Over SSH, and in Remote-SSH editors¶

The short version: dictating into an SSH session needs no setup, because the remote machine is not involved.

YazSes injects text at the OS level, into whatever window has focus on your machine. If that window is a terminal with an SSH session in it, the characters go down the same pipe as the ones you type. The remote host cannot tell the difference, and there is nothing to install on it.

What works with no setup¶

Remote-SSH editors are the case people expect to be hard and is not. VS Code Remote-SSH runs its UI locally and only the server side is remote, so the text field you are typing into is a local window like any other.

What does not work¶

• A remote graphical application — an X11 or VNC window belonging to the other machine. Focus is on your side; the keystrokes are delivered by the remote X server. That is the case yazses remote exists for.
• A serial console or an out-of-band KVM. No text field, no injection.
• A terminal doing its own bracketed-paste handling may interpret a burst of characters differently from typed ones. If the text arrives mangled, switch the injector:

[injection]
backend = "type"     # auto | type | clipboard | wtype

clipboard is the one to avoid here: pasting is a no-op in most terminals.

The remote agent — for a remote GUI¶

When the window genuinely belongs to the other machine, run a small agent there and let your local daemon deliver into it.

On the remote host:

pipx install yazses           # or: pip install yazses
yazses-agent --listen 9875

Real output:

INFO yazses-agent using injector: XdotoolInjector
INFO yazses-agent listening on 127.0.0.1:9875

Note the address. The agent binds 127.0.0.1, never 0.0.0.0 — it is not reachable from the network, on purpose. Your dictation reaches it through an SSH tunnel and nothing else, so an open port on a shared host cannot be used to type into your session.

On your machine:

yazses remote user@host

That opens the reverse tunnel and points injection at the far end. Dictate as normal; the text appears in the remote GUI's focused window.

yazses status      # state: remote_active while it is connected

Choosing which side transcribes¶

Transcription happens where the microphone is — your machine — in both modes. The remote host never receives audio, only the finished text. That is deliberate: sending audio over the network is the thing this project exists to avoid, and it also means a small remote box needs no model and no CPU budget.

The one thing to check on a laptop¶

The tunnel is bound to the SSH session. Suspending the laptop, changing network, or letting the connection time out drops it, and dictation silently returns to typing into the local window:

yazses status | grep -i remote

Re-run yazses remote user@host to reconnect.

What was not tested for this page¶

• The full yazses remote flow against a real second machine. There is no second host here, and sshd is not running on this one. What was verified is that yazses-agent starts, selects an injector, and binds loopback-only — confirmed both by connecting to it and in remote/agent.py.
• VS Code Remote-SSH and JetBrains Gateway are listed on the architectural argument that their windows are local, not from a session in each. That argument is solid, but if one of them surprises you, say so — a report beats a deduction.