Cursor iOS Public Beta with Remote Control, Claude Code v2.1.21x tools, Codex Remote GA (mobile agents and CLI tools)
TL;DR
From 17 to 24 July Cursor, Claude Code, and OpenAI Codex released mobile betas, session limits, safety tools, and remote access features that let builders run agents across devices and IDEs.
What shipped
From 17 to 24 July the three main AI coding environments pushed updates focused on mobile control, team controls, and cross-device workflows. Cursor emphasized iOS access and marketplace rules. Claude Code and Codex added session safeguards and remote pairing options that affect how agents run in production.
Cursor
Cursor shipped fifteen updates centered on its iOS public beta and team marketplace controls. The releases add remote agent management from phones, side chats, and Slack multi-repo support that compete directly with Claude Code and Codex CLI. Admins gain group-based access rules while developers get router-based model selection for cost and speed tradeoffs.
- •Cursor iOS Public Beta Cursor released its iOS app in public beta on paid plans so users can launch cloud agents and control desktop sessions remotely with push notifications.
- •Team MCP Distribution Version 3.10 lets admins push Team MCP servers to cloud agents, IDE, and CLI while limiting marketplace access by organization groups.
- •Side Chats and Search Version 3.11 adds persistent side chats for parallel exploration plus local transcript search via Cmd+K and updated project pickers.
- •Slack Execution Plans Slack integration now shows plans before tasks run, supports mid-task repo switches in multi-repo setups, and shares context across channels.
- •Auto Mode Router Auto mode uses Cursor Router to send requests to frontier or cheaper models depending on task complexity and chosen cost or speed goals.
- •Unified Customize Page A new Customize page centralizes plugin, skill, MCP, and hook management with marketplace leaderboards and third-party repo imports.
Claude Code
Claude Code delivered thirty-four point releases that tighten session safety, add accessibility options, and refine subagent behavior. The changes introduce the EndConversation tool, manual skill triggers, and filesystem sandbox toggles that reduce runaway loops compared with earlier Cursor and Codex versions. Builders gain clearer permission prompts and live timers for long tool calls.
- •Fork and Subtask Updates Version 2.1.212 copies conversations into background sessions via /fork, adds /subtask, and caps searches plus subagents per session.
- •EndConversation Tool Version 2.1.214 adds the EndConversation tool to stop abusive sessions and requires Docker daemon prompts for safety.
- •Manual Verify and Review Version 2.1.215 stops automatic /verify and /code-review runs so users must trigger them explicitly with slash commands.
- •Filesystem Sandbox Toggle Version 2.1.216 introduces sandbox.filesystem.disabled to bypass isolation while keeping network controls and fixes long-session stalls.
- •Emoji Autocomplete and Caps Version 2.1.217 adds emoji shortcode autocomplete, limits concurrent subagents to twenty, and warns on transcript save failures.
- •Background Code Review Version 2.1.218 runs /code-review as a background subagent and improves MCP connection status reporting.
- •Subagent Streaming Flag Version 2.1.211 adds --forward-subagent-text for stream-json output and fixes model inheritance bugs across sessions.
- •Live Tool Timers A live elapsed-time counter now appears on collapsed tool summaries for long-running calls.
- •Vim Insert Remaps The vimInsertModeRemaps setting maps two-key sequences such as jj to Escape in insert mode.
- •Screen Reader Mode An opt-in screen reader mode renders plain text and is enabled via flag or settings file.
- •Deprecated Task Mode The Task tool mode parameter is ignored and subagents inherit the parent permission mode.
- •Accessibility and Vim Settings Version 2.1.208 adds screen reader mode and vimInsertModeRemaps plus corporate launcher support.
- •Background Dialog Fixes Version 2.1.209 restores /model and other dialogs in background sessions.
- •Live Timers and Isolation Version 2.1.210 adds elapsed-time counters and fixes git-mutating command isolation in subagents.
- •Text Forwarding and Layout Version 2.1.211 streams subagent text and fixes parallel session logout plus terminal layout bugs.
OpenAI Codex
OpenAI Codex released thirty-three updates that integrate Codex into ChatGPT desktop and iOS apps while expanding remote access. New Record & Replay skills, workspace browsers, and multi-folder projects give builders mobile and cross-device control that parallels Cursor iOS and Claude Code terminal features. CLI improvements add Bedrock support and import paths from competing environments.
- •iOS Workspace Browser ChatGPT for iOS adds a file browser, directory picker, global diff controls, and LaTeX rendering for Codex tasks.
- •EEA Rollout Computer Use, Chrome extension, Memories, and Chronicle previews reach users in the EEA, UK, and Switzerland.
- •Record and Replay Skills The Codex desktop app on macOS turns workflows into reusable skills with Record & Replay and thread handoff.
- •Host Personalities ChatGPT for iOS adds Friendly and Pragmatic personality settings per host plus direct goal editing in the composer.
- •Codex Remote GA Codex Remote reaches general availability with QR pairing so mobile users can control desktop hosts and provision Droplets.
- •Desktop App Integration Codex now lives inside the ChatGPT desktop app on macOS and Windows with inline editing and PR Chat for GitHub reviews.
- •Mermaid and Forms iOS renders Mermaid diagrams and interactive forms inside Codex tasks with better thread navigation.
- •Voice and Multi-Folder Projects ChatGPT Voice and multi-folder local projects arrive in the desktop app for cross-thread coordination.
- •Terminal Visualization Links Codex CLI displays secure clickable inline visualization links inside the terminal UI.
- •Multi-Agent V2 Stable Codex CLI stabilizes multi-agent V2 with configurable sub-agent models, reasoning levels, and concurrency controls.
- •Local Audio and Realtime V3 The CLI accepts local audio inputs and runs streaming Realtime V3 voice conversations in the terminal.
- •Amazon Bedrock Login Experimental Bedrock support and custom endpoints appear with GPT-5.6 Sol as default model.
- •Import from Rivals The /import command migrates settings, MCP servers, plugins, and memories from Cursor and Claude Code.
- •Paginated Thread History Codex CLI 0.145.0 adds paginated history, session resume, search, and long-term memories.
- •Version 26.608 Updates The Codex app gains import flows from Claude tools and a filtered plugins marketplace screen.
- •Version 26.609 Updates Rate-limit banking, Developer mode for Browser use, the /init command, and faster Browser performance ship.
- •CLI 0.144.0 Approval Modes Writes app-approval mode and interactive MCP auth appear with usage credit details.
- •Installer and Code-Mode Fixes Standalone installer issues and code-mode host exposure are resolved in the CLI.
- •Guardian Policy Rollback The previous Guardian auto-review policy and request format are restored after a regression.
- •Dangerous Command Detection Codex CLI 0.144.5 adds more forced rm forms and clearer rejection messages.
- •CLI 0.144.6 Metadata GPT-5.6 model context windows are corrected to 272,000 tokens with refreshed instructions.
What this means for you
For Vibe Builders: You can now launch and steer agents from your phone with Cursor iOS and Codex Remote while using Record & Replay to turn repeated workflows into reusable skills. Side chats and manual review toggles in Claude Code give you safer ways to explore ideas without breaking main threads. Test the iOS betas and remote pairing this week to move daily tasks off the desktop.
For Non-techies: SMB owners gain phone-based control over coding agents through Cursor iOS and Codex Remote so tasks can start or continue without sitting at a computer. Slack plans and multi-repo support reduce back-and-forth when teams review work. Try the public betas to see how mobile oversight changes daily AI use.
For Developers: Session limits, EndConversation, and sandbox toggles in Claude Code plus Bedrock login and import tools in Codex CLI give you concrete controls for production reliability. Cursor Router and multi-agent V2 options let you benchmark cost versus capability across environments. Evaluate the new permission prompts and subagent caps against your current stack before rolling out.
What to watch next
Watch Cursor 3.12 for deeper iOS agent hooks and Claude Code 2.1.219 for expanded screen-reader and vim features. Track Codex CLI 0.146 for additional import paths and multi-folder project polish. Check release notes mid-week for any surprise model routing or safety updates.
Harsh’s take
The week shows three vendors racing to lock users into their mobile and team surfaces rather than pure model quality. Cursor pushes phone control hardest while Claude Code focuses on safety rails that prevent loops and Codex emphasizes cross-app continuity. The practical effect is more fragmentation: each environment now requires its own import step and permission model.
Builders should pick one primary environment this week and run a side-by-side test of the new remote or subagent features against their existing workflow. Measure actual session success rate and time-to-first-useful-output instead of feature checklists. Only then decide whether switching surfaces is worth the migration cost.
by Harsh Desai
More AI news
- Daily RoundupFlux 3 video with audio, Vercel MCP deploys code, and fresh agent tools
Vendors added deployment and evaluation tools while new models and agent products appeared across gateways and marketplaces.
- LaunchRunway launches Runway Media Router through Runway Dev platform
Runway launched Runway Media Router on its Runway Dev platform. The router provides API access to third-party image, video and audio models.
- FeatureDependabot adds three-day cooldown before version updates
Dependabot now defaults to a three-day cooldown that delays version update pull requests.