Cursor 3.8 Automations, Claude Code v2.1.178 teams, Codex Remote GA (agent tools for builders)

What shipped

Across this week Cursor, Claude Code, and OpenAI Codex released updates centered on agent management, permission controls, and workflow automation. The changes target faster reviews, safer command execution, and easier handoffs between local and remote environments. Builders now have more concrete options to run agents without constant manual oversight.

Cursor

Cursor delivered twenty updates focused on Design Mode, Bugbot reviews, cloud agents, and automations. The releases emphasize multi-select editing, three-times-faster reviews, and under-ten-minute cloud setup with isolated subagents. These features let users trigger agents from Slack or GitHub and manage plugins through a new marketplace page.

• •Cursor Design Mode Cursor Design Mode now supports multi-select and voice input so agents can edit multiple UI elements at once while users narrate changes during runs.
• •Bugbot speed update Bugbot runs over 3x faster and 22% cheaper with Composer 2.5, letting developers run local /review checks before pushing to GitHub or GitLab.
• •Cursor 3.7 Agents Window Cursor 3.7 adds cloud setup in under 10 minutes, /in-cloud subagents, and local-to-cloud session handoffs for long-running CI or PR tasks.
• •Cursor 3.8 /automate skill Cursor 3.8 introduces the /automate skill with Slack emoji triggers, five GitHub triggers, and a computer use tool for local agent sessions.
• •Cursor Customize page Cursor added a Customize page with plugin marketplace, prebuilt canvases for Hex and Atlassian, and import options from GitLab or Azure DevOps.
• •Design Mode multi-select Design Mode now lets users multi-select elements to match styles across component groups and keep voice input active during agent work.
• •Bugbot local reviews Developers can run Bugbot and Security Review locally with the /review command before pushing code, syncing results with GitHub and GitLab.
• •Cloud snapshots Cursor supports reusable snapshots that set up cloud environments in under 10 minutes and run isolated /in-cloud subagents with session handoff.
• •Cursor Automations triggers Cursor Automations now include Slack emoji triggers, five GitHub triggers, and a computer use tool so cloud agents can generate demos automatically.
• •Customize marketplace The new Customize page shows a leaderboard for plugins and skills plus prebuilt canvases and team marketplace imports from multiple hosts.
• •SDK custom tools TypeScript and Python SDKs now support local.customTools, auto-review configs, JSONL stores, and nested subagents up to several levels deep.
• •Voice input in Design Mode Continuous voice input lets users queue UI edits without pausing agent runs, while multi-select handles group style changes.
• •Composer 2.5 Bugbot Bugbot powered by Composer 2.5 finds 10% more bugs, runs three times faster, and supports pre-push local reviews via the /review command.
• •/in-cloud subagents The /in-cloud command launches isolated background agents for tasks like fixing CI while sessions move between local machines and cloud VMs.
• •/automate skill rollout The /automate skill lets users configure agents in plain language with Slack and GitHub triggers plus computer use for cloud agents.
• •SDK nested subagents Updated SDKs add custom tools, auto-review routing, and support for deeply nested subagents plus JSONL metadata storage.
• •Multi-select UI edits Design Mode multi-select now matches styles across groups and voice input queues instructions without waiting for prior agent runs to finish.
• •Pre-push Bugbot reviews Users can configure Bugbot to review only new changes and run reviews locally before pushing with the /review command.
• •Cursor 3.7 cloud config Cursor 3.7 enables fast cloud environment setup and /in-cloud subagents that run background tasks with seamless local-to-cloud handoffs.
• •Slack and GitHub triggers Cursor Automations adds the /automate skill plus Slack emoji and five GitHub triggers for local agent configuration.

Claude Code

Claude Code issued fifty updates centered on agent teams, auto-mode safety, and terminal fixes. Releases block destructive commands unless requested, add parameter-based permissions, and introduce /config key=value syntax plus MCP login commands. The changes improve reliability for WSL2, streaming, and subagent handling across many point releases.

• •Agent teams simplification Claude Code v2.1.178 removes TeamCreate tools in favor of implicit teams and adds parameter-based permission rules for subagents.
• •Mid-stream drop fixes Version 2.1.179 preserves partial responses during connection drops and fixes mouse-wheel scrolling in WSL2 sessions.
• •Inline config syntax Version 2.1.181 adds /config key=value syntax and upgrades the bundled Bun runtime to 1.4 for faster long-paragraph streaming.
• •Auto-mode safety blocks Version 2.1.183 blocks git reset --hard and terraform destroy commands unless explicitly requested by the user.
• •Stream-stall hint change Version 2.1.185 updates the stall message to Waiting for API response and triggers after 20 seconds instead of 10.
• •MCP CLI commands Version 2.1.186 adds claude mcp login and logout commands for headless environments and improves subagent permission prompts.
• •Sandbox credential setting Version 2.1.187 introduces sandbox.credentials to block sandboxed commands from reading secrets and adds org model restrictions.
• •Conversation rewind Version 2.1.191 adds /rewind to resume from before a /clear and improves MCP server reliability with automatic retries.
• •Auto-mode shell classify Version 2.1.193 adds autoMode.classifyAllShell to route all commands through the classifier and OpenTelemetry logging.
• •Model enforcement setting The enforceAvailableModels managed setting constrains default models to the allowed list in managed environments.
• •Localized session titles Session titles now generate in the conversation language and Bedrock credential caching extends based on expiration.
• •Implicit agent teams Team creation tools are replaced with implicit teams and parameter-level permission rules improve auto-mode safety checks.
• •WSL2 scroll fix Partial responses are preserved on drops, mouse-wheel scrolling works in WSL2, and only one promotional banner shows per session.
• •Bun 1.4 upgrade The /config syntax sets options from the prompt, sandbox Apple Events are supported, and streaming for long paragraphs improves.
• •Destructive command blocks Destructive git and terraform commands are blocked in auto mode with model deprecation warnings and new help commands.
• •Stall hint timing The hint now reads Waiting for API response and waits 20 seconds before showing instead of 10 seconds.
• •MCP authentication CLI CLI commands for MCP login and logout are added along with status filtering in workflows and automatic bash responses.
• •Sandbox secret blocking The sandbox.credentials setting prevents reading secrets and org model restrictions appear in the model picker.
• •General reliability fixes Version 2.1.190 delivers bug fixes and reliability improvements across the terminal agent experience.
• •Scroll acceleration toggle Version 2.1.174 adds a setting to disable mouse-wheel scroll acceleration and fixes model picker visibility for some plans.
• •EnforceAvailableModels Version 2.1.175 introduces the enforceAvailableModels setting to keep default models inside the allowed list.
• •Multilingual titles Version 2.1.176 generates titles in the conversation language and improves Bedrock credential caching.
• •Simplified teams update Version 2.1.178 removes setup tools, adds parameter permission rules, and improves pre-launch subagent safety checks.
• •Connection handling Version 2.1.179 preserves partial responses on drops and fixes plugin loading in remote sessions.
• •Inline config release Version 2.1.181 adds /config key=value, upgrades Bun to 1.4, and improves streaming performance.
• •Auto mode safety Version 2.1.183 blocks destructive commands and adds deprecation warnings for older models.
• •Stream-stall update Version 2.1.185 changes the stall message and extends the silence threshold to 20 seconds.
• •MCP login commands Version 2.1.186 adds CLI MCP authentication and status filtering for workflows.
• •Credential security Version 2.1.187 blocks sandbox commands from reading secrets and adds org model restrictions.
• •Fable 5 normalization Version 2.1.173 normalizes Fable 5 model names by stripping context suffixes on Windows.
• •VS Code attribution Version 2.1.174 adds detailed usage attribution in VS Code and disables scroll acceleration.
• •Managed model rules Version 2.1.175 constrains default model resolution to the allowed models list.
• •Bedrock caching Version 2.1.176 improves Bedrock credential caching and fixes background session bugs.
• •Agent team changes Version 2.1.178 streamlines teams and adds Tool(param:value) syntax for permission rules.
• •Partial response save Version 2.1.179 keeps partial responses during drops and fixes subagent focus issues.
• •Bun runtime upgrade Version 2.1.181 upgrades Bun to 1.4 and adds sandbox.allowAppleEvents for macOS.
• •Command safety blocks Version 2.1.183 blocks destructive git and pulumi commands unless requested.
• •Stall message tweak Version 2.1.185 updates the stall hint to Waiting for API response and extends timing.
• •Workflow status filter Version 2.1.186 adds status filtering and teammateMode settings for workflows.
• •Safe mode flag A --safe-mode flag disables customizations for troubleshooting and /cd changes directories mid-session.
• •Fable 5 model Claude Fable 5, a Mythos-class model safe for general use, is now available.
• •Nested sub-agents Sub-agents can now spawn up to five levels deep with marketplace search and reduced idle CPU.
• •Scroll toggle A setting disables mouse-wheel scroll acceleration and adds detailed VS Code usage attribution.
• •Language titles Titles generate in the conversation language and Bedrock caching improves.
• •Permission rules Parameter-based rules replace TeamCreate tools and improve auto-mode safety evaluation.
• •Drop handling Partial responses are preserved on mid-stream drops and WSL2 scrolling is fixed.
• •Config syntax /config key=value sets settings from the prompt and Bun 1.4 improves streaming.
• •Stall timing The stall hint now waits 20 seconds and shows a retry message.
• •Fable 5 launch Claude Fable 5 brings advanced Mythos-class capabilities safe for general use.

OpenAI Codex

OpenAI Codex released twenty-nine updates covering CLI 0.142, ChatGPT iOS features, Sites preview, and Record & Replay. Releases add usage credit redemption, Windows proxy support, Face ID lock, and migration flows from Claude Code. Remote control, branch management, and LaTeX rendering expand day-to-day coding options across mobile and desktop.

• •Codex CLI 0.142.0 CLI 0.142.0 adds usage credit redemption, remote plugin organization, token budgets, and multi-agent delegation controls.
• •iOS per-host personalities ChatGPT iOS 1.2026.167 adds per-host personality settings and direct goal editing in the composer.
• •Windows proxy support CLI 0.142.1 introduces opt-in Windows system proxy support including PAC and WPAD rules.
• •MCP tool search CLI 0.142.2 enables default MCP tool search and macOS system proxy support.
• •Codex Remote GA Codex Remote reaches general availability with secure QR pairing and a DigitalOcean plugin for remote workspaces.
• •Face ID lock ChatGPT iOS adds optional Face ID or passcode lock plus a settings screen for follow-up behaviors.
• •Sites preview Sites plugin enters preview so users can create, deploy, and inspect hosted websites and dashboards.
• •Activity insights Profile section gains activity insights and share cards for reviewing Codex usage highlights.
• •Branch and worktree support ChatGPT iOS adds branch choice, worktree creation, and a Codex profile screen with usage stats.
• •Migrate from Claude Migration flows import setups from Claude Code and Claude Cowork into Codex.
• •Rate-limit banking Rate-limit reset banking and Developer mode for Browser use with CDP access are added.
• •Workspace file browser ChatGPT iOS adds a workspace file browser, directory picker, and LaTeX rendering.
• •EEA rollout Computer Use, Chrome extension, and Memories now reach users in EEA, UK, and Switzerland.
• •Record & Replay Record & Replay on macOS turns workflows into reusable skills with bulk actions for history.
• •Personality settings Per-host Friendly and Pragmatic personalities plus goal editing appear in the composer.
• •Usage redemption CLI 0.142.0 restores Linux TUI rendering and adds usage credit redemption via /usage.
• •iOS 1.2026.167 Version 1.2026.167 improves autocomplete and fixes Face ID unlocking issues.
• •Bedrock integration Codex now runs OpenAI models through Amazon Bedrock with AWS-managed authentication.
• •iOS 1.2026.146 Version 1.2026.146 adds Face ID lock and Windows SSH connection support.
• •Sites plugin Sites preview lets users build and deploy websites with environment variable management.
• •Activity cards Version 26.602 adds share cards and improves Computer Use startup readiness.
• •Profile stats Version 1.2026.153 adds branch support and inline review comments on the profile screen.
• •Plugin revamp Version 26.608 revamps the plugins screen with category filters and expanded Settings search.
• •Developer mode Version 26.609 adds rate-limit banking, /init command, and customizable macOS Dock icons.
• •File browser Version 1.2026.160 adds workspace file browser, diff controls, and MCP approval choices.
• •Regional expansion Computer Use and Memories roll out to EEA, UK, and Switzerland users.
• •Record & Replay bulk Version 26.616 adds bulk actions for automation history and local-remote thread handoff.
• •Noise relay upgrade CLI 0.141.0 upgrades remote executors to encrypted Noise relay channels.
• •Workflow recording Record & Replay turns macOS workflows into reusable skills with thread handoff support.

What this means for you

For Vibe Builders: You can now trigger agents from Slack emojis or GitHub events in Cursor, set parameter-based rules in Claude Code, and migrate setups from Claude into Codex without writing code. These changes let non-coders run reviews locally, hand off sessions to the cloud, and replay recorded workflows on macOS. Test the /automate skill or Record & Replay this week to ship a small automation before the next release cycle.

For Non-techies: For your business, AI coding tools moved from chat to action this week with Face ID locks on iOS, Sites for quick website previews, and safer auto-mode blocks on destructive commands. SMB owners can now review activity insights, connect Windows machines over SSH, and migrate existing Claude setups into Codex. Start with the Sites preview or rate-limit banking to handle routine tasks without extra developer time.

For Developers: On the platform side, Cursor SDK nested subagents, Claude Code parameter permissions, and Codex multi-agent delegation plus Noise relays point to more reliable local-first agent stacks. Evaluate the new /review command and enforceAvailableModels setting against your current CI and Bedrock workflows before integrating. Watch the next Codex CLI and Claude Code point releases for further MCP and proxy refinements.

What to watch next

Watch for expanded MCP tool search defaults in Codex CLI, deeper Fable 5 usage in Claude Code, and additional Sites environment controls. Track whether Cursor marketplace leaderboards drive plugin adoption and whether Claude Code auto-mode denial reasons reduce support tickets.