Practical guides · 02
What your AI terminal sends, and how to check it.
Local and cloud models take different paths. Choose how much context a cloud model sees, then read what a request actually carried.
4 minute read
An AI gives better terminal help when it knows a little about what you just did: the last command, its exit code, maybe the error it printed. Where that context goes depends on the AI you picked. Before you connect a model, it helps to know which path a request takes and how to check one afterward.
1. Know which path you chose.
A local model, such as one served by Ollama on your Mac, receives requests without them leaving the computer. It needs an initial engine and model download; after that, it can work offline.
A cloud connection, such as OpenAI, Anthropic, Gemini or OpenRouter with your own key, sends the request directly from your Mac to the provider you chose. There is no AITerm-hosted proxy in between. That provider’s own terms and usage charges apply.
Neither choice is right for every task. A local model keeps the request on your Mac; a cloud model may suit a harder question. The point is to choose on purpose.
2. Choose how much a cloud model sees.
In AITerm, Settings ▸ Privacy has a setting called “How much the AI sees.” It caps the session context sent to cloud providers:
- Nothing: no ambient session context.
- Recent commands only: recent command lines and exit codes, without their output. This is the default.
- Recent commands + output: recent commands, exit codes and the end of their output.
The setting only limits what cloud models receive. A local model gets the full session context, because that context stays on your Mac.
If you are working near configuration files, logs or anything else that might hold sensitive values, keep the level at Nothing or Recent commands only, or switch to a local model for that task.
3. Treat redaction as a safety net.
Before a request goes to a cloud provider, AITerm redacts common secret patterns, including API keys and tokens, Authorization headers, passwords in KEY=VALUE form, credentials inside connection strings, private-key blocks and long hex or base64 strings.
That catches a lot of accidents, but pattern matching cannot recognize every secret. A password with no label, an internal hostname or a customer name in a log line may look like ordinary text. Choose the context level first and let redaction back it up.
4. Read what was actually sent.
After an AI request that carried context, the eye chip on the input bar shows what the provider received. For a cloud provider, it shows the redacted version, produced by the same step that sends the real request. For a local provider, it shows the raw context, since nothing left the machine.
If the AI you switched to cannot start, nothing is sent: AI stays off until you fix that connection or pick another, and the chip says Not sent — AI is off.
A five-minute check.
- Pick the AI you plan to use and note whether it is local or cloud.
- Open Settings ▸ Privacy and choose a context level for cloud requests.
- In a folder with nothing private in it, run a harmless command such as
ls. - Ask a question about that output.
- Open the eye chip and compare what was sent with what you expected. Adjust the level if it carried more or less than you want.
Try it in AITerm.
AITerm keeps provider keys in the macOS Keychain, and your command history stays on your Mac. The core app is free on Apple Silicon Macs with macOS 13 or later and includes AI command suggestions, /fix, /explain, tabs, split panes, saved workspaces, Ollama and supported bring-your-own-key connections. Cloud provider charges are separate.
Keep learning.
- Three checks before you run an AI-generated command
- What is included in Free and Pro
- Safety settings and command review
Prepared with AI assistance from the current AITerm documentation. The steps follow the handbook; this is not a recorded app demonstration.