Tether Nexus · engineering preview

Your machines, as one computer.

Ask for what your work needs — a GPU with enough memory, ffmpeg, the fastest free cores. Tether decides which machine runs it, and whether any data has to move.

Download How it works

Illustration, not a screen capture. Nothing on this site renders a number that is not in the measurements it links to.

What changes

Ask for a capability

Name what the work needs. Tether answers with a machine that reports it, or with a plain refusal and the reason. An address is an implementation detail, not a request.

Move the work, not the bytes

If a machine already holds the data, the job goes to it and a result comes back. Data leaves a machine only when moving it is cheaper than moving the computation.

Every result carries a receipt

What ran, where it ran, how long it took, and a digest of the output. A receipt is a checkable record, not a log line you have to trust.

The request

One line of intent. One plan.

The planner reads the capabilities each machine reports, the cost of each path, and the fixed cost of starting a job somewhere else. Then it prints the decision it took.

You ask

gpu.memory_bytes >= 32GiB
tool:ffmpeg
cpu.free_cores >= all

It answers

plan: move_compute
where: the machine that holds the data
bytes: none moved
receipt: written on that machine

Measured, and printed with its loss

Four numbers, each one a file you can open.

These are the repository's own measurements, not projections. Where a run lost to running locally, it is printed here too.

1.28× faster when one job was split across two machines source: replay.json
2.86× faster on a machine that was already saturated source: measured-not-claimed.json
1.96× slower on a quiet machine — so the job stayed local source: measured-not-claimed.json
33 units is where sending work out starts to pay; 32 stay here source: when-it-helps.json

No cloud in the path

Down is a state, not an outage.

There is no Tether account server between your machines. Pairing is an SSH path you already trust, and a job runs on the machines at both ends of it.

If a donor stops answering, the job finishes where it started and says that it fell back. If a cloud service is down, the machines on your desk still work together.

Start with the machines you already own.

macOS and Windows for now. No public installer is published yet — the download page prints the exact state and the commands that work today.

Download Read the docs