UTS

One machine runs the AI agent. Every machine is within its reach.

An AI agent is bounded by what it can see and that boundary is usually a single machine.

UTS pushes it out to the edge of the subnet: deploy Claude Code/Codex/Gemini on one machine and it can survey, query and collect across all of them, including PC, COTS Server, even Proprietary Hardware Testbed.

Nothing runs on the other side. No agent/daemon/package. SSH is ALL YOU NEED.

Unifying The Sky: One Claude/Codex/Gemini Deployment, Multiple Backends Support

Quick Start

(0) Before Everything:

# rename
mv host.example.json host.json
# then customize host.json

Every command that touches the network says who it is talking to, by the name in hosts.json. There is no default: -H for some, -a for all.

(1) Basic Commands:

./uts hosts                                   # what the names are
./uts status -a                               # are the machines alive
./uts status -H test-a                        # just that one
./uts ls -H test-a,test-b '~/data/'           # several

./uts ls -H test-a '~/data/'                  # how many, how big, what types, how recent
./uts peek -H test-a '~/data/*.csv'           # right shape? several runs mixed together?
./uts pull -H test-a '~/data/*.csv' --dry-run # what would be fetched
./uts pull -H test-a '~/data/*.csv'           # fetch into .uts/ and record it
./uts pull -H test-a '~/data/*.csv' --to ./raw

./uts push -H test-a ./setup.sh ./lib --to '~/bin/'   # --force to overwrite

(2) Let the agent use it:

./install-skill    # links the skill into ~/.claude/skills and uts onto your PATH

Then open Claude Code in any project and ask about a machine — "are my boxes up", "how far along is the training run on the GPU host" — without mentioning uts. Both links point back at this checkout, so editing the code or the skill takes effect immediately.

(3) Further Development:

See User Manual.