Coding agents, in workspaces of their own.
Describe a task and smedja launches a real machine to do it in — checked out, set up and running — then keeps it for as long as the work does. Two jobs never share a working tree.
In early access: sign-in is by invitation for now.
A workspace per task
Its own checkout, its own disk, its own shell. Pause it and the disk keeps; resume it and the working tree is exactly where the agent left it.
Your machine or ours
A managed runtime, or a launcher you run on a computer you own. Same workspace either way — the choice is one flag at launch.
Everything has a CLI
Anything the app can do, smedja can do — including from inside a workspace, where an agent can open the next one. Your own agents get the same operations over MCP.
Four moves, and you are watching work happen.
Connect a repository
Sign in with GitHub, then connect a repository account. Signing in alone grants no access to your code.
Add an agent
Claude Code, Codex or Gemini CLI, on your own credential — checked with the provider before it is stored.
Describe the task
Pick a template and say what you want. The message waits until the machine is up, so you can close the tab.
Pause it, park it
Pause to give back the compute and keep the disk. Archived work stays searchable, and comes back out when you need it.
Built for teams that share repositories, not machines
An organisation is the people; a project is where the work is. Everyone in a project can drive every workspace in it, admins decide who is in the room, and an action you cannot take is simply not on screen.
Colour here means status and nothing else.
Want in?
smedja is in early access. Tell us what you would run on it and we will open the door.