You run agents all day.
“Five sessions, three projects, and I keep switching windows to see who’s stuck.”
See every agent at once, keep the machine fast, and let the guards stop the bad commands.
Webmux runs Claude Code, Codex, Grok Build and Antigravity side by side on your own machine — sharing the same tools, guarded against bad commands, and on your phone when you leave the desk.
Free plan available. Your first 7 days include Pro, no card. Keep your own agent subscription.
macOS 13+, Linux, and Windows through WSL2.

Morning: review the agents’ completed work from your phone.
Who it’s for
Webmux is for every developer who works with coding agents, alone or in a team. Pick the line that sounds like you.
“Five sessions, three projects, and I keep switching windows to see who’s stuck.”
See every agent at once, keep the machine fast, and let the guards stop the bad commands.
“My agent asked a question at six. I saw it at nine the next morning.”
Answer it from your phone, on the train, at the gym or on the couch.
“The provider limit hits in the middle of the task.”
Fewer tokens per command, the habits that cost the most, and a plan the agent shows you before it builds.
“Every developer sets up their agents a different way.”
One setup committed with your code, the same guards on every machine, seats in one pool.
“A day of WSL, drivers and paths before the first agent runs.”
Install Webmux in WSL2 with one command today.
Work from anywhere
Your agents run on your own machine, not in a browser tab. When one stops on a question at six, you answer it from your phone on the train — not at nine the next morning.
How
A real terminal on your phone, with the same screens as at your desk. Remote access goes through your tailnet, your local network or the Webmux Relay, end-to-end encrypted.
Drawers that slide in from the edge, a key row you can edit — Esc, Tab, arrows, Ctrl+C — and an input bar for long prompts.
And
Every agent, the same tools
Add an MCP server, a skill or a guard once, and every agent on your machine has it. Claude down or out of quota? Switch the pane to Codex and keep going with the same tools.
How
One row per server or skill, one box per agent. Tick a box to give it to an agent that lacks it.
Installed for each agent from one page. Your own hooks and a rule you edited are never overwritten.
Every agent’s instruction files checked: one real AGENTS.md per folder, a file too large for an agent and a dead link flagged.
And
Mistakes undone
A wrong rm , a reset over uncommitted work, a crash mid-task. Webmux refuses the dangerous commands before they run, and brings any conversation back.
How
rm, unlink, rmdir and find -delete are refused, even through sh -c or xargs. Your agent is handed webmux trash instead, so you can restore what it removed. Claude Code and Codex.
reset --hard, clean, stash drop, a forced switch and branch -D are refused before they run. Claude Code and Codex.
« Resurrect » restarts it on the same session; its last screen stays on display meanwhile.
One search — Cmd+P — across pages, projects, panes, settings and your Claude and Codex conversations.
A new agent can work in its own worktree; « Apply » brings its changes back into your working copy.
And
Several agents at once
Put agents side by side, across projects, with the app you are testing and the plan they follow in the same window. You steer; they work.
How
Split a screen into panes: Claude Code, Codex, Grok Build and Antigravity next to each other, each with its status.
Every pane laid out at its real size, filtered by type, live.
A browser with Chrome DevTools in a pane, and your agent drives it: clicks, typing, console, network.
Files read, edits, commands, git and pages on one timeline. Nothing runs again.
The plan open beside the agent, the step in progress marked.
Where a name is used and what a change reaches, answered for every agent by one shared language server per project.
And
This interactive product example keeps the original English pane content.
More done per plan
Every token your agent reads counts toward its limit. Webmux trims what reaches it, and shows you which habits cost the most.
How
Searches, git reads, listings, tests and lints reach your agent summarised, on Claude Code and Codex. The full output stays on disk.
Switching models mid-thread, sending a message during an answer, reading a file twice: an alert says what it cost, and a tip comes back when the habit does.
Read and search without waste, test the edge cases and reproduce a bug before fixing it, report clearly: rules Webmux installs into Claude Code and Codex.
Sessions never compacted, effort set too high, a costlier model, files read twice: each one priced on your real week, with its fix.
And
A machine that stays fast
Agents leave dev servers, busy ports and temporary files behind. Webmux shows what is still running and what can be cleaned up.
How
Every dev server found, with its page title: stop it, or jump to its project.
What a closed pane left running is listed, heaviest first, with why nothing needs it.
Temporary terminal and agent uploads expire after 7 days; full command-filter outputs after 3. Your source project is not deleted.
And
Design before code
A wrong screen costs a redraw; a wrong build costs a rewrite. Your agent shows you the screen, the options and the data change first, so you correct a drawing, not a release.
How
Your agent draws the page; it opens in every tab you have, phone included.
Numbered options, keys 1 to 9, each one drawn across the whole screen.
Who acts, in what order, and where the time goes.
What leaves in red, what arrives in green.
A method your agent follows: a plan sized to the task, reviewed before it runs.
Your code stays home
Webmux runs where your code is. Remote access is end-to-end encrypted, and a local model handles what should never leave.
How
llama.cpp, with the tested models marked by what your machine can run; GGUF files from Hugging Face.
Whisper on your machine turns speech into text; tuned to your project’s own words.
And
Cloud coding agents still send the prompts you give them to their providers. Choose a local model for work that must stay on your machine.
One setup for the team
Keep shared MCP servers, skills and project rules in versioned files. Select a file to see what your teammates receive.
Claude Code reads this project’s shared server declarations after the developer approves the project.
Codex reads project configuration from this file. Webmux presents the supported settings beside those of the other agents.
A project skill packages the instructions and resources your teammates need for the same task.
The project’s hook declarations travel with its configuration. Personal settings remain separate.
Codex’s project hooks use its own file format. Webmux keeps the provider-specific settings visible.
Version project conventions so a teammate can review the same instructions with the code.
A project can carry Webmux Apps that open beside its terminals and agents. Running project content remains subject to trust.
Project configuration is versioned with your code. A teammate must trust the project before its configuration can run.
Any machine
Install Webmux in WSL2 with the same command as on macOS and Linux.
How
Ubuntu or Debian in WSL2, with the same install command as on macOS 13+ and Linux.
Webmux keeps your terminals and your agents running on your own machine, and opens them in any browser — your desk, your laptop, your phone.
$0
forever
Terminals, sessions and editor — free forever.
$5/mo
excluding tax · €6/mo including VAT in the eurozone
or $50/yr · €60/yr
Remote access to your machine.
$15/mo
excluding tax · €18/mo including VAT in the eurozone
or $150/yr · €180/yr
Run a fleet — more machines, more browsers.
from $5/seat/mo
excluding tax · from €6/seat/mo including VAT in the eurozone
PRO $5/seat · MAX $15/seat
Pool PRO & MAX seats for your team.
Every new account starts with a 7-day free trial. Annual billing includes 2 months free.
Your machine runs the workload. A subscription adds Webmux capabilities; it does not include credits for your coding agent or a hosted machine.
$25.00 per month
Anthropic and OpenAI each ship a desktop app for their own agent. Webmux runs both agents in one workspace, on your machine, and lets you reach it from anywhere.
Those apps are made for everyday use. Webmux is made for developers who run many agents at once, in as many windows as they need: it keeps the machine’s CPU and memory in view, and the quota of each agent account with the time it resets.
| What changes | Webmux | Claude Code desktop | Codex in ChatGPT desktop |
|---|---|---|---|
| Reach it, keep it running | |||
| Your agent keeps running when you close the window | SSH and cloud sessions¹ | Cloud tasks; local not documented² | |
| Your own machine from any browser, nothing to install | PRO | The agent only, through Remote Control¹ | Cloud sandboxes only² |
| Your own machine from your phone | Terminals and agentsPRO | Remote Control or Dispatch¹ | Codex chats, from a Mac or Windows host² |
| When the agent runs in the cloud instead of on your machine | |||
| Your personal skills, settings and MCP servers | All of them | Only what the repository commits¹ | Repository skills; personal ones stay local² |
| Your work in progress, committed or not | A fresh clone of the pushed branch¹ | A checkout of a branch or commit² | |
| Your local servers and the rest of your network | An allowlist of public hosts by default¹ | No internet while the agent works, by default² | |
| A test suite that runs for an hour or more | As long as your machine is on | 2 hours per background command; pauses when idle¹ | Maximum run time not documented² |
| The machine it runs on | Yours, every core and all its memory | 4 vCPUs, 16 GB of RAM, 30 GB of disk¹ | 2 to 4 vCPUs, 8 to 16 GiB, by plan² |
| Your agents | |||
| Claude Code, Codex, Grok Build and Antigravity in one workspace | Claude only³ | Codex only³ | |
| Skills and MCP servers shared between your agents | PRO | Reads AGENTS.md¹ | Imports Claude’s setup and keeps it in sync² |
| The agent drives the workspace: opens panes, shows pages | Through Webmux’s MCP serverPRO | Its sessions and Browser pane¹ | Its tasks and browser tabs² |
| One language server shared by all your agents | TypeScript and Rust, compact answersPRO | A plugin you install; sharing not documented¹ | Not documented |
| Command output cut to what matters before the model reads it | Claude Code and Codex, the whole output kept 3 daysPRO | Truncation only¹ | Truncation only² |
| Context and plan-usage meters | Per session and per account | Usage ring¹ | /status and a web dashboard² |
| Panes | |||
| Local Browser: your app beside the agent, DevTools for you | Chrome DevTools built inPRO | Browser pane; DevTools not documented¹ | Built-in browser; DevTools not documented² |
| Web Browser: the agent reads pages that turn bots away | Camoufox, with saved profilesPRO | Its browser or Chrome; stops at CAPTCHAs¹ | Its browser or an extension; anti-bot not documented² |
| Proto: pages your agent shows you, on every device | PRO | HTML in the Browser pane; artifacts¹ | HTML previews (rolling out)² |
| Tasks: agent turns, commands and your approvals, chained | On a canvasPRO | Scheduled tasks; workflows chain agents, no approval between steps¹ | Scheduled tasks; no step pipeline² |
| HTTP client | PRO | Not documented | Not documented |
| Files, editor and git | Monaco editor, git panel | File tree, spot edits, diff view¹ | File tree; stage, commit, push² |
| Several sessions side by side | Any layout of panes | Split panes and worktrees | Parallel chats and worktrees |
| Live CPU and memory of the machine | Not documented | Not documented | |
| On your machine | |||
| Voice input | Whisper, on your machinePRO | Dictation in the desktop app¹ | Dictation; Voice runs in the cloud² |
| Local models: chat | llama.cppPRO | Not documented on the device¹ | Local chat in the CLI (--oss)² |
| The browser’s traffic through your VPN | A WireGuard tunnel you pickMAX | Not documented | Cloud tasks only (Tailscale)² |
| The tools agents need, installed for you | Language servers, Camoufox, Whisper, the GitHub CLI… on one page | Offers to install gh¹ | Cloud setup it prepares; local scripts you write² |
| Platform and price | |||
| Runs on | macOS, Linux, Windows (WSL) | macOS, Windows, Linux (beta) | macOS, Windows, Linux (preview) |
| What it costs | FREE costs nothing; PRO from $5/mo excl. tax, €6/mo incl. VAT in the eurozone — plus your agent’s own plan | In Claude Pro ($20/mo) and up | In every ChatGPT plan, Free included |
PRO MAX Built in · text names a condition · Not documented: the vendor’s pages do not say · PRO / MAX the Webmux plan a row needs. Checked on 2026-10-03.
1 · Claude Code desktop. A local session goes offline when you quit the app; SSH and WSL sessions keep running (release notes, 2026-09-15) and cloud sessions run on Anthropic’s machines. Remote Control links claude.ai/code or the mobile app to a session on your computer — you steer the agent, not a shell — on Pro, Max, Team and Enterprise. Claude can list, message and archive your other sessions and open your dev server in the Browser pane; the Browser pane and Claude in Chrome stop at CAPTCHAs, and a DevTools inspector for you is not documented. Code intelligence is a plugin whose language server you install yourself; a server shared between sessions is not documented. Command output is truncated, not filtered — filtering is a hook you write. Local scheduled tasks fire while the app is open; cloud routines run on Anthropic’s machines; a workflow chains agents in one script, with no approval between its steps. How the Browser pane reaches the network is not documented. The app offers to install the GitHub CLI. A cloud session works on a fresh clone of your pushed GitHub branch, on a machine of about 4 vCPUs, 16 GB of RAM and 30 GB of disk; your ~/.claude instructions, personal skills and the MCP servers you added on your machine stay there, and only what the repository commits comes along. Its network is an allowlist of public hosts by default. A background command gets at most 2 hours, and the machine pauses after a few minutes without activity (cloud environments and tools reference, read 2026-10-08). Desktop · Cloud environments · Tools reference · Remote Control · Code intelligence · Scheduled tasks · Workflows · Memory · Release notes · Pricing
2 · Codex in the ChatGPT desktop app. The Codex app became part of the ChatGPT desktop app on 2026-07-09. Cloud tasks keep working while your computer sleeps; whether a local chat survives quitting the app is not documented. On the web, chatgpt.com runs cloud tasks; your own computer is reached from the ChatGPT mobile app or another Mac or Windows device, and remote access stops when the host closes the app. The app imports Claude Code’s setup and keeps it in sync. No language server is documented anywhere in the docs; tool output is cut to a token budget. Scheduled tasks run in a local project while the app runs; no pipeline of steps is documented. Local models are documented for the CLI (--oss); images and Voice run in the cloud. A VPN is a setting of cloud environments, and Codex prepares a cloud environment itself, installing its dependencies and tools; a local setup is a script you write. How the built-in browser meets anti-bot pages is not documented. The review pane stages, reverts, commits and pushes. A cloud task checks out a GitHub branch or commit; your computer’s local files, running processes and VPN access are not transferred, and personal skills are not synced. Internet access is blocked while the agent works, by default. The machine has 2 vCPUs and 8 GiB on Plus, 4 vCPUs and 16 GiB on Pro and Business; a maximum run time is not documented (cloud environments and internet access, read 2026-10-08). Changelog · Cloud environments · Internet access · Remote connections · Import · Scheduled tasks · Browser · All docs · Review · Pricing
3 · One vendor each. Each app’s terminal pane can start another agent’s CLI, but only its own agent gets the app’s sessions, review and remote access.
Webmux. Webmux’s own filter hands Claude Code and Codex a summary of the tests, searches and git reads they run unasked. The paid panes and tools carry their plan’s badge; the rest is in FREE. Webmux runs the agents you already pay for — a plan never includes model credits.
Questions
The questions developers ask first.
No. Webmux runs the Claude Code, Codex, Grok Build or Antigravity you already use, on your own plan. It includes no agent credits and adds no fee on tokens.
No. Webmux runs on your machine. Remote access goes through your tailnet, your local network, or the Webmux Relay, which forwards end-to-end encrypted data it cannot read.
No. Webmux is a small server you install beside your projects. Your agents, files and terminals stay on your machine, and any browser you approve opens them.
Claude Code, Codex, Grok Build and Antigravity, side by side. The guards and the shared rules cover Claude Code and Codex.
Yes, with no time limit: terminals, splits and the editor, your agents side by side, every account’s quota, and the machine in view. Every new account also gets 7 days of Pro, no card.
You move to the free plan on your own, and nothing is charged: no card was asked. The guards and the filter stay installed but pause, silently, and come back the moment you upgrade. Webmux tells you once, the day the trial ends.
Yes, in WSL2 (Ubuntu or Debian), with the same install command as on macOS 13+ and Linux.
Webmux writes the MCP servers, skills, guards and rules into files in your repository. Commit them, and every teammate who pulls gets the same tools.
Start on the free plan, with 7 days of Pro on top. No card.