Wayfinder Try it free

Help

Set up the assistant

Ask about a host in plain words, and an assistant looks into it with commands it runs there, then answers with the lines that show it. It uses a model you choose, on your own account or your own network. It is off until you set one up.

Choose a model

With Ollama running on this Mac, there is nothing to fill in. Open the assistant in an SSH tab (⌘J) or look at the welcome screen: Wayfinder picks the Ollama model best suited to investigating, one that can run commands, and offers it: Use qwen2.5:14b on This Mac, for example. One click sets it up, and nothing leaves the Mac.

To choose yourself, click Choose a Model… there, or open Wayfinder › Settings… › Assistant, and pick where the model comes from:

Choosing a model from the list tests it at once; Test does the same for a name you typed. A key is kept in this Mac's Keychain. To use Ollama on this Mac, install it and fetch a model once:

ollama pull qwen3:14b

Ask

  1. In an SSH tab to the host, press ⌘J, or choose Shell › Ask About This Host. The first time, the panel asks you to accept that the assistant is AI, that it can be wrong, and that what you run on your systems is your responsibility. Nothing is sent until you do; Settings › Assistant shows when you accepted and can withdraw it.
  2. Ask, for example, why the load is high or why a service will not start. Include what the terminal shows is ticked; untick it to leave the screen out.
  3. Commands that only read run at once. Anything else waits with the reason, for you to Run or Skip. A command that deletes, stops, restarts or reconfigures something has a red card, and its Run asks once more.
  4. The answer shows as the model writes it, and quotes the lines that show the cause. A fix it suggests has Put at the Prompt, which types it into the terminal without running it.

Or start from the terminal: choose an error or a log line, right-click it and choose Ask the Assistant About This. With nothing chosen, Ask the Assistant About the Screen asks what the terminal shows. While a debug session is recorded, each question takes along what was typed and run in every tab since it started, so the assistant does not suggest what you already tried; untick Include what this debug session recorded to leave it out.

What it may run by itself

Only commands Wayfinder can tell only read, such as ls, df, ps, journalctl and systemctl status. Anything that could change the host, runs as another user, reaches another host, or reads what looks like a secret waits for you. A second check of its own knows what is hard to take back, such as rm, systemctl restart, kill or writing over a file: such a command never runs by itself, even if the first check were wrong about it. Each command runs for at most 20 seconds. To approve every command yourself, turn off Run commands that only read without asking in Settings › Assistant. Every command it runs is in Window › Activity Log, marked as the assistant's.

What it costs

Under each answer, Wayfinder says how many tokens it took and how many came from the provider's cache, which Anthropic bills at about a tenth of the price. The model sees at most 8,000 characters of each command's output, and when you ask a follow-up, the outputs of questions already answered go again shortened. A new conversation (the pencil at the top of the panel) starts from nothing.

What the model is sent

Settings › Assistant lists it all: your question, the host's name and your user on it, a few facts about its system, the commands and their output, the terminal's screen when that is ticked, and the conversation so far. Nothing is sent until you ask. What looks like a password, token or private key is replaced with [redacted] first, and your saved passwords are never part of it. The Security page has the full rules.

When Test fails

Wayfinder says what the server answered and what to check. A busy or rate-limited server is asked again three times before you are told, as long as the server says to wait, up to half a minute.

Back to Help