Remote kernels
An agent can do their work on another machine while Albedo stays on yours. The kernel, which is where the agent runs commands, starts jobs, and reads and edits files, runs on that host over SSH. The conversation, memory, and everything else about the session stay with your daemon.
On this page: Start a session on another machine, What the host needs, What runs where, When the connection drops, and Visiting a machine without moving there.
Nothing is installed on the host by hand, and it does not need Albedo.
Start a session on another machine
- A new session: in the folder picker of the terminal or web client, type
host:/path, or@hostto browse that machine's folders. Hosts from your~/.ssh/configare offered as you type. - A session you already have: run
/cd host:/absolute/path. The terminal client asks "Copy memory and move the agent to this machine's daemon?". Answer no to run only the kernel there, which is what this page describes. Yes hands the whole session to an Albedo daemon running on that host.
A location is [user@]host:/absolute/path. host:~/project works too, and the
~ is resolved on the host. The host can be an alias from your SSH config.
Then talk to your agent the way you would on your own machine:
you're on the build box now, so the gpu is yours. can you run the benchmark there and tell me what you make of the numbers?
this checkout on staging is a bit behind. could you catch it up and see if the tests still pass over there?
What the host needs
- SSH that already works for you. Albedo connects the way your terminal does, with your config, keys, and agent.
- Python 3.11 or newer as
python3.
On the first visit Albedo copies its kernel to ~/.albedo-remote on the host,
and Albedo says it is copying while that happens. Later visits reuse it.
Albedo never answers a passphrase, a second factor, or a host key question for
you. When a host needs one, Albedo says the host needs you to sign in. The
terminal client opens that sign-in in your terminal, and anywhere else Albedo
gives you the ssh command to run. One sign-in is enough: Albedo keeps a single
connection per host open and reuses it, and when your SSH config names a
ControlPath, your own ssh and git share it.
What runs where
On the host:
- commands and background jobs
- reading, searching, and editing files
- the project's
AGENTS.mdand its skills - the changes and file listings the clients show you
On your machine:
- the conversation and the requests to your model provider
- memory and the journal
- the work ledger, agents, and mail
- your provider accounts
The agent is told which machine they are on, so paths in their answers are that machine's. Two things differ from a local session: you can attach images to a message but not other files, and worktrees cannot be created on a remote host.
When the connection drops
The kernel runs on the host, not inside the SSH connection. If the network drops, your laptop sleeps, or the daemon restarts, the kernel and its jobs keep going, and the agent picks them up again when the host is reachable. What the kernel had in memory is still there.
i'm closing the laptop for the train. keep the training run going on the gpu box and tell me how it went when i'm back?
Visiting a machine without moving there
An agent working on your machine can also reach out to another one for part of a task, without moving there. Ask for it in passing:
nginx on deploy@box keeps restarting and i can't work out why. can you hop on and have a look with me?
Use a remote kernel when the project itself lives on the other machine, and a visit when the agent only needs to look at or fix something there. See Remote, Colab, browser, and desktop for visits and for Google Colab runtimes, which work the same way.
See the Kernel and Workspaces references for implementation details.
Previous artifacts Next remote, colab & ui