The AI terminal that reads the room.
A native terminal for macOS, Linux and Windows. It watches your screen, runs your shells, edits your code, and talks to any model you choose — or none at all.
Speaks to the models you already pay for — or runs fully offline
Context-aware
It reads your screen, not just your commands.
Stuck on an error you can't even describe? Don't write the perfect prompt. Just say what's happening. AZMX AI pulls the live terminal buffer, the open file, and your selection into the conversation — then tells you exactly what broke and where.
You approve the fix. It applies it. The flow never leaves the terminal.
Always on, never loud
Keeps watch while you work.
Tail the prod log. Watch the type-checker. Notice when the build slows down. Ask once, and AZMX AI remembers — quietly. If nothing's wrong, you won't hear from it. If something is, it surfaces in the smallest possible voice, the first time only.
It reads the room. It doesn't break it.
Your data, your call
Your model. Or no model at all.
AZMX AI is bring-your-own-key. OpenAI, Anthropic, Google, Groq, xAI, Cerebras, any OpenAI-compatible endpoint — the agent loop, tools, and per-hunk diffs don't care which. Charges hit your account, on your terms. We never see the traffic.
Or point it at LM Studio and never leave the machine. Same agent. Same diffs. Zero round-trips.
Real desktop app. Real PTY. Real keychain.
No Electron tax. No Chromium ghost in the background. AZMX AI uses the OS window, the OS keychain, and the OS shell — a Rust core under a fast React UI, signed and notarized where the OS asks for it.
A terminal that grew up.
Multi-tab terminal
xterm.js + WebGL, native PTY (zsh/bash/pwsh), split panes, shell integration, true-color, link detection.
Code editor
CodeMirror 6 with vim mode, inline AI autocomplete, and per-hunk AI edit diffs you accept or reject.
File explorer
Material icons, fuzzy ⌘P quick-open, keyboard nav, inline rename, per-file git status.
AI agent & sub-agents
Multi-step agent with approval-gated tools, named sub-agents, voice input, and MCP servers.
Connections
GitHub PRs/issues, Kubernetes pods, AWS EC2, SSH hosts — using your existing auth, never a key paste.
Project memory
AZMX.md for config + a local memory tree the agent recalls from.
You'll let it read your files. So let's be honest.
Keys live in the OS keychain
Written through keyring, service azmx-ai. Not on disk, not in localStorage, not in logs.
It won't touch the obvious traps
A built-in deny-list refuses .env*, .ssh/, and credential paths — on both reads and writes.
It pauses before it acts
Reads auto-run. Writes, deletes, and shell commands always wait for an in-UI approval. You see it before it happens.
It doesn't phone home
No analytics, no crash reports, no account. The binary hits one URL on its own: the update manifest. Block it if you like.
Open it up. It's listening.
The download is small. The setup is short. The first useful exchange takes about a minute.