Orchestrator v0.14.6: switch agents mid-conversation
Move a thread from Claude Code to Codex without starting over, resume automatically when a usage limit resets, and merge PRs or comment on diffs from your phone.
Source: Orchestrator on X →New this week
Notable AI coding launches and updates from roughly the last 30 days (research cut-off 2026-10-27). Every item has a primary source link. We do not invent dates.
baton: hand a session to another agent when you hit a limit
Open-source (MIT) tool that moves one coding session between Kiro, Claude Code, Codex and Cursor with the task, plan and files intact and secrets redacted. Early, tested on macOS.
Source: launch post on X →Cursor Agent SDK is available
Define agents as code, with prompts in Markdown and tools in TypeScript, deploy them on Cursor's hosted infrastructure or your own, and measure every change with built-in evals.
Source: Cursor on X →DeepSeek V4.1 Flash lands in Cursor
Cursor says every effort level sits on its CursorBench cost frontier, at $0.60 to $0.98 per task. Turn it on in Cursor Settings, Models.
Source: Cursor on X →GPT-6.1 Terra in GitHub Copilot, coming to the API and Codex
GPT-6.1 Terra is generally available in GitHub Copilot, and OpenAI says it's coming to the API and Codex as a fast, lower-cost model for computer use and high-volume workflows.
Source: GitHub on X →Devin Security relaunched
Cognition's new Devin Security uses Claude Mythos and frontier cyber models to help teams find, prioritise and fix urgent vulnerabilities.
Source: Cognition on X →Supabase: scoped Personal Access Tokens
You can now limit a Supabase access token to specific organisations, projects and permissions, so a CI pipeline or coding agent only gets the one project it needs.
Source: Supabase on X →Bolt adds four more free business website templates
New free templates for HVAC services, landscaping, general contractors and home cleaning, ready to customise in Bolt.
Source: bolt.new on X →Replit now works fully in Japanese
Replit's interface, prompts and output are now all available in Japanese.
Source: Replit on X →Boxi: disposable local microVMs for coding agents
Open-source sandboxes on your Mac for Claude Code, Codex and Cursor that start in 20 to 50ms, run fully locally and keep secrets out of the box.
Source: launch post on X →Cursor agents can now control your computer
Cursor agents can now drive desktop apps, such as the iOS Simulator, to test the features they build.
Source: Cursor on X →Refreshed Devin Desktop app with an agent command center
Cognition released a refreshed Devin Desktop app built around an agent command center for managing your agents.
Source: Cognition on X →Chroma is joining Replit
The team behind the Chroma vector database is joining Replit to build a context layer for agents.
Source: Replit on X →Claude Models on Google Cloud (Anthropic managed) is GA
The full Claude Platform feature set is now generally available on Google Cloud, using Google Cloud IAM credentials, billing and commitment retirement.
Source: Claude on X →Claude apps switch to auto mode by default next week
From next week, sessions in Claude web, desktop and mobile run in auto mode by default on Pro, Max and Team plans. An automated check blocks risky actions instead of asking you to approve each step, and you can switch back to manual anytime.
Source: Claude on X →Replit Monitoring Agent
An on-call agent for AI-app builders that can see, fix and deploy changes to your Replit app autonomously.
Source: Replit on X →OpenAI Decisions API is generally available
Classify text and images, evaluate inputs and decide what happens next in near real time, without generating a full model response. Pitched as faster and cheaper for apps and agents.
Source: OpenAI Developers on X →Supabase: give your app its own MCP server
Supabase apps can now expose their own MCP server, so your customers can use your app from ChatGPT or Claude with the account they already have.
Source: Supabase on X →Cognition open-sources a security agent plugin
An agent plugin that helps secure systems against recent cybersecurity threats. Install it in any agent or use it in Devin today.
Source: Cognition on X →Bolt Builders Week: free small-business website templates
Free templates for restaurants, home services, lawn care, and wedding and event planners that you can customise in Bolt.
Source: bolt.new on X →Security alert: LiteLLM 1.107.0 shipped a credential stealer
A malicious LiteLLM 1.107.0 was published to PyPI and Docker on 21 Oct (now pulled), so if you installed it, pin 1.106.2 and rotate your keys. What to do.
Source: LiteLLM incident report →Claude Security: continuous re-scans as your code changes
Connect your repositories once and Claude runs a deep first security scan, then re-scans as changes land so findings track the current state of your code. Public beta on Claude Enterprise.
Source: Claude on X →Replit Agent can incorporate your business via Stripe Atlas
Through the Stripe MCP integration, you can now ask Replit Agent to incorporate your company with Stripe Atlas without leaving your project.
Source: Replit on X →Replit: share a full Agent conversation with teammates
Share a Replit conversation with people in your workspace so they can read the work behind your findings, decisions and next steps instead of a pasted summary.
Source: Replit on X →v0 adds Slack MCP
v0 can now read Slack threads and files and post progress updates back to Slack, limited to channels your account can access.
Source: v0 on X →Supabase .NET SDK v8, now officially supported
The Supabase .NET SDK is now officially supported, with v8 putting every package in one repo on a single version number.
Source: Supabase on X →LunoPeak goes open source: local dashboard for AI coding sessions
LunoPeak, a local-first desktop dashboard that tracks sessions, agents, token costs and config across Claude Code, Codex and Cursor, is now MIT-licensed on GitHub, with v2.0 released the same day. No account or telemetry.
Source: Timothy Murphy on X →Hallmark v2: an anti-slop design skill for coding agents
Together AI's open-source Hallmark design skill hit v2.0 with 48 themes, redesign and audit commands, and a browser lint that flags the tells of AI-made pages (purple gradients, an eyebrow over every heading). Works in Claude Code, Codex, Cursor and OpenCode; install with npx skills add nutlope/hallmark -g. MIT-licensed.
Beacon: open-source shared memory for coding agents
Asymptote Labs launched Beacon alongside a $5.5M seed round. It captures sessions from Claude Code, Cursor, Codex, OpenCode and 20+ other agents and turns fixes and workflows that worked into reusable skills, so context follows you when you switch tools. Runs locally by default.
Source: Asymptote Labs on X →Vercel durable-harness: long-running coding agents
Vercel's new durable-harness runs Pi, OpenCode, Codex, Claude Code or your own agent harness durably on Vercel, so agents resume automatically after failures and can run for hours, days or months.
Source: Vercel on X →OpenAI API: 12-hour prompt caching
OpenAI now lets developers choose 12-hour prompt caching in the API, so persistent agents can keep shared instructions, tools and context cached and pay the lower cached-input rate even when requests are hours apart.
Source: OpenAI Developers on X →Supabase: coding agents can query your project logs
Supabase released a skill that lets your coding agent query Supabase logs and check for error spikes, security issues, slow queries and growing usage, on a schedule you choose.
Source: Supabase on X →Bubble universal import for vibe-coded apps
Bubble launched universal import: bring an app you built with a vibe coding tool into Bubble as a full app, then host it, scale it and edit it visually.
Source: Emmanuel Straschnov (Bubble) on X →v0 iOS: Browser Use annotations
v0's iOS app now supports Browser Use annotations.
Source: v0 on X →Mercury 3 diffusion model live on OpenRouter
Inception's Mercury 3 is a diffusion LLM that OpenRouter says generates over 1,000 tokens per second with a 1M-token context. Inception reports it beats GPT-6 Luna and Claude Haiku 5.5 on IFBench, GPQA Diamond and SWE-Bench Atlas (vendor claims). Launch pricing is 50% off at $0.10 per million input tokens and $0.35 per million output tokens.
Source: OpenRouter on X →MiniMax M3.1 Flash free on OpenRouter for 72 hours
OpenRouter's stealth model Space Bunny has been revealed as MiniMax's M3.1 Flash and is free for 72 hours from launch. After that it's $0.15 per million input tokens and $0.60 per million output tokens. MiniMax reports 73.67% on DeepSWE and says it plans to release the weights later.
Source: OpenRouter on X →Devin efficiency-based auto-approvals and Devin Coach
Cognition added efficiency-based auto-approvals so a team's most productive engineers get more AI budget instead of hitting blanket usage caps, plus Devin Coach to help everyone else use Devin more effectively.
Source: Cognition on X →Superset Plugins: one setup for every coding agent
Superset launched Plugins: install an MCP server, skill or login once and every agent (Claude Code, Codex, Cursor) gets it on every machine, so you stop redoing the same setup each time you switch agents.
Source: Superset team on X →Claude Haiku 5.5 in Replit's model selector
Replit's weekly recap says Claude Haiku 5.5 is now available to try in the Replit Agent model selector, alongside the new Replit Solutions Partner program and Bring Your Own Cloud for enterprise.
Source: Replit on X →Devin Fusion tops the Artificial Analysis Coding Agent Index
Cognition says Devin Fusion, running an Opus 5.5 lead with a GPT-6 Luna sidekick, scored 68 on the Artificial Analysis Coding Agent Index, beating Claude Code with Opus 5.5 at a 31% lower cost per task. This is Cognition's own claim.
Source: Cognition on X →Claude Code warns before a resume eats your usage limit
Claude Code 2.1.304 shows a warning on Pro and Max plans when resuming an inactive conversation would use a noticeable share of your 5-hour limit, so you can start fresh instead. It also adds a guided /auto-mode-setup.
Source: Claude Code changelog →Vercel Shop (GA): agent-ready Shopify storefronts
Vercel's open-source Next.js storefront template for Shopify is generally available, built on Shopify's rebuilt Hydrogen SDK. Product pages are machine-readable, agents can search products and build carts over MCP, and the template ships skills that guide coding agents through setup.
Source: Vercel blog →Codex CLI 0.163.0
Send a follow-up while Codex is still answering to steer it, search every setting from a new /config panel, and let Codex run independent read-only tools in parallel. Daybreak is now on by default for eligible accounts.
Source: Codex release notes on GitHub →Muse Wonder from Meta Superintelligence Labs in Cursor
Meta's Muse Wonder model is now selectable in Cursor. Cursor says it scores 46.1% on CursorBench at default effort, the highest of any Muse model, priced at $4 per million input tokens and $20 per million output tokens.
Source: Cursor on X →Vercel AI Gateway session view for coding agents
AI Gateway logs now group requests into sessions, showing cost, turns, cache hit rate and which coding agent (Claude Code, Codex, Cursor) ran each conversation, with session affinity to keep prompts cached.
Source: Vercel changelog →Reflection AI Beam free on Vercel AI Gateway
Reflection AI's open-weight agentic coding model Beam is free for a limited time on Vercel AI Gateway, so you can try it inside Claude Code, Codex or Cursor. Use the free model ID if you don't want it to roll over to paid pricing.
Source: Vercel changelog →DeepSeek V4.1-Flash-1017 + Search & Fetch APIs
Open-weights Flash update with better coding scores and shorter, faster outputs at the same API price. DeepSeek also opened its own Search and Fetch APIs for agents.
Source: DeepSeek on X →Aceternity UI Pages
Complete light/dark landing pages you install with one shadcn CLI command, plus an MCP server so Claude Code, Codex, Cursor or Devin can pull pages while building.
Source: Manu Arora (Aceternity) on X →Codex & Work get more effective usage
OpenAI says Codex now stretches your plan further by reusing context across subagents, caching context longer, and routing requests for more cache hits.
Source: Tibo (OpenAI Codex) on X →Cursor cloud agents build interactive canvases
Ask a Cursor cloud agent for a dashboard or report while it works, then share the finished canvas with your team as a link.
Source: Cursor on X →Gemini 4 Argon in Google AI Studio Build
Google AI Pro subscribers can now build with Gemini 4 Argon in AI Studio’s Build mode and playground.
Source: Google AI Studio on X →Replit Bring Your Own Cloud (BYOC)
Run Replit development and production inside your own Azure, Google Cloud, or AWS account for more isolation and control.
Source: Replit on X →Runway inside Replit
Connect Runway to a Replit project to generate images, video, and other assets for your site without leaving the editor.
Source: Replit on X →Supabase Middleware
Write per-request logic (like checking who’s calling) once and run it on any fetch runtime, with bridges for Hono, H3, Elysia, NestJS, and TanStack Start.
Source: Supabase on X →Bolt free for everyone this weekend
Bolt is free from 12:00am PT on October 17 for the weekend, with a build contest (submit by October 20 for a shot at $10K).
Source: bolt.new on X →Claude Code Review (GA)
Anthropic’s automated GitHub PR reviewer left research preview for Claude Team and Enterprise. Reviews finish in about 6–10 minutes with Quick, Standard, or Deep per repo; eligible Enterprise orgs can claim Code Review credits.
Source: Anthropic →OpenUI 1.0
MIT generative UI runtime so coding agents render charts, forms, and dashboards from your components instead of text walls. Agent Skill works with Claude Code, Codex, Cursor, and Copilot.
Source: OpenUI docs →OpenRouter Ori Harness
CLI that runs Claude Code, Codex, OpenCode, Hermes and more on OpenRouter with one login—pick any OpenRouter model with the ori claude --model command.
Source: OpenRouter →Claude Opus 5.5 Fast Mode (GA on Platform)
Fast mode for Opus 5.5 is out of research preview on the Claude API (up to 2.5× output speed at premium rates). Fast mode inside Claude Code is unchanged and still uses extra usage credits.
Source: Claude Platform docs →OpenAI Agents API (public beta)
Codex-style harness and sandboxes for building long-running cloud agents via API. No extra Agents API fee—pay for tokens and tools.
Source: OpenAI →Cursor Projects (beta)
Coordinator agent that delegates multi-PR features, migrations, and recurring codebase gardening across cloud subagents.
Source: Cursor blog →Cognition SWE-2
New Devin coding model available in Devin Desktop and CLI, rolling out to Web and Fusion.
Source: Cognition →Gemini 4 Argon in GitHub Copilot (GA)
Google’s frontier model lands in Copilot IDE, CLI, and coding agent. Billed at provider list pricing for Pro+, Max, Business, and Enterprise.
Source: GitHub Changelog →Codex CLI refresh (DevDay 2026)
Voice control, /agents view, and cloud Codex improvements from OpenAI’s DevDay recap. Exact DevDay calendar day not independently re-verified.
Source: OpenAI DevDay recap →Google Antigravity (public preview)
Agent-first IDE listed on Google Cloud’s free AI tools page with a free individual plan during preview (rate limits apply).
Source: Google Cloud →Hostinger Connector in Claude & Cursor marketplaces
Agent connector for deploying and managing Hostinger sites from chat. Product details reported via marketplace/directory coverage—confirm on Hostinger.
Source: Hostinger →