Cursor Router and iPad app, Claude Code 2.1 series with Opus 5, Codex CLI 0.147 and Daybreak tiers (agent tools for builders)

What shipped

From 7 to 14 August, three major coding agents released overlapping updates focused on routing, mobile reach, and security controls. Cursor emphasized model selection and workspace plugins. Claude Code pushed frequent version bumps with new sandbox rules. OpenAI Codex introduced tiered access and desktop integration.

Cursor

Cursor released repeated updates to its Router for dynamic model selection across cost and intelligence modes. The company also pushed an iPad client with PR review tools, Google Workspace plugins, and a localized Start plan for India at 649 rupees per month. Slack integration gained multi-repo planning support.

• •Cursor Router in Auto mode Cursor Router now evaluates prompt complexity and routes tasks across frontier or efficient models using Cost, Balance, or Intelligence presets.
• •Cursor Start plan for India Cursor Start launched at 649 rupees monthly with UPI billing, giving Indian developers daily model access, Cloud Agents, and iOS remote control.
• •Cursor for iPad release The iPad app added split-screen PR reviews, Apple Pencil markup, multi-agent sidebar chats, and a shared task inbox with iPhone.
• •Google Workspace plugins New plugins let Cursor agents search Drive files, draft Gmail messages, and schedule Calendar events without leaving the editor.
• •Background builds for Cloud Agents Cursor now spins up ready environments ahead of time so Cloud Agents start 10 times faster and reach first token three times quicker.
• •Slack integration upgrade Cursor in Slack now posts task plans first, handles multiple repositories, and works across channels and threads.
• •Router with team admin controls The updated Router adds admin settings that let teams lock optimization modes for consistent routing behavior.
• •India Start subscription details The 649-rupee plan bundles Grok 4.5, Composer, cloud agents, and marketplace plugins with local payment support.
• •iPad PR review features Paid users on iPad get pinned sidebar chats, full PR code review, and an Inbox that tracks agent tasks across devices.
• •Google Drive and Gmail access Cursor agents can now read and edit Drive documents or handle email directly through marketplace plugins.
• •Multi-repo Slack planning Slack users can switch repositories mid-prompt and receive execution plans before any code runs.
• •Router complexity analysis Cursor Router selects models by scanning task demands and offers three preset optimization modes with admin overrides.
• •Localized India pricing launch Developers in India receive the Start plan at 649 rupees via UPI, including Grok 4.5 and mobile remote control.
• •iPad full review suite The iPad client supports multi-PR sessions, Apple Pencil markup, and unified task tracking on every paid plan.
• •Calendar and email plugins Agents inside Cursor can update schedules and manage Gmail using the new Google Workspace marketplace entries.
• •Slack cross-channel context The integration now gathers context from multiple channels and renders messages more cleanly before running tasks.
• •Auto mode model routing Router powers Auto mode by matching task complexity to the best available model under Cost, Balance, or Intelligence settings.
• •UPI billing for Start plan Indian users pay 649 rupees monthly through UPI for access to cloud agents, Composer, and the iOS app.
• •iPad layout optimizations Screen-specific layouts, multi-PR support, and an Inbox for background tasks arrived on all paid iPad plans.
• •Drive file management inside Cursor Agents search, edit, and organize Google Drive content without switching to a browser tab.

Claude Code

Claude Code issued over forty updates across versions 2.1.221 to 2.1.229, adding self-hosted runners, cross-session messaging, and Focus view in VS Code. Claude Opus 5 became the default model with a 1M context window and strict sandbox network rules. Marketplace wildcards, spend-limit warnings, and subagent forking changes also shipped.

• •Marketplace owner wildcards Owner wildcard entries now let teams manage marketplace settings without listing every user individually.
• •Self-hosted runners for teams Team and Enterprise plans gained self-hosted runner environments plus cross-session SendMessage support.
• •Gateway spend-limit warnings Usage warnings now include reset details and workspace trust prompts for agents in untrusted directories.
• •Stability and bug fixes Multiple releases improved process stability, reduced terminal crashes, and fixed streaming display issues.
• •Subscription tier evaluation fix Feature flags now evaluate correctly even when sessions start with expired tokens.
• •Synced skills hardening Skills pulled from claude.ai are now sanitized to block command shadowing or unauthorized execution.
• •SSE keepalives added Long thinking pauses send keepalive pings to prevent upstream timeouts on Vertex and Bedrock.
• •MCP OAuth redirect fix Sign-in now succeeds when servers use pre-registered OAuth clients such as Slack.
• •Subagent forking default on Subagent forking inherits full conversation and cache data; @mentions let users reference other sessions.
• •Claude Code 2.1.221 update Focus view in VS Code, file-based credential masking, and prompt auditing tools arrived together.
• •Worktree isolation improvements Version 2.1.222 strengthened git worktree security and refined raw diff views.
• •Owner wildcard settings Version 2.1.223 added marketplace owner wildcards and context window enforcement updates.
• •Self-hosted runner support Version 2.1.224 introduced runners, zip archive plugins, and removed the 200-subagent limit.
• •Spend-limit and Focus view Version 2.1.225 added gateway spend warnings and VS Code Focus view folding.
• •Maintenance patch 2.1.226 General reliability fixes addressed session handling across operating systems.
• •Slash-command menu polish Version 2.1.227 improved menu styling and reduced event-loop stalls during file checks.
• •Cross-session messaging fixes Version 2.1.228 resolved interactive freezes and improved Vertex credential handling.
• •Remote control documentation Version 2.1.229 added hook support for self-hosted runners and SSE keepalives.
• •Claude Opus 5 default Opus 5 now ships with a 1M context window, fast mode, and strict sandbox allowlists.
• •General reliability updates Bug fixes targeted background session glitches and unexpected failures on multiple platforms.
• •VS Code Focus view toggle Per-turn tool summaries can now be collapsed behind a single view in VS Code.
• •Git worktree isolation Destructive commands are now isolated across subagents to protect main checkouts.
• •Marketplace org wildcards Wildcard entries simplify marketplace management for organizations.
• •Self-hosted runner launch Web, mobile, and desktop runners support zip archive plugin sources.
• •Workspace trust prompts Agents now prompt before acting in untrusted directories.
• •Maintenance release fixes Stability patches improved session resume and compaction handling.
• •Slash command performance Menu character matching and event-loop speed improved after subscription checks.
• •Interactive redrawing fixes Windows path issues and skill isolation received targeted patches.
• •Background code review subagents /code-review now runs as a background subagent with screen-reader support.
• •Opus 5 1M context window The new default model adds nested subagent forwarding up to depth three.
• •Credential masking for sandboxes Linux and WSL sandboxes gained file masking mode for secrets.
• •Worktree isolation enforcement Auto mode safety checks now cover cross-session messaging.
• •Owner wildcard marketplace entries Managed settings accept wildcards to reduce manual configuration.
• •Archive plugin sources via HTTPS Zip files can now be installed as plugin sources.
• •Gateway spend-limit details Warnings display cap reset timing for Team and Enterprise users.
• •Feature flag evaluation fix Expired login tokens no longer block correct tier detection.
• •Background subagent reviews Code review commands execute without blocking the main session.
• •Opus 5 fast mode support Strict network allowlists accompany the new model rollout.
• •Focus view per-turn summaries Tool activity is aggregated behind collapsible summaries in VS Code.
• •Git worktree security patches PreToolUse hook issues and raw blob diff views were corrected.
• •Bash permission bypass fix Command security validation gaps received targeted patches.
• •Cross-session SendMessage Sessions on different machines can now exchange messages directly.
• •Spend-limit usage warnings Gateway caps now surface reset details inside the CLI.
• •System reliability improvements Internal fixes addressed background runtime stability across releases.

OpenAI Codex

OpenAI Codex released CLI 0.147.0 with portable agent plugins, persistent conversation sections, and Cursor skill imports. Daybreak Blue and Daybreak Red tiers gave approved defenders access to GPT-5.6 Sol and Cyber models. Desktop and iOS apps gained multi-folder projects, voice steering, and inline Mermaid support.

• •Codex CLI 0.147.0 release Portable agent plugins, persistent sections, and --approve-for-me flag arrived with MCP 2026-07-28 support.
• •Daybreak Blue defensive tier Approved users access GPT-5.6 Sol for defensive workflows under the new Daybreak program.
• •External agent import command /import now pulls settings, skills, and projects from Claude Code or Cursor into Codex.
• •Linux desktop app preview ChatGPT desktop app reached Ubuntu, Debian, and Fedora on both x64 and ARM64 via deb and rpm packages.
• •Computer History on macOS Pro and Enterprise users can opt in to capture app and web activity into searchable timelines.
• •CLI 0.147.0 portable plugins Agent plugins can now move between machines while keeping conversation sections intact.
• •Daybreak Red red-team tier Approved defenders receive GPT-5.6 Cyber for authorized exploit validation.
• •Agent plugin portability Skills and plugins import from Cursor and Claude Code via the new /import command.
• •Daybreak security access tiers Blue and Red tiers separate general defensive work from specialized cyber tasks.
• •iOS task management update Version 1.2026.181 lets users manage Codex tasks inside conversations with git diff filters.
• •Desktop app Codex integration Inline Markdown editing and PR Chat for GitHub reviews landed in the macOS and Windows apps.
• •iOS inline visualizations Version 1.2026.188 added progress indicators and composer control visibility for Codex tasks.
• •Mermaid diagram support Version 1.2026.195 introduced interactive Mermaid diagrams inside Codex tasks on iOS.
• •ChatGPT Voice in desktop app Voice steering powered by GPT-Live now works across multiple project threads.
• •iOS voice and autocomplete Version 1.2026.202 improved voice selection limits and plugin mention autocomplete.
• •Sign in with ChatGPT beta Partner sites including Airtable, GitLab, and Vercel now accept ChatGPT sign-in.
• •Desktop multi-repo review Address bar history and multi-repository diff inspection arrived in update 26.727.
• •GPT-5.4 retirement notice GPT-5.4 and mini will stop working in Codex on 31 August; switch to GPT-5.6 Terra or Luna.
• •Daybreak access tiers rollout Approved security defenders can now request Blue or Red model access for engagements.

What this means for you

For Vibe Builders: You can now route tasks automatically in Cursor, run reviews on iPad, and import skills from other agents into Codex without writing code. Claude Code stability fixes mean fewer session crashes when you chain subagents. Start with the Cursor Start plan if you are in India or test Daybreak tiers if security workflows matter for your projects.

For Non-techies: SMB owners gain easier access to agent tools through Cursor iPad support and Codex desktop integration that handles PR reviews and voice steering. India pricing at 649 rupees lowers the barrier for daily model use. Watch for Daybreak tiers if your work involves security reviews or compliance checks.

For Developers: Production teams should benchmark Cursor Router against Claude Code subagent forking and Codex portable plugins before standardizing on one stack. Self-hosted runners and spend-limit warnings in Claude Code 2.1.22x series provide clearer controls for enterprise usage. Retire GPT-5.4 references in Codex configs before the 31 August cutoff and test MCP 2026-07-28 protocol support in the latest CLI.

What to watch next

Track whether Cursor Router reaches parity with Claude Code subagent depth and whether Codex desktop voice features expand to more languages. Monitor adoption of Daybreak tiers and self-hosted runners for signals on enterprise readiness.