Pick your agent per tab and run them side by side: chat, build entire apps, preview them live, generate images and video, and talk by voice — all in one window. Antigravity and other ACP-compatible agents plug in too. You never have to touch a terminal: if you've used an AI coding assistant, you already know how to drive it — and the real terminal is still in there for when you want it.
Free · open source · macOS app coming · Linux available
A factual rundown of what shellX v0.2.12 does today — no jargon, updated every release.
/build command keeps it working across many steps until the job is verified done
Ask shellX to build a website or an Expo app and it runs it for you — a local preview server, right inside the window. Then it does what a chat box never can: it takes a screenshot of the running app and inspects it with its own eyes. A blank screen, a button in the wrong place, a console error — the agent sees it and fixes it before it ever tells you it's done.
Type /build "make this production-ready" and shellX
runs a long-horizon build: it writes a plan, works through it,
and refuses to claim it's finished until the work is verified. A
reviewer subagent checks the code, a verifier runs the gates, and
anything with a UI has to come back clean from the live preview.
Every step is logged as a receipt — plan writes, file changes,
checkpoints, completions — so you can read exactly what happened.
build.md — edit it, share it, version it/pause and /resume any time
Image and video generation are built into every session — so you can be deep in a Claude Code or Codex build, ask for a picture, and get it without switching tabs. shellX routes the request to Grok Imagine on your Grok account, or to GPT Image through Codex, drops the result straight into your conversation, and the agent uses it right away. Grok Imagine also makes short clips with native synced audio — all collected into their own gallery.
Open Trace on any session and shellX draws its activity as a live graph — every file it read, searched, wrote or deleted, every git move, every subagent it spawned, every build receipt, with the media it made linked right in place. Resize the evidence panels for a records-heavy run, and search the whole trace by path, command, query, tool, source, or timestamp. Running several agents at once? Trace rolls their activity into one report you can watch from outside the window.
Sixty-one built-in tools the agent can call directly — the same set whether it's running on your PC, in WSL, or over SSH. Voice, vision, files, processes, secrets, and more. Toggle them per session.
Real microphone in, real spoken answers out — not a dictation gimmick. Push-to-talk for one-shot prompts; voice-chat mode re-arms the mic after every reply so you have an actual conversation while your hands stay on the keyboard. Uses your Grok account for speech, and now reads back Claude and Codex replies too — nothing extra to set up.
→ push-to-talk · voice chat mode · per-tabshellX captures your screen, sends it to Grok's multimodal model, and returns a description the agent can act on. Verify a deploy. Catch the dialog blocking your terminal. Audit a UI you just shipped — the same eyes that check its own live previews.
→ full desktop, active window, or a named windowRead any file. Write any file. Fast search across whole project trees. Scoped to a working folder, and every call is auditable.
List, inspect, signal, and read output from any process the agent can see. A build hung? It finds it and stops it without you ever leaving the chat.
Grab the desktop, a window, or the live app preview, then pipe it into vision_describe for a closed loop: see → reason → fix. This is the engine behind the live preview check.
An encrypted vault backed by your OS keyring. The agent calls a secret by name — the value never appears in chat, logs, or transcripts.
Every request is logged with response codes and byte counts. Allowlist hosts per session. The agent stays on-rails.
Dispatch isolated subagents from the running session. Each gets a
fresh context window and a role-baked prompt, and reports back to
the parent. The foundation under /build.
Discover and install MCP servers from the public registry with a one-click UI, project-scoped or global. Any MCP server you already use elsewhere works here unchanged.
Conversations are saved and searchable across history. Resume a week-old session by name. Export any session as a single zip for a teammate or for CI replay.
shellX can run the agent on your Windows machine, inside a WSL distribution, or on a remote server over SSH — and carries its full toolset along wherever it goes. And when you do want a terminal, it's a real one: a true PTY, run vim or htop, no fake echo.
Real PTY via ConPTY. Run vim, htop, anything interactive. No fake terminal, no command echo simulator.
→ Settings · Connections · Local → binding ConPTY 200x60 → host tools on 127.0.0.1 ready · grok-4.3 you> read C:\src\index.ts ok · 142 lines
Spawn the agent inside any installed distribution. The host toolset streams in — same tools, Linux paths.
→ Settings · Connections · WSL · ubuntu-24.04 → entering distro → tools tunneled: Windows → WSL ready · grok-4.3 you> grep -r 'TODO' ~/repo ok · 7 matches
Any Linux box with SSH. Grok runs there; tools are tunneled back. No agent on your laptop, no inbound ports on the server.
→ Settings · Connections · SSH · prod-01 → ssh forward: localhost ←→ remote → pty: 200x60 (xterm-256color) ready · grok-4.3 you@prod-01> tail -f /var/log/app.log ok · streaming
Everything you can click in shellX, another program can drive over HTTP — bearer-token gated, loopback-bound, origin-checked. Spawn a session, send a prompt, run a build, capture a screenshot, archive the workspace as a zip — without a human ever touching the window. The foundation for CI hooks and headless agent fleets.
/connect · /prompt · /build/* ·
/preview/work/* · /autonomy ·
/screenshot ·
/state/{header,sessions,subagents} ·
/permissions/:reqId/respond ·
/diagnostics · /tabs/:id/archive ·
vault · plugins · sessions/history.
If a human can click it, an agent can drive it from outside.
Subscribe with a single WebSocket. Every agent frame, every tool call, every build receipt, every permission request, every prompt completion lands typed and tagged with tabId.
When the agent asks for permission and autonomy is "Confirm", the
request lands on the event stream with a reqId. POST a decision to
/permissions/:reqId/respond — your orchestrator is
the user.
One POST captures the working tree, the session scratch dir, every emitted event, and the active plan as a single zip. Drop the bundle in CI, replay deterministically.
shellXagent binds 127.0.0.1, requires a per-install bearer token minted from the OS random source, and enforces an origin allow-list server-side. Your machine, no inbound ports.
Self-test the running install — files, tools, screenshot, vault, sessions, connections, settings, auth, and preview setup. Returns a structured pass/fail report your CI can gate on.
full surface @
docs/API.md
Every design choice has a corresponding rejection. These are ours.
Free. Open source. No account required to install. Bring an agent — sign in to Grok, Claude Code, Codex, or Antigravity to start talking; new Grok accounts receive free credits.