Codex Remote goes GA as Cursor and Claude Code polish mobile and terminal workflows

Remote control from anywhere

The biggest story this week is mobile access to coding agents. OpenAI declared Codex Remote generally available with authenticated one-to-one QR pairing between the ChatGPT mobile app and any connected Mac or Windows host. A new DigitalOcean plugin lets Codex provision a Droplet and connect it as a remote workspace in one flow, and connections established since June 8 carry forward seamlessly. Meanwhile, Cursor launched its own iOS mobile app for all paid plans — launch cloud agents from your phone, speak commands with voice input, and track progress via Live Activities on the lock screen. The same release shipped Remote Control for directing a desktop agent from the phone. Two platforms betting that agentic coding doesn't stop when you step away from the keyboard.

Performance and reliability across the big three

Anthropic's Claude Code v2.1.191 delivered a meaningful quality-of-life batch: streaming CPU usage dropped roughly 37%, the scroll position no longer jumps to the bottom while you're reading earlier output during a response, and stopping a background agent from the tasks panel is now truly permanent — no more resurrecting agents. The team also added /rewind support for resuming from before /clear and fixed hook matchers with comma-separated patterns. An earlier v2.1.186 brought CLI authentication for MCP servers via claude mcp login and logout — useful over SSH with --no-browser — and made bash ! commands trigger automatic Claude responses. v2.1.187 added sandbox.credentials to block sandboxed commands from reading secrets and enforced org-configured model restrictions across every entry point.

xAI's Grok Build saw a flurry of polishing releases. The dashboard in v0.2.64 now displays the current directory and branch with Ctrl+L to switch location and Ctrl+W to dispatch new agents into fresh git worktrees. v0.2.65 lets grok -w --ref <branch> create worktrees from a specific ref instead of HEAD. On the configuration side, v0.2.62 made prompt and /usage warnings correctly reflect prepaid credits and auto top-up status, and made TODO lists persist after compaction. v0.2.68 enabled hot-reloading MCP servers without restarting sessions — a notable developer-velocity win. The jumbled v0.2.63 fixed hook matcher over-matching, while v0.2.66 and v0.2.67 added --json-schema for headless output constraints and custom sandbox profiles with per-file kernel-deny rules.

Codex itself shipped v0.142.2 with MCP tool search enabled by default, macOS system proxy, PAC, and WPAD support for authentication, and fixes for Amazon Bedrock credential recovery and PowerShell AST safety checks. A smaller v0.142.1 added the same proxy support for Windows.

Deeper MCP and automation integration

Cognition's Devin added session pinning and background start from the Cmd+K palette (June 25), then followed up with archive/unarchive from the same palette and run-once automation schedules (June 26). An "out of date" banner now surfaces stale MCP installations with a one-click marketplace refresh.

Claude Code's v2.1.195 fixed a tricky voice-dictation problem where silence was captured after the default input device changes, and corrected hyphenated hook matchers (code-reviewer) from substring-matching to exact-matching — use mcp__brave-search__.* to match all tools from a hyphenated MCP server instead.