Subshell vs. Alternatives

The real question in front of you is not "which product"; it is how much machinery you want around a coding agent CLI. There are two honest answers that aren't Subshell: you already have tmux and SSH, or you let a hosted product run the agent for you. This page treats both fairly, because both are good at what they are for.

Running it yourself: tmux and SSH

The incumbent, and not really a competitor: Subshell is tmux underneath. So the comparison is about the layer around it.

What it does better: durability is already there; that is tmux's entire job. SSH reaches the pane from anywhere you have a terminal. There is nothing between you and the CLI (no server to run, no data model, no update channel), and the agent runs with your machine exactly as you left it: every credential, dotfile, and half-configured dev database in place. For one person on one machine who is mostly at that machine, it is complete, and adding software to it needs justification.

What it costs: the pane is durable but invisible. Nothing tells you the agent has a question: you find out when you next look, which is how the SSH-from-dinner evening starts. From a phone you are inside a terminal emulator squinting at 80 columns, because the terminal is the whole interface. Involving another person means sharing the machine or the account, not a link. Agents can't see each other: five panes are five unrelated processes, and coordinating them is your calendar. Machine number two means your ssh config and your memory, and every morning's launch is the same typed incantation.

What Subshell adds on top of the same tmux: a browser client on every device, showing the real pane as a live terminal you can type into, with scrollback replay on attach. Push notifications when a pane needs attention. Per-pane credentials and an MCP server so agents message each other over encrypted channels and can launch and steer their siblings. Sharing by grant (read-only or interactive) instead of by machine access. Nodes so many machines sit under one control plane, and presets so the launch you type every morning is chosen from a list.

The honest cost of Subshell here: you now run a server. It is one binary and it asks for little, but it is a piece of infrastructure with its own trust story, where bare tmux has none.

Hosted agent products

Claude Code's hosted web agents, OpenAI's Codex cloud, Devin-style services: the vendor runs the agent on a machine they built, against your repository, and you watch results in their UI.

What they do better: zero infrastructure, in both directions. You sign up and it works: nothing to install, patch, back up, or keep reachable from your phone. The environment is fresh and isolated per task, which is a feature when you want an agent to poke at a repo without living inside it. If your honest preference is "I do not want to own any machine for this," these products answer that directly and Subshell does not: there is no Subshell cloud to sign up for.

The axis they all share: the agent works where they run, not where you do. Your repository connection, your credentials, and the agent's working environment live inside their boundary, and what the agent sees is the snapshot or clone they set up, not your build box with its local services, your internal-only codebase that cannot leave the network, or the machine your work actually depends on. When the product's flow doesn't fit, there is no terminal to type into; you stay in the product. And the meter runs on their cloud (seats and usage) whether or not any machine of yours is involved.

Subshell is the mirror image on every one of those points: the pane runs on your machine (the laptop, the mini in the office, the GPU box), with the CLI's own credentials staying on that machine; it is the real CLI in a real pane, so every escape hatch you know still works, including talking to the agent directly at any moment; there is no per-seat charge and no service between you and the process, because you are the service. What it costs is the flip side: you own the machine, the uptime, and the network the whole thing rides.

Who each is for

• tmux and SSH by hand: one person, one machine, and you are usually at it. Subshell earns its keep the first time you leave a task running and want it from elsewhere, or want two people watching without sharing a password.
• Hosted agent products: you want work done without owning infrastructure, the code can live on their platform, and a fresh clone is the right starting point more often than your actual machine.
• Subshell: the machine matters. Internal code that stays on your network, dev environments with real state, hardware the agent needs. You want the agent working after the laptop closes, the whole fleet under one control plane, and (from the sofa, on any device) the same terminal you'd see at the desk.

See also

• What is Subshell: the feature set in one page
• Security Model: what self-hosting actually commits you to
• Get Started: the shortest path to a running subshell

Edit on GitHub