TALOS

Workflows / 03 Run remote

talos-sky

Outposts for your coding agent.

talos-sky gives the agent on your laptop, Claude Code or Codex, an outpost on every machine you can reach over SSH: a rented GPU, a cloud VM, a lab server. Through it the agent runs commands and long jobs, copies files both ways, has cheap rules watch the logs, and sees renders, videos, 3D models and metrics as they are made.

  • v0.5
  • private preview
No keys on hosts
The outpost runs no model. Model calls leave from your laptop, with your agent's own login.
python3 is enough
No pip, no network and no repository access needed on the host.
Agent and terminal
Ten MCP tools, and a tsky command for each of them.
Survives sleep
Events from while the laptop was away arrive when it comes back.

How it works

The laptop opens every connection. On the host, the outpost keeps jobs, watches and an event log, and listens only on a private socket reached through SSH. Below is one background job, end to end.

One background job: start it, watch it, hear about it, look at what it made.
Laptop
Claude Code · Codexmain agent, plus an outpost subagent on a small model for multi-step remote work
tsky mcpone live link per host · pushes events into the session
tsky view · forwardpreview board and tunnels, on 127.0.0.1 only
exec · put · get · watchjob ends · alerts · previews
Host · any ssh-able machine
Outpostno model · no API keys · needs only python3
Jobsown process group · survive restarts and upgrades
Watcheslog patterns · disk · memory · each hit with a GPU snapshot
Event lognumbered · written before delivery · replayed after sleep
preview/png · mp4 · glb · ply · metrics
  1. tsky add copies the outpost over SSH and starts it. Upgrades swap atomically and roll back if the outpost does not come back.

  2. The agent calls exec in the background. The outpost starts the job under its own supervisor and hands back a job id at once; the job keeps running if the outpost restarts.

  3. The agent asks the outpost to follow train.log for nan|Traceback. Disk and memory are watched from the start. The rules run on the host, without a model.

  4. The loss goes NaN. The watch hit is written to the numbered event log first, with a snapshot of load, memory, disk and GPU.

  5. The event is pushed into the live conversation and the agent decides what to do. Events from while the laptop slept are replayed when it reconnects.

  6. Renders, video, 3D models and metrics the job writes to preview/ come back for the agent to look at. TensorBoard, Rerun or viser come through a localhost tunnel.

What it does

Jobs
Run a command, or start a long one in the background and get a job id. Kill it, tail it, or let it finish while you are away.
Files
Copy files and folders both ways over one shared SSH connection per host. Nothing is overwritten unless asked.
Watches
Cheap rules on the host: a log line matching a pattern, a full disk, low memory. Each hit arrives with load, memory and GPU state.
Previews
Anything a job writes to preview/ reaches the agent: images, video, 3D models drawn from four sides, metrics as curves.
Remote viewers
Bring TensorBoard, Jupyter, Rerun, viser or a simulator page to a localhost URL through a tunnel that reconnects by itself.
Two modes
Operate, gated by your agent's permission prompts, or read-only: allowlisted inspection commands, no shell, no uploads.

Use it

Everything the agent does through MCP, a person can do from the terminal
tsky add gpu-1 --ssh ubuntu@gpu-1.example.com   # installs and starts the outposttsky exec gpu-1 -b -- python train.py           # a long job, with a job idtsky watch gpu-1 --file train.log --pattern "nan|Traceback"tsky peek gpu-1 --label render                  # the latest render, inlinetsky forward gpu-1 6006                         # TensorBoard on localhost

Safety

  • The outpost runs no model, and tsky stores no credentials on either side.
  • Everything a host sends is treated as untrusted: capped, cleaned, and never able to approve a tool call.
  • The laptop opens every connection. The outpost listens only on a private Unix socket reached through SSH.
  • The preview board and tunnels listen on 127.0.0.1 only.
  • Copies unpack only regular files inside the target, and overwrite nothing unless asked.
  • Read-only mode is a guard rail, not a sandbox: rented machines get a low-privilege user and no long-lived credentials.

Limits and next

Limits today
Events reach Claude Code live through its research-preview channels; other agents read them with the events tool. Previews over 32 MiB are not fetched. Renting a GPU goes through your own provider CLI, after you confirm the offer.
Next
3D previews you can turn around, a shared preview board for a team, hosts straight from a scheduler, sandboxes from ready images such as gmsh and CalculiX, and storage that outlives a host.