You kick off a coding agent in your terminal, then wander off to make coffee. When you come back, the screen has scrolled past a wall of output and you have no idea what it actually did. Did it edit the right file? Did it run something you would not have signed off on? The terminal is a firehose, and firehoses are hard to supervise.
Companion is an answer to that problem. It is a free, open-source web and mobile interface that sits in front of Claude Code and Codex, the command-line coding agents from Anthropic and OpenAI. Instead of squinting at a scrolling terminal, you drive the whole thing from a browser tab, or from your phone.
Quick facts
- License: MIT (you can use it, fork it, and ship it commercially, for free)
- Primary language: TypeScript
- Last active: actively maintained, with a v0.95.0 release in April 2026 and around 2.3k stars (as of 4 Jul 2026)
What it is
Claude Code and Codex are agents that live in your terminal. You give them a task in plain English, and they write code, run commands, and edit files on your machine to get it done. Powerful, but the terminal is a cramped place to watch an agent work, especially when several are running at once.
Companion wraps those agents in a proper interface. It runs a small local server on your computer (localhost:3456) and gives you a chat window, a live timeline of everything the agent is doing, and buttons to allow or block sensitive actions. The clever bit is how it connects: the Claude Code CLI exposes an undocumented websocket path (--sdk-url), and Companion talks to it directly, streaming the agent's events into your browser as they happen.
The real-world problem it solves
Say you ask an agent to refactor a chunk of your app. It starts editing files, then decides it needs to delete an old config and run a database migration. In a raw terminal, that migration might fly past before you even register it. With Companion, execution pauses and asks: this command wants to run, allow it? You get to be the adult in the room for the actions that actually matter, without babysitting every harmless file read.
Where it's useful
- A freelance developer running three client projects at once launches a separate session per project and watches all of them from one dashboard, instead of flipping between three terminal windows and losing track of which is which.
- A team lead reviewing an agent's work uses the tool timeline to see exactly which files were touched and which commands ran, so a risky change gets caught before it lands rather than during the next incident.
- Someone away from their desk keeps a long-running task moving from their phone, approving the occasional permission prompt from the couch instead of racing back to the laptop.
- A cautious first-timer who does not fully trust an AI on their machine yet turns on approvals for anything destructive, so the agent cannot delete or overwrite without a human tap first.
Why it stands out
The honest headline feature is visibility. Companion does not just show you the agent's final answer. It shows the whole execution flow: the streaming response, each tool the agent reaches for, the exact parameters it passed, and what came back. When an agent does something surprising, you can usually see why.
Two other things are worth calling out. It supports both Claude Code and Codex, so you are not locked into one vendor's agent. And it survives restarts: if the server or the process goes down mid-task, you can restore the session rather than starting over. There is also a background service mode, so the server keeps running after a reboot and you just open the tab whenever you need it.
One caveat worth being upfront about: this leans on a reverse-engineered, undocumented protocol. That is genuinely impressive engineering, but undocumented paths can change under you when the underlying CLI updates. Handy to keep in mind if something breaks after a Claude Code upgrade.
How to get started
You need Bun and at least one of the agents (Claude Code or Codex) installed. Then the fastest path is a single command that runs it without installing anything permanently:
bunx the-companion
Open http://localhost:3456 and you are in. If you want it always available, install it globally and register it as a background service so it starts on boot:
bun install -g the-companion
the-companion install
the-companion start
On first run it generates an auth token and stores it at ~/.companion/auth.json, so the interface is not wide open to anything on your network. There is a full docs site and a Docker image too if you would rather run it that way.
Takeaway
If you have started handing real work to Claude Code or Codex but feel uneasy about what they do when you are not looking, Companion is a low-cost way to get your eyes back on the process. It is free, it is MIT-licensed, and it turns a scrolling terminal into something you can actually supervise. Worth a bunx on a spare afternoon.