cellular or you go through a tunnel. Mosh keeps your session alive across IP changes and predictively echoes keystrokes, so it actually feels usable on a train.
And then for the actual on-phone client, the two real options are Blink Shell on iOS and Termius on iOS or Android. Blink is the one with first-class mosh and Tailscale integration, it's basically the iPad power user terminal. Termius is more polished, more cross-platform, has a nicer key management UI, but it's freemium and the agentic-CLI-friendly features sit behind the paid tier.
Now, and this is the part the listener was poking at, all of that is, in their words, dressed up SSH. You're still in a terminal. You're still typing on a phone keyboard. Claude Code's TUI is gorgeous on a thirty-two inch monitor and a bit cramped on a six-inch screen. So is there anything that's actually mobile-native, that treats the agent as the primary citizen and gives you a real mobile UI on top?
This is where it gets interesting because Anthropic themselves shipped something for this in February 2026. It's called Remote Control. The model is, your Claude Code session keeps running on your local machine, but the iOS app, the Android app, and claude.ai/code can all attach to that session. The traffic is brokered through the Anthropic API over TLS. So you're not opening a port, you're not running a tunnel, you just authenticate on your phone, pick a session, and you're in.
And critically that's not just SSH-with-lipstick. The mobile app actually understands turns, plans, tool calls, file diffs. You can read what Claude proposed, hit approve, send a follow-up, all in a real touch UI.
That's the good part. The honest caveats. One, it's still a research preview as of when we're recording. Two, it requires a Claude Max subscription, which is in the hundred to two hundred dollar a month band, with Pro access promised but not fully there. Three, and this is the big one for the listener's question, the session still runs on the local machine. It is not Claude Code on a server. It is Claude Code on your desktop, viewed from your phone.
Which does cover the sysadmin case beautifully, because the whole point of sysadmin Claude is that it has to be on the box being administered. But it doesn't help if your goal is, I want to spawn a brand new Claude instance from a beach in Eilat, with no laptop on, that can clone a repo and start working.
For that you need the agent to actually live on a server. So now we're in case two, the dev work case. Let's walk the categories.
Category one. Hosted dev environments with Claude Code installed. GitHub Codespaces, Gitpod, Coder. You spin up a container in the cloud, your repo is checked out, Node and Python and your tools are pre-installed, Claude Code runs inside it, and you access it through a browser. The Anthropic engineering blog actually wrote a piece about why their own engineers use Coder for this. Auth is whatever your dev platform gives you, usually GitHub OAuth or Google SSO.
That's a respectable answer. The mobile UX is meh, because you're still talking to a code-server style VS Code in a phone browser, which is not great, but it's not pretending to be something it isn't. The agent is genuinely on a server, you can close your laptop, the job keeps running, and you can come back to it from any device.
Category two. Sculptor from Imbue. This one is interesting because it's the only thing in the list that was built from the ground up to be a UI for parallel coding agents. Each agent runs in its own Docker container, the containers are pre-baked so startup is seconds rather than minutes, and there's a pairing mode that bidirectionally syncs the agent's container with your local IDE. It supports Claude Code and Codex.
My honest take on Sculptor is that it's the right shape for the desktop power user who wants to run five Claudes in parallel without them stepping on each other. It is not yet the answer for, I am on my phone in line at the post office. The UI is a desktop app. The mobile story is not really there.
Category three. Wrappers and mobile-first frontends specifically for Claude Code. There are a few of these now. Happy is one, AgentsRoom is another, Nimbalyst is another. The pattern is, you install a daemon on your server or workstation, the daemon registers with the wrapper's cloud, and a phone app talks to that cloud. You get push notifications when Claude finishes a turn, kanban boards for parallel agents, visual diffs, all the touch-friendly stuff.
Verdict on these. They are the most genuinely mobile-native options. They are also the riskiest from a trust standpoint. You're handing a third-party startup a persistent connection into your dev environment. Read their security model carefully, look at where the broker servers live, look at whether they store transcripts. For a personal side project, fine. For client code under an NDA, think harder.
Category four. Cloudflare Access in front of code-server. This is the homelab approach. You run code-server, which is VS Code in a browser, on a VPS or your home server, you put Cloudflare Access in front of it for OAuth, and you install Claude Code inside that VS Code instance's terminal. You get a browser-accessible IDE with an agent in it, gated by a real identity provider.
Auth model on that one is actually very good. Cloudflare Access lets you bind to Google, GitHub, Okta, whatever, with device posture checks and short-lived tokens. The downside is the same as Codespaces, which is that the mobile UX is a desktop IDE in a phone browser. Functional, but not joyful.
Category five, which the listener mentioned. Sandboxed VPS web terminals. Things like Wetty, ttyd, GoTTY, all behind some auth proxy. Honestly, I'd skip these. They're fine as building blocks but you're basically reinventing what code-server already does, with worse ergonomics.
Category six, and this is the one we should call out as a serious contender. Anthropic's own Routines and Managed Agents, which they announced earlier this month. Routines are scheduled or event-triggered Claude Code jobs that run from the Claude web app rather than from your local terminal. Managed Agents is Anthropic's hosted runtime for long-horizon agent work. Both of those are genuinely the agent on Anthropic's infrastructure, not on your laptop.
The catch with Managed Agents and Routines is that they are not yet a general-purpose, sit-down-and-iterate-with-it experience. They are more about, here is a defined task, run it on a schedule, report back. Which is great for some workflows and not what the listener is describing if they want to spawn a Claude and have a back and forth conversation with it from a coffee shop.
So let me try to land this. If I were Daniel today, advising someone with the listener's two use cases, I would split it.
For the sysadmin case. Tailscale plus SSH plus tmux plus mosh, with Blink or Termius on the phone. Add Anthropic Remote Control on top if you have a Max subscription, because it gives you the touch UI for the most common interactions and you can drop down to the SSH session for anything weird. That stack is, by a wide margin, the most mature, the most secure, and the cheapest. The auth model is your Tailscale identity plus SSH keys, which is hard to beat.
For the conventional dev work case. Codespaces or Coder or a Cloudflare-Access-fronted code-server, depending on your preference and your wallet. Run Claude Code inside that. For the mobile experience specifically, layer one of the third-party Claude Code mobile clients on top, with eyes open about the trust model. Or, if you trust Anthropic's first-party path, just use Remote Control pointing at a long-lived session running on a small VPS where you've checked out your repos.
And what's missing. What I think the listener is really fishing for and not finding.
What's missing is a hosted, multi-tenant, mobile-first agentic CLI service where you log in, you tap a plus button, a fresh ephemeral container spins up with your repos and your secrets and your dotfiles, Claude Code is already running, and you talk to it through a real touch UI. No SSH, no tmux, no daemon on a home server. That product does not really exist yet. Sculptor is the closest in spirit but is desktop-first. Codespaces plus Claude Code is the closest in plumbing but the mobile UI is bad. Managed Agents is the closest in cloud-execution model but the interaction loop is wrong.
Which means the honest answer is, in 2026, you still pick two of three. Server-resident, mobile-native, agentic. Pick two.
That's a fair summary. And my prediction is that gap closes within the year, because every single one of these vendors is racing toward the same target, and the path is obvious.
Before we wrap, let me hit a few specific things the listener asked about that we glossed. Devbox-style hosted environments. There's a whole category of these now, things like Daytona, DevPod, Coder workspaces, Replit's agent runtime. The pattern is the same in each case. A control plane that provisions a containerized or VM-based dev environment on demand, with your toolchain pre-installed, keyed to your identity. The differences are at the margins. Daytona and DevPod are open source and self-hostable, which matters if you don't want a third party hosting your code. Replit is the most polished consumer experience but is also the most opinionated about how you work. Coder is the enterprise-grade pick if you actually need to run this for a team.
And the auth question across all of them is roughly the same shape. SSO via your identity provider, short-lived workspace tokens, and either an HTTPS endpoint or an SSH endpoint into the workspace. Where they differ from running Claude Code on your home box is that the workspace is ephemeral, which is a feature when you want isolation and a bug when you wanted that long-running tmux session that's been computing something for three hours.
code-server specifically deserves a clean call-out. It is open source, self-hostable, and surprisingly mature. If you want the cheapest possible answer to, I want VS Code in a browser on a VPS with Claude Code in the terminal, code-server is it. Stick it behind Cloudflare Access or Tailscale Funnel and you're done. The honest knock is that the mobile experience is a desktop IDE in a phone browser, which we already said.
Mosh deserves one more sentence too because the listener mentioned it. Mosh is not a full alternative to SSH. It rides on top of SSH for the initial handshake and key exchange, then switches to its own UDP protocol for the actual session. It is purpose-built for exactly the mobile case, where your IP changes, your latency varies, and your connection drops. If you're going to do anything serious with an agentic CLI from a phone, mosh is not optional, it's table stakes.
One closing nuance for the listener. Whatever you pick, do not put your agent behind plain port-forwarded SSH on a public IP. Use Tailscale, use Cloudflare Access, use the vendor's brokered tunnel, but do not leave a Claude Code process listening on the open internet behind a password. The blast radius if that gets popped is your entire dev environment plus whatever credentials Claude has access to.
Agreed. Identity-based access, short-lived tokens, and a tunnel rather than an open port. That is the non-negotiable floor for any of this.
Good answer. Thanks for the prompt, listener. We will see you next time.
Take care.
