An SSH terminal with an AI assistant that can run on your own model

What the console does

The SSH console in the RDP FIDO client is a terminal with a side panel. On the “Assistant” tab you describe a task in words — “find what is filling the root partition” — and get commands back, each in a block with three buttons: “Insert” types the command into the terminal without Enter, “Run” executes it, “+ Save” keeps it in your list of commands.

Underneath it runs your system ssh, so keys, ssh-agent and ~/.ssh/config work as they always did. The console is part of the client for Windows 10 / 11 and Linux, the free edition included.

Connect it to a local model

1. Start a server that speaks the OpenAI API. Ollama serves it at http://HOST:11434/v1, LM Studio at http://HOST:1234/v1. If the model runs on another machine, make the server listen on the network and not only on localhost.
2. Open the console: click an SSH tile in the client, or run RdpFidoClient.exe console --host 203.0.113.10 --user admin (on Linux: rdpfido console SESSION).
3. On the “Assistant” tab press the gear, choose the OpenAI-compatible service and enter the address. The console asks the server for its models and shows them as a list; an API key is optional.
4. If the model is slow, raise “Wait for the answer, seconds” in the same settings: the default is 600, the maximum 3600.

# Ollama on a machine in your LAN
OLLAMA_HOST=0.0.0.0 ollama serve
ollama pull qwen2.5-coder:14b

# address for the console:  http://192.168.1.50:11434/v1

With a model inside your network neither your questions nor the terminal text leave it: requests go from your computer straight to the address you entered, not through our servers.

For Claude, choose the Anthropic service and enter an API key; the address and the model are already filled in.

What the model sees — and what it never sees

• Nothing from the terminal by default. The assistant gets your question and knows which host and user the session is for.
• The screen, only while “Sees the console” is ticked. Then the last 200 lines go with each question. The rule is enforced by the program itself, not by the panel's web page.
• Passwords: never. Saved passwords are typed by the program, the terminal does not echo them, and they are not part of the text sent to the model.
• The API key is stored encrypted — DPAPI on Windows, the desktop keyring on Linux — and is sent only to the address it was saved for.

One caution no program can remove: what is on the screen is on the screen. If you print a file with secrets and then ask a question with “Sees the console” ticked, those lines go to the model. With a cloud model, untick it first — or use a local one.

Who presses Enter

• By click only (the default): the assistant suggests, you insert or run.
• Insert the first one: the first suggested command lands on the command line; Enter is yours.
• Run the first one: the first command is executed at once. The console warns about this mode when you switch it on.

The assistant is instructed not to open with a destructive command and to warn before one that removes data, restarts a service or changes access. It is still a language model, and small local models make more mistakes than large ones: read a command before you run it.

Describe the server once

Prompts are standing notes for the assistant: which distribution this is, what is installed, what must not be touched. Ticked prompts go with every question — for this server or for all of them. Conversations are saved per session, so yesterday's investigation can be continued today; opening an old conversation types nothing into the terminal.

What it does not do

• No file transfer and no tabs: one terminal per window.
• No autonomous agent: it suggests commands, and at most runs the first suggestion if you switch that mode on.
• On Linux the console window needs a Chromium-family browser; without one the page opens in the default browser.