Title: Running Claude Code, Codex, Grok and Ollama side by side on one machine
If you use more than one coding agent, you probably have a few terminals open, each with its own session and no idea what the others are doing. An hour in, you're piecing together which one did what.
I built JunctionBox AI for that situation. It's a freemium app that runs your agents side by side on your own computer, on Linux, macOS or Windows, and it doesn't need an account. This post goes through how it handles the agents and what I put around them, and there's a download link at the bottom so you can try it out.
The agents
You can create agents that run on the Claude Code, Codex or Grok CLI tools you're already signed in to, so those need no separate API key and no API billing. API keys for Anthropic, OpenAI and xAI work too, and so do local or remote Ollama and llama.cpp servers.
One team can hold a Claude agent, a Codex agent, a Grok agent and a local model. You can also send the same question to several of them and compare the answers side by side, which is a quick way to see which one suits which job.
Teams, memory, and schedules
Agents can be grouped into teams and hand work to one another, within boundaries you set. That saves you from copying output from one window into another.
Memory is optional. Identities, profiles and approved memories give an agent context that lasts past a single session, so you aren't explaining the project again every time.
Schedules run an agent on a timer, from a webhook, or from a Telegram message. Jobs keep going when you close the browser and survive a restart. Run history, retries and token use sit in one place, which matters for anything that ran while you were away, because you can see afterward what it actually did.
The sandbox
On Linux and macOS the agents run in a sandbox, and each one gets a permission profile: restricted, workspace or unrestricted. Unrestricted is there for the times you want an agent to have full access. Privileged operations can be reviewed before they run, and there's an audit history of what happened.
Windows has no equivalent sandbox yet, so agents there run with your own account's rights. The settings page says so next to the permission profile.
Your other machines
Agents can run on your other Linux, macOS and Windows machines too. You can open terminals on them and share their screens from the same place.
A local model doesn't have to run on the computer you're sitting at, either. If you already have an Ollama or llama.cpp server somewhere, point JunctionBox at it. Or put the models on your stronger machine and use them from the one you're on.
The TV side
On Roku, Google TV and Fire TV there's a launcher that replaces the home screen with your streaming apps, the weather, your lights and your Ring cameras. This entire piece of the project started because I was tired of seeing Roku Advertising on it's home screen. It also puts your agents on the big screen. From the couch you can open a Claude Code, Codex or Grok terminal on your computer, open a shell or screen sharing on your other machines, chat with your agents, watch them work and approve what they ask. The terminals, shells and agent chat sit behind a PIN you set.
The rest of the TV side is a daily Morning Briefing, sports game schedules for your favorite teams, reminders, a picture frame, music, books with read-aloud, and an arcade of four games you play with the ordinary remote.
Most of everything else is plugins. There are 25 plus plugins at this point and I have been adding new ones pretty frequently. You can install them from a zip file, covering things like home control (smart plugs, smart bulbs and A/C), video, comic and book production, iMessage and Slack clients, and you can write your own plugins with a plugin builder. There's no marketplace and no review step.
Tradeoffs and what's still rough
It's closed source but I have made it freemium, meaning that JunctionBox AI Core is free to use with no signup. The free core allows two active agents at a time, and JunctionBox AI Pro is $49 a year for unlimited agents and full access to all plugins.
The agents use the resources of whatever machine they run on, so several agents plus a local model on a small laptop will feel like exactly that. You don't have to keep it all in one place, though. Run agents on your other computers, and run the models on a stronger machine or a server you already have. A local model is only as quick as the hardware it runs on.
For the Smart TV piece I tested the most on Roku, and less on Google TV and Fire TV, so I'd like to hear how those behave on real devices. The project is young and has rough edges, but I would love to get some feedback on the project. From the console in JunctionBox or from the website https://junctionboxai.com there is a feedback form for bug reports or feature requests.
Trying it
Builds for Linux, macOS and Windows are here:
https://github.com/junctionboxai/junctionbox-releases/releases/latest
Top comments (0)