A terminal that pays attentionComing soon
Most of the time it's just a terminal — a real zsh or bash running on a pty. The AI sits on top as a layer, following what you type and what comes back, then finishes commands, explains failures, and turns plain language into commands — without ever running anything on its own.
Your shell, your .zshrc, your habits — all unchanged. Interm only steps in at
the moments you'd otherwise reach for a search engine.
Pause while typing and a grey completion appears. Tab accepts it.
When a command fails, the cause and a fix show up in a panel.
Describe what you want in plain words; get a command to confirm.
Start a line with ? — or press ⌘K — to ask the AI.
What you type decides where it goes. The first character — or the shape of the sentence — routes it to the shell, the AI, or the command converter.
PATH and it runs three words or more.
Stop typing for a moment and a grey completion appears right after your cursor. It's based on the current directory, the files in it, and what you've just been running.
Press Tab to accept. Keep typing something different and it simply disappears — it never gets in the way.
Latency is the whole experience here, so you can point suggestions at a separate, faster model than the one used for conversations.
Whenever a command ends with a non-zero exit code, Interm reads the command and its output and opens a diagnosis panel with the likely cause and a command to fix it.
Click To terminal and the fix is placed on your prompt — you still press Enter yourself. Drag the panel by its header if it's covering something.
Write what you want in plain language and press Enter. Interm turns it into a command and shows it in a confirmation bar above your prompt.
Enter runs it, Tab drops it onto the prompt so you can edit first, Esc cancels.
If a value is missing — a username, a host — it's left as a placeholder like <user>, already selected so you can just type over it.
A terminal can delete a disk in one line, so the guardrails don't depend on the AI behaving well. They're enforced locally, by the app itself.
Every command the AI produces lands in a confirmation bar or on your prompt. It only runs when you press Enter.
Right before running, the app re-checks the command itself. Even if the AI called it safe, destructive patterns bump it to dangerous and the bar turns red.
Keys are encrypted with the OS keychain (safeStorage) before being written to ~/.interm/config.json. An exported ANTHROPIC_API_KEY works too.
Interm generates its own rc files that load your original config first, then add a few hooks. Your .zshrc is never edited.
Switch providers anytime in Settings (⌘,). Whichever you choose, the model list is fetched from the server, so you pick from a menu instead of memorizing long model IDs.
claude-haiku-4-5.Accept the ghost suggestion.
Close a suggestion, panel, or confirmation bar.
Ask the AI — whatever you'd typed becomes the question.
Open Settings: provider, models, keys, font.
Open a new window.
vim, less and other programs, Esc passes through as usual.
It opens like any terminal, with your usual shell and prompt. The window title follows your current directory.
Press ⌘,, choose a provider, and enter an API key — or point it at LM Studio / Ollama on your Mac to stay fully offline.
Type as usual and let suggestions appear. Start a line with ? to ask a question, or simply write what you want in a sentence.
brew install --cask font-d2coding.