Show HN: AITerm – a macOS terminal with an AI command loop and a safety gate

AITerm - Native macOS AI Terminal

New — native macOS app

Your terminal,

supercharged with AI

A native macOS terminal that speaks plain English. The AI proposes commands you approve, fixes and explains errors, and runs multi-step tasks — on your own AI: local Ollama, your API key, or your Claude / ChatGPT subscription.

⬇ Download for macOS

⚖️ How it compares

or install with Homebrew brew install --cask vega-llc/aiterm/ai-term

Download not starting? Direct link

Plain English: list the 5 largest files in this folder

AITerm proposed: find . -type f -exec du -h {} + | sort -hr | head -n 5

Risk: Green — review, edit, then press Return

$ npm test

Failed: exit 1

/fix: reads the failed output and proposes the next command

Proposed fix: npm install && npm test

💬

Natural Language Commands

Describe what you want in plain English. AITerm proposes an editable shell command and waits for your approval.

🔍

Error Diagnosis

Use /fix and /explain on failed output to get a correction or a plain-English read of what happened.

🔒

Privacy First

Run fully on-device with Ollama or the managed Apple-Silicon (MLX) engine. Local requests never leave your Mac, and cloud requests go straight to your provider — no middle server — with secrets redacted first.

☁️

Cloud APIs

Bring your own key for OpenAI, Anthropic, OpenRouter, Groq, Gemini and more. Pro adds Amazon Bedrock, Azure OpenAI, and your existing Claude / ChatGPT subscription via the official CLIs.

📝

Security Gate

Every command path goes through one safety check with green / amber / red risk labels and an auditable decision log.

⚡

History & Power Tools

Search command history, switch models, pull Ollama models, open the command palette, and share cleaned-up fix cards.

🗂️

Tabs — Many Terminals, One Window

Open as many tabs as you need — each is its own real shell with its own working directory, scrollback, and AI session. ⌘T to open, ⌘1–9 / ⌘⇧[ ] to switch; each tab's title shows its folder, and a command you approve in one tab only ever runs in that tab.

⬛

Splits — Side by Side

Split any tab into resizable panes — each its own real shell — and lay them out however you work. ⌘D splits right, ⌘⇧D down, ⌘⌥arrows move focus; drag the divider to resize. A new split opens in the folder you were already in, and your keystrokes only ever go to the focused pane.

💾

Saved Workspaces

Your tabs, splits, and working directories come back exactly as you left them when you relaunch. Save named layouts and reopen a whole project setup — every terminal in the right folder — in a single step.

🤖

Agent Mode — Manual or Autopilot, Always Fenced

Give it a goal. With /agent it works step by step — proposes each command, you approve every one. With /auto (Autopilot) it runs on its own, but only auto-executes commands the safety policy classifies as safe for the current project; risky, destructive, network, sensitive, and build/dependency steps pause for your one-tap approval or block outright. Both pass the same safety gate and your Safety Profile — so on a locked-down machine, anything destructive is blocked outright, even in Autopilot.

📋

Runbooks — Save & Replay

Save a multi-step sequence once — or capture it straight from commands you just ran with ⌘⌥R — with {{variables}} you fill in at run time, then replay it with a single pre-flight approval. Destructive steps are flagged up front, and if one fails you can /fix and retry from exactly where it stopped. Chain steps by feeding one step's output into the next with {{step1.output}}.

🔄

Automatic Updates

When a new version is ready, a subtle “Update ready — Install & Relaunch” pill appears in the status bar — no modal interrupting your session, and no silent background install that never lands on an app you leave open for days. Signed, notarized, delivered as tiny delta updates.

✨

Propose, Then Confirm

Plain English streams into an editable command with a green / amber / red risk badge — press Return to run, edit to change, Esc to discard. Risky commands never run without your confirmation; in Autopilot only provably-safe steps run on their own.

⌨️

One Unified Input

Type a shell command or plain English in one place — no modes. A local classifier routes it; start with “?” to force AI, or a space to force a literal command.

🩺

Fixes & Explains

/fix reads a failed command's exact output — plus your repo, branch, and language — and proposes a correction shown side by side with the original, saveable as a runbook in one click. /explain (⌘⇧E) gives a plain-English read of any output — line by line if you want.

🧠

Bring Your Own AI

Run it your way — free: local & private with Ollama, or your own key for OpenAI, Anthropic, OpenRouter, Groq, Gemini, DeepSeek and more (everything stays on your Mac). Pro plugs in your existing Claude or ChatGPT subscription via the official CLIs, plus Amazon Bedrock, Azure OpenAI, and a managed on-device Apple-Silicon (MLX) engine.

🔒

One Security Gate

Every command — typed, translated, or AI-suggested — passes a single, type-enforced safety check, red-teamed against ~30 evasion tricks. Switch Safety Profiles (Personal → Work → Production → Locked Down) to make the gate as strict as the machine you're on. Audit log of every decision.

↩️

Rollback Suggestions

Before you run something destructive, AITerm offers a conservative way back when one safely exists — an inverse command (git checkout -, mv back, rmdir) or a backup-first step (cp -a … .bak). It stays quiet rather than inventing a fake “undo.”

🛡️

Private by Design

Local models never leave your Mac. Cloud requests go straight to your provider on your account — no middle server — with keys, tokens, and secrets redacted first.

🖥️

A Real Terminal

Full PTY terminal: vim, top, ssh, colors, Ctrl-C, and tab-completion all work — with AI on top, not bolted on the side.

🧰

Built-in Power Tools

Command palette (⌘K), model manager to pull and switch Ollama models, fuzzy history search (by command or the plain English you typed), live git-branch status, themes, share-a-fix cards, and onboarding.

The native AI terminal — free forever. Private local AI, set up for you on first run. Bring your own cloud key — everything stays on your Mac. Pro adds your Claude / ChatGPT subscription + automation. No account needed to start.

⬇ Download for macOS See plans

Apple Silicon · macOS 13+ · notarized by Apple · auto-updates

or install with Homebrew brew install --cask vega-llc/aiterm/ai-term

Tabs & panes

New tab⌘T

Split the pane right⌘D

Split the pane down⌘⇧D

Close pane (then tab, then window)⌘W

Move focus between panes⌘⌥ ← → ↑ ↓

Next tab⌘⇧]

Previous tab⌘⇧[

Jump to tab 1–9⌘1 – ⌘9

AI & tools

Command palette⌘K

Command history⌘Y

Model manager⌘⇧M

Runbooks⌘⇧R

Save recent commands as a runbook⌘⌥R

Explain the last output⌘⇧E

AI commands

Multi-step task, you approve each step/agent

Autopilot — safe steps auto-run, risky ones pause/auto

Fix the last failed command/fix

Explain the last command or output/explain

Type raw in the terminal — AI gate offClick

At the prompt

Recall previous commands↑ ↓

Complete a file pathTab

Run the proposed commandReturn

Discard a proposalEsc

Interrupt the running command⌃C

App

Settings⌘,

Show this shortcuts reference⌘/

Quit AITerm⌘Q

Free

$0 forever

The native macOS app — the AI terminal that actually runs commands. No account needed.

AI command loop — propose, approve, run

/fix & /explain any command or output

Tabs, split panes & saved workspaces

Safety gate, safety profiles & rollback plans

Private local AI — set up on first run, then fully offline, no key

Bring any cloud key — OpenAI, Anthropic, OpenRouter, Groq, DeepSeek, xAI, Mistral, Gemini & more

No proxy — keys & prompts never leave your Mac

⬇ Download for macOS

Available now

Pro

$99 / year · per seat

Just $8.25/mo, billed annually — save ~17%.

Early-bird price — locked in for life. 14-day free trial, no card — the trial starts automatically the first time you open the app.

Everything in Free

Your AI subscription — use Claude Max / ChatGPT Plus via the official CLI, run locally (no proxy, no per-request metering)

Runbooks — save, parameterize & replay multi-step sequences; capture from history; output chaining

/agent & /auto — run multi-step tasks: approve each step (/agent), or let Autopilot auto-run the safe ones (/auto) — both behind the same safety gate

Managed MLX engine — a second on-device engine, Apple-Silicon-native, set up for you

Amazon Bedrock & Azure OpenAI connections

Up to 3 of your machines per seat

Priority support

Not ready? Just download the app — every new install starts the 14-day Pro trial, no card.

Small team? Buy a seat per person. Need 10+ seats or invoicing? Email us.

Pro checks in with our license server about once a month to keep your subscription active — automatic and invisible as long as you go online now and then. Only your license is validated; your commands and files never leave your Mac. If your Mac stays fully offline for more than about six weeks, Pro pauses until you reconnect once.

AITerm Warp Wave Cursor

Native macOS PTY terminal ✓ ✓ ✓ editor

BYO cloud key stays 100% on your Mac 1 ✓ ✗ ✓ ✗

Private local AI, one-tap setup 2 ✓ ✗ ~ 3 ✗

Free AI with no credit meter 4 ✓ ~ 4 ✓ ✗

Tiered dangerous-command safety classifier 5 ✓ — — ~

Your Claude / ChatGPT subscription as the app's AI engine 6 ✓ ~ ✗ ✗

Runbooks — capture, parameters, output chaining 7 ✓ ~ ✗ ✗

Price Free · $9.99/mo Free · $20/mo Free (OSS) Free · $20/mo

✓ yes · ~ partial · — none / undocumented · ✗ no

1. Per Warp's own docs, BYOK prompts and your key pass through Warp's backend on each request. Warp's local-model support works the same way: your model must be exposed at a public tunnel URL, and every prompt still round-trips through Warp's cloud to reach a model running on your own machine. AITerm's BYO key and local model never leave your Mac.

1. AITerm downloads a small local engine and model for you on first run — then it runs fully offline.

1. Wave supports local models (Ollama / LM Studio), but you install and run the engine yourself.

1. Warp meters its own AI by credits (1,500 credits/mo on the $20 entry paid tier), and per Warp's pricing FAQ the Free plan now includes no bundled AI at all — on Free you must bring your own key or endpoint, which is unmetered but still routes through Warp's backend (hence the ~). AITerm's local and BYO-key AI is never metered and never proxied.

1. AITerm ships a tiered dangerous-command classifier (explicit confirmation for catastrophic commands) in a real PTY. Cursor gates agent shell calls via sandbox + allowlist + a classifier subagent (its "Auto-review" mode, which Cursor itself calls best-effort guardrails). Warp uses static regex allow/deny lists plus per-action approval — no risk-scoring classifier documented. Wave's AI can't execute commands yet (listed as coming soon).

1. Warp now hosts the official Claude Code / Codex CLIs in its terminal too (plus a separate cloud-harness mode) — but they run alongside Warp's own credit-metered Agent. In AITerm Pro, your Claude / ChatGPT subscription is the app's AI engine: the same propose→approve→run loop and safety gate, powered by the official CLI running locally. (Wave's "Claude Code integration" is tab-badge notifications, not an AI backend.)

1. Warp Drive has parameterized workflows / notebooks (limited on Free) — but no runbook run-history capture or step-to-step output chaining.

Wave is open-source and, like AITerm, local-first and private — there we differentiate on the safety classifier, runbooks, and a zero-setup native Mac experience, not on privacy. Sources: warp.dev/pricing & Warp docs, waveterm.dev docs, cursor.com/pricing & docs (re-checked July 2026).

Ready to get started?

Download the app, drag it to Applications, open it — the setup assistant gets you to your first AI command in about a minute. Fre

[truncated for AI cost control]