Codex Gains Remote Control as Cursor and Lovable Push Agent Tools

What shipped

On 24 May, AI vendors focused on agent workflows and cross-platform access. OpenAI shipped the largest set of Codex updates including mobile control and spreadsheet sidebars. Cursor, Lovable, and smaller Mac tools added automation layers that let non-coders and engineers move tasks out of chat windows into live apps and repos.

Hugging Face trending

Three new models appeared on Hugging Face Hub with strong download momentum. OBLITERATUS led with a text model, Stability AI followed with audio generation, and CohereLabs placed an image-text model in the same list. Builders can test these directly in the hub without new infrastructure.

• •Qwen3.6-27B-OBLITERATED OBLITERATUS published a text generation model on Hugging Face that users can download and fine-tune for custom chatbots. It gives vibe builders a ready starting point instead of training from scratch.
• •stable-audio-3-medium Stability AI released a text-to-audio model on the hub that converts prompts into sound clips for video or podcast work. Non-technical users can generate tracks without separate audio software.
• •command-a-plus-05-2026-bf16 CohereLabs added an image-text-to-text model that handles combined visual and text queries. Developers can run it for document analysis tasks that mix screenshots and instructions.

Product Hunt picks

Five Mac and video tools launched on Product Hunt. DockFlow and Freu AI target workflow switching, ModelHub focuses on local models, Runway Agent handles video edits, and Stitch 3.0 offers UI generation on a canvas. These give immediate install options for daily use.

• •Freu AI Freu AI automates any Mac app through a CLI with zero recurring fees, letting users script clicks and data moves without monthly SaaS bills.
• •ModelHub ModelHub adds a menu bar app that runs local LLMs on Mac and keeps models accessible without terminal commands. It simplifies switching between models for quick tests.
• •Runway Agent Runway Agent edits and adds sound to videos through chat prompts, replacing separate timeline tools for short marketing clips.
• •Stitch 3.0 Stitch 3.0 generates and iterates UI screens on a live canvas inside Google tools, giving designers a direct path from prompt to editable mockup.

Vendor launches

OpenAI delivered twelve Codex and ChatGPT updates that cover remote sessions, financial data views, and spreadsheet sidebars. The changes expand Codex from chat helper to background task runner while adding safety and memory controls for personal accounts.

• •Codex richer context and Goal Mode Codex added macOS window screenshots, a released Goal Mode, browser annotations, and remote locked computer use for secure background jobs that run without the screen on.
• •ChatGPT personal finance links Pro users in the US can now connect bank accounts through Plaid inside ChatGPT to track spending and ask budget questions on web or iOS.
• •Codex mobile remote access Users can monitor and approve Codex sessions on macOS from iPhone or Android without returning to the desktop.
• •File Library for Free and Go tiers ChatGPT opened its File Library to free and Go plans with storage caps from 500 MB to 100 GB depending on the plan.
• •Inline web images for Free users Free users now see more web images placed next to answers when using the 5.5-Instant model on visual topics.
• •Trusted Contact safety feature Personal account holders can name a contact who receives alerts if safety systems flag self-harm concerns.
• •Memory personalization and sources Plus and Pro users gain better recall from past chats plus a tool to inspect and edit what data shaped each reply.
• •GPT-5.5 Instant default model GPT-5.5 Instant became the new default, delivering higher accuracy and better image handling than the prior 5.3 version.
• •ChatGPT Excel and Sheets sidebar ChatGPT now appears as a sidebar in Excel and Google Sheets worldwide to help write formulas and clean data.
• •Codex Appshots and remote use on macOS Codex now sends active window screenshots for context, runs Goal Mode on locked Macs, and improves browser reliability for long tasks.
• •Codex CLI 0.133.0 release The CLI update turns goals on by default, adds storage tracking, and fixes TUI startup bugs for smoother remote control.

Other

Cursor shipped five automation and Composer updates, Lovable released ten features including mobile apps and an MCP server, and OpenClaw added ten SDK and performance fixes. The combined releases give builders more ways to run agents across repos, phones, and external clients.

• •Cursor Automations in Agents Window Cursor now manages automations inside the Agents Window with multi-repo support and five new marketplace templates.
• •Cursor Jira integration Cursor can receive Jira tickets and update them with pull requests after completing scoped tasks from the ticket text.
• •Composer 2.5 in Cursor Composer 2.5 raises intelligence and instruction following for long tasks and offers standard plus fast pricing tiers.
• •Cursor 3.4 Agents Window upgrades Cursor 3.4 adds full-screen tabs and compact chat views to reduce clutter during agent runs.
• •Cursor cloud agent environments Teams can define Docker-based environments with secrets and caching for cloud agents that span multiple repos.
• •Lovable branded app URLs Business workspaces can publish apps under custom subdomains instead of the default Lovable domain.
• •Lovable Google Maps and Semrush connectors Lovable apps can now pull maps data, SEO keywords, and TikTok metrics through three new connectors.
• •Gemini 3.5 Flash in Lovable Lovable switched its built-in AI to Gemini 3.5 Flash for faster coding and agent steps inside projects.
• •Lovable SEO and AI Search tab A new tab gathers sitemap checks, speed audits, and keyword research in one place for site owners.
• •Lovable Design Guidance Lovable generates three visual directions or asks guided questions before any code is written.
• •Lovable Telegram chat Users can message Lovable on Telegram to build apps, run SQL, or check analytics without opening a browser.
• •Lovable MCP Server preview External clients such as Claude Desktop can now control Lovable projects through the new MCP server.
• •Lovable mobile app release The Lovable iOS and Android apps let users start projects and view analytics from a phone.
• •Lovable Workspace Skills Reusable markdown playbooks teach Lovable recurring tasks such as launch checklists across a workspace.
• •Lovable Wiz security scan Wiz now runs SCA and SAST scans inside Lovable's Security view for every project.
• •OpenClaw Gateway performance fix OpenClaw cut model-listing latency from twenty seconds to five milliseconds by reusing channel reads and pre-warming auth state.
• •OpenClaw external meeting-notes plugin A new plugin captures meeting audio from Discord and allows read-only CLI access to transcripts.
• •OpenClaw embeddingProviders contract The Plugin SDK now registers embedding providers as a reusable surface beyond memory adapters.
• •OpenClaw chat session picker search The Control UI added search and pagination so older chat sessions remain reachable without full reloads.
• •OpenClaw Grok OAuth and search Grok integration now reuses OAuth profiles for web search and adds model aliases with default timeouts.
• •OpenClaw row-level session helpers Plugins can read and patch sessions directly without loading the entire store shape.
• •OpenClaw npm shrinkwrap Published packages now ship with shrinkwrap to lock dependency graphs for reproducible installs.
• •OpenClaw Claude 4.x migration One-million-token handling moved to generally available Claude 4.x models and dropped the retired beta flag.
• •OpenClaw public memory artifacts Memory files and daily notes now surface through the active memory bridge without internal core access.
• •OpenClaw video edit routing fix Video edit requests now hit the documented endpoint and respect private-network settings.

Industry news

Reports covered Anthropic model supply questions, automated algorithm discovery, and default model pitfalls in Copilot and Gemini. Additional notes addressed ByteDance training methods, Amazon wearables, and real-time AI security gaps at large providers.

• •AutoTTS scaling algorithm discovery Researchers used a coding agent to find a control algorithm that cuts compute by seventy percent while matching accuracy and cost only forty dollars to run.
• •Default model selection warning Microsoft Copilot produced country stereotypes on identical datasets with swapped labels, showing that users must manually pick thinking models for data work.
• •ByteDance long-document training study A seven-billion model answered questions on image-heavy documents more reliably than larger models when trained on questions instead of full transcriptions.
• •Armin Ronacher issue quality note GitHub issues written by AI often contain confident but inaccurate root-cause guesses that waste maintainer time.
• •Usborne Creepy Computer Games remake Simon Willison fed a 1983 PDF into Claude to recreate the Mad House game as an interactive web version.
• •Google AI security transition note Google and other providers are still defining live security practices for agent systems in production.

What to watch next

Watch for wider rollout of Codex remote computer use and any new Claude 4.x context limits. Track whether Lovable MCP reaches stable release and whether Cursor adds more Jira-style connectors. Check Hugging Face for follow-up fine-tunes of the three trending models.

Sources

Hugging Face trending

• •Qwen3.6-27B-OBLITERATED by OBLITERATUS trends on HuggingFace
• •stable-audio-3-medium by stabilityai trends on HuggingFace
• •command-a-plus-05-2026-bf16 by CohereLabs trends on HuggingFace

Product Hunt picks

• •DockFlow launches on Product Hunt
• •Freu AI launches on Product Hunt
• •ModelHub launches on Product Hunt
• •Runway Agent launches on Product Hunt
• •Stitch 3.0 by Google launches on Product Hunt

Vendor launches

• •Enhance Codex with richer context, goal mode, and remote locked use
• •Launch Appshots, Goal Mode, and Remote Computer Use on macOS

Other

• •Introduce Cursor Automations in the Agents Window
• •Configure branded app URLs for Business and Enterprise workspaces
• •Optimize Gateway startup and model-listing performance

Industry news

• •Anthropic may keep supplying Claude to the NSA despite being flagged as a supply chain risk by the Pentagon
• •Researchers let Claude Code discover AI scaling algorithms that humans probably wouldn't have designed
• •Why you shouldn't leave model selection on default in Copilot, Gemini and other AI tools
• •ByteDance study finds that asking LMMs questions beats making it transcribe text for long document training
• •I tried Amazon’s Bee wearable and am both intrigued and slightly creeped out
• •Quoting Armin Ronacher
• •Mad House: Usborne Creepy Computer Games
• •Everyone is navigating AI security in real time: even Google