Claude Code web UI alternatives: I tried seven, then built the eighth.

The official answer, and the first one to try, because if you are on Pro or Max it is already included. You connect GitHub repositories at claude.ai, describe the job, and Anthropic runs it on their own infrastructure. Each task gets an isolated sandbox with network and filesystem restrictions, git goes through a proxy that can only reach the repos you authorised, and you can fire several tasks at once across different repos. Each one comes back as a branch and a pull request with a summary of what changed. It is in the iOS app too, as part of the same research preview.

For a well defined, sandboxable change this is the best tool on the page. "Add validation to this endpoint, write the test, open a PR" is exactly the shape it wants, and delegating three of those while you do something else is a genuinely good afternoon.

Where it stops is where your machine starts. The sandbox has no route to your local Postgres, your .env, the MCP servers you spent an evening configuring, or the half finished branch that has never been pushed anywhere. It is GitHub only. And it shows you the tasks you started there, which means the run currently going in your terminal does not exist as far as it is concerned. It is a delegation product. It was never trying to be a window onto your own computer.

Not a web UI, and I am including it anyway, because a good share of the people searching for one actually want this. The VS Code extension and the JetBrains plugin are not stripped down wrappers: they drive the same engine as the CLI, share the same conversation history, respect the same CLAUDE.md, and use the same permission modes. What they add is everything a terminal is bad at, which is mostly reading. Diffs render in the editor's own diff viewer. You can @-mention a file with a line range. Your language server's diagnostics get handed to Claude without you copying anything. Because Cursor is a VS Code fork, the extension loads there too, as it does in the other forks.

If your complaint about Claude Code is "I cannot read these diffs in a terminal", install the extension and stop reading this post. You are finished in two minutes and it costs nothing.

Where it stops: it is a panel in an editor on one computer. Close the lid and the run is not somewhere else, it is over there. No phone, no fleet, nothing to glance at from the sofa, and no answer at all for the machine under your desk that does the heavy jobs.

A Mac app for running a team of agents at once, each in its own git worktree. You point it at a repo, it clones locally, and every task gets a fresh isolated checkout in about ten seconds with its own branch, terminal, diff and review path. It drives Claude Code, Codex, Cursor and OpenCode, connects to GitHub and Linear, and the agent can open pull requests and answer review comments. It is free and you bring your own subscription. Engineers at Linear, Vercel, Notion and Stripe use it, which is the kind of sentence I normally distrust and in this case seems to be plainly true.

If the thing costing you time is five agents trampling each other in one checkout, Conductor is the sharpest tool here. Worktrees are the correct answer to that problem and it does them in one click instead of the four commands I usually fumble.

Where it stops: macOS only, and it is a desktop app. Everything happens on the Mac in front of you, so a Linux box under the desk is not part of the picture and neither is a train. It is an orchestrator for one machine, not a control plane for your machines.

The board shaped one. Vibe Kanban puts your agents on a kanban board with To Do, In Progress, Review and Done, and supports well over ten of them: Claude Code, Codex, Gemini CLI, GitHub Copilot, Amp, Cursor, OpenCode, Droid, Qwen. The MCP server is the clever part, because a planning ticket can ask an agent to decompose the work and generate the downstream cards itself. It runs locally and you drive it from a browser tab, so of everything on this list it is the closest to a web UI that is not trying to be a chat window.

Worth knowing the history before you build a workflow on it: Bloop, the company behind it, announced a shutdown in April 2026. The project is Apache-2.0 and community maintained since, the hosted features were removed, and the architecture is now fully local. That is a real answer rather than a red flag, but you should hear it from a blog post and not from a 404.

Where it stops: it is one machine's local server, and the unit of work is a ticket rather than a conversation. Reaching it from a phone is a tunnel you build yourself, and nothing pages you when a card stalls on an approval.

The one most people mean when they type this search. siteboon's project, somewhere north of eight thousand GitHub stars, reads ~/.claude/projects, the same directory the CLI writes its sessions into, and serves them through a React front end over a WebSocket. That one design decision is why it can show you a session you started in a terminal, and it is the decision I respect most in this whole roundup. There is a file tree with syntax highlighting, a built in shell, a layout that survives a phone, and it drives Codex, Cursor CLI and OpenCode as well. Free, self hosted, with a hosted version if you would rather not.

Where it stops is the part that never makes it into a README: self hosting means you own the exposure. Getting to your laptop from a phone means a tunnel, a certificate and an authentication story, and "expose the thing that can run arbitrary shell commands on my machine to the open internet" is a sentence worth finishing slowly. Beyond that it is one machine at a time, and there are no push notifications, so you still have to be the one who goes and looks.

The privacy answer, and the best pure phone experience here. You start sessions with their happy wrapper instead of claude, and the session syncs to a real iOS and Android app and a web client. Everything is end to end encrypted with TweetNaCl, the same primitives Signal is built on, so the relay forwards ciphertext and cannot read your code even in principle. MIT licensed, free, and there is realtime voice, which sounds like a gimmick until you use it on a walk.

Where it stops: the wrapper is the catch. Sessions have to be started through happy, so the plain claude you fired in a terminal an hour ago is not in the app. And an app is not a fleet board. It is one conversation at a time, on a phone.

This is what I ran for a year, and it is a completely respectable answer. tmux keeps the session alive when the lid closes. Tailscale puts your machines on one private network without you touching a router. ttyd or code-server puts a terminal in a browser tab. Total cost nothing, total vendor risk nothing, and every piece is software that will outlive all of us.

Where it stops, and I say this with love, is everywhere. tmux has no notion of a notification. It will hold a finished run in a detached session for six hours and never say a word. A permission prompt on a phone sized terminal is eighteen columns of wrapped text. A diff is ransom mail. And you now maintain a small piece of infrastructure whose worst day is you debugging your own VPN config on a Sunday instead of the bug you sat down to fix.

I can date the exact minute this stopped being funny. I started a refactor on the Linux box, walked away for an hour, and came back to find the agent had spent fifty eight of those sixty minutes sitting on a permission prompt, waiting for a keystroke I was never going to see, because the window was behind a browser on a different machine.

So, the conflict of interest. Termdeck is the browser and phone front end for the CLIs you already have, Claude Code, Codex and Grok, running on the machines you already own. It does not host a model, it does not want your API key, and it never proxies your code anywhere.

The decision that matters is the same one CloudCLI made, and I would defend it to the death: the CLI's own session file on disk is the source of truth, and Termdeck renders that file. So a chat you started in a terminal appears in the browser (there is a group on the board called, literally, "open in a terminal"), a chat you started in the browser reopens in your terminal with claude --resume, and the second another process starts driving a session Termdeck locks that composer to view only, because I once typed into both ends of the same conversation and forked its history in half, which is as unpleasant to untangle as it sounds.

The rest is the fleet. Every machine on one board, sessions grouped by whether they need you. Push notifications fired from the machine itself, so an approval reaches your phone with no browser open anywhere and you can allow, deny or take the run over from wherever you are standing. Diffs docked next to the conversation that produced them. A usage ledger across every account and model. One click switching between two Claude subscriptions when the first one hits its limit, mid chat, without losing context. The agent dials out over an ordinary WebSocket, so there is no inbound port, no forwarding rule and no VPN, which is the whole reason I stopped maintaining one.

Honest limits, since I am the one keeping score. It is not an IDE and does not want to be; if you want to write code by hand, go and write it by hand. Every machine needs a small agent installed, which is one command but it is still one command. And the free plan is one connected machine, one session at a time and 30 browser turns a month, which is plenty for watching and approving and not enough to work from the browser all day. Paid starts at $9.99. Every new account gets 14 days of Pro first, with no card.

If the job is a clean GitHub task, use Anthropic's own Claude Code on the web. It is included in your plan, it runs somewhere safe, and it hands you a pull request. Nothing on this list beats it at that.

If you are always at your desk, install the VS Code or JetBrains extension. Two minutes, free, and it fixes the only real complaint most people have about the CLI.

If you run five agents on one Mac, use Conductor. Worktrees per task is the right idea and it is the best implementation of it.

If you want the source and one machine is enough, self host Claude Code UI. It reads the CLI's own session files, which is the property that makes a browser UI trustworthy.

If you only care about the phone, use Happy Coder. End to end encrypted, free, and the app is genuinely nice.

If you have a desktop, a Linux box and a work laptop, and the recurring cost in your week is agents quietly waiting on you while you are in a different room, that is the thing I built Termdeck for, and I did not build it because I spotted a market. I built it because I lost that fight badly enough to write software about it.