v0.1.77 — Claude Code goes GUI, your session goes mobile, and mobile QA gets real
Released 2026-09-12. Signed + notarized on macOS (Apple Silicon + Intel); Windows + Linux published. Gated on full smoke + both E2E journeys.
A substantial release. Three things land together: the Claude Code agent moves out of the terminal into a rich point-and-click chat; you can drive your desktop coding session from your phone; and the built-in mobile QA automation (iOS + Android) gets a batch of real correctness fixes after field testing on Android.
Claude Code is a GUI chat now (not a terminal)
The Claude Code tab used to be a raw terminal — you couldn’t attach a file or click to approve an edit. It’s now a first-class chat, the same surface SMART-Code already had:
- A real chat — input box, scrollable transcript, streamed responses, tool calls rendered inline.
- Attach files and paste screenshots — drop a file onto the pane or paste a screenshot (⌘V). The agent reads them with its own tools.
- Click to approve — when Claude wants to edit a file or run a command, an Approve / Deny card appears with the diff. Reads run unattended; writes and shell commands always ask first.
- Multi-turn, per project — each project keeps its own Claude Code conversation; New chat starts fresh.
- Nothing removed — a ⇄ Terminal toggle flips back to the raw
claudeCLI instantly, and the dedicated Terminal tab is still a full terminal.
Under the hood it drives the real claude CLI headlessly and maps its live stream into the GUI — same Claude Code you trust, with a human interface. Codex and Gemini get the same treatment next.
Dropping the terminal doesn’t cost you any access to your machine — the agent runs as a local process with full filesystem and command access either way. You only trade the CLI’s own TTY prompts for GUI buttons (with the ⇄ Terminal escape hatch right there).
Remote control — drive your desktop from your phone
Your coding session now follows you. A signed-in desktop with a project open registers a live relay session; your phone pairs to it and can send prompts, watch the model stream and diffs live, and approve writes remotely.
New: no project linking required — your own desktop shows up on your phone with zero linking. Just sign in on both. See AF Mobile IDE for the step-by-step.
You can now drive both engines from the phone — SMART-Code and the real Claude Code CLI (headless, with write-approval gated the same way).
Mobile QA automation — Android + iOS fixes
v0.1.75 shipped the built-in mobile QA suite (YUKTI). Field testing on Android surfaced four bugs where a flow could silently do the wrong thing — or nothing — and still report PASS. All are fixed, and we audited the iOS path for the same classes of bug.
| Bug (Android) | Fix |
|---|---|
Only the first step ran, yet PASS. The step loop piped through while read, so a step’s adb shell consumed the same stream and swallowed the rest. | Steps now feed a separate file descriptor (read <&3). All steps run, and a failing step fails the flow (exit 1). |
Wrong element tapped. Lookup took the first visible-text substring and ignored resource-id — “Email” hit the “or use email” caption, “Password” hit “Forgot Password?”. | Lookup now ranks candidates: resource-id + exact label + editable/clickable controls beat a static caption. |
Passwords with special characters failed. type passed the string to the device shell unquoted, so a * was glob-expanded (8-char password → 11 chars). | The typed text is quoted for the on-device shell; spaces still map to %s. |
| Android ran the iOS path. Platform defaulted to iOS and nothing set it. | Platform is resolved from the project config (auto-detected on scaffold) — Android runs adb/uiautomator/Gradle, iOS runs the Simulator. |
iOS audit: only the element-lookup bug reproduced on iOS — now fixed with the same identifier-aware, field-preferring ranking (honoring AXUniqueId and element type). The step-loop fix is shared; the quoting bug never applied to iOS (it types character-by-character with no device shell).
Upgrade
- Desktop: update to v0.1.77 (auto-updates on relaunch). macOS arm64 + x64, Windows, Linux AppImage.
- Mobile: nothing to install — open agentfoundry.me/app and sign in. Add it to your home screen.
Known issue: the Linux AppImage may not launch on very old distributions (e.g. Ubuntu 20.04) due to a glibc mismatch in a native module; a broader-compatibility Linux build is in progress. macOS and Windows are unaffected.