Cursor Origin hosting and agent swarms, Claude Code fixes, Codex GPT-6-Astra (agent tools ready now)
TL;DR
Cursor rolled out native code hosting and multi-agent orchestration while Claude Code and OpenAI Codex shipped targeted fixes and new model access for terminal workflows.
What shipped
From 4 to 11 September, three coding platforms released updates focused on agent orchestration, self-hosted execution, and model access. Cursor introduced the largest set of changes with Origin hosting and cloud agent features. Claude Code and OpenAI Codex followed with stability patches and new model options.
Cursor
Cursor shipped fifteen updates centered on Origin code hosting and Cloud Agents. The changes let teams run agents on self-hosted machines, start projects without GitHub, and coordinate subagents across VMs. Integrations with Vercel and Buildkite appear in multiple releases for preview and CI use cases.
- •Cursor Origin hosting Cursor Origin now hosts repos natively with two-way GitHub sync and pull request tools plus Vercel and Buildkite links for previews and pipelines.
- •Cloud agent subscriptions Agents can subscribe to PR and Slack events and run long goals with the /goal command while steering messages arrive without interruption.
- •Project start without GitHub Cloud Agents create new projects from scratch and save directly into Origin repos with live browser previews and one-click Vercel publish.
- •Self-hosted agent machines Execution now runs on internal servers with dynamic team pools, hibernation, and Linux or Mac computer use while secrets stay inside the network.
- •Projects for multi-agent work Coordinator agents delegate tasks to parallel subagents and keep shared context across environments while reacting to Slack or PR triggers.
- •Pre-warmed background builds Pre-cloned repos and install scripts cut Cloud Agent boot times by up to 10x with automatic fallback to prior snapshots.
- •Origin beta for paid plans Early beta adds native repos, bidirectional GitHub sync, in-repo agents, and app extensions for Vercel and Depot.
- •Custom modes and subagent VMs Persistent modes, isolated virtual machines for subagents, and the /goal command support long-running automated workflows.
- •No SCM required for agents Prompting starts immediately and saves to Origin repos with port-forwarded previews and direct Vercel publishing.
- •Self-hosted execution pools Dynamic scheduling and external sandbox links keep tool runs inside corporate networks on Linux and Mac.
- •Background build snapshots Pre-built environments deliver 10x faster starts and resilient recovery from the last successful install state.
- •Origin beta repo sync Native repositories include two-way GitHub and PR syncing plus built-in browsing agents for paid users.
- •PR and Slack monitoring Agents watch threads or pull requests, run isolated subagents, and accept mid-task steering updates.
- •Scratch project creation Agents generate code straight into Origin repos without third-party SCM and include live browser previews.
- •Team pool scheduling Self-hosted machines support auto-hibernation and sandbox integrations while keeping all execution on internal infrastructure.
Claude Code
Claude Code released twenty-three updates that fix gateway regressions, add output controls, and improve prompt caching. The releases target VS Code stability and headless operation for teams running unattended sessions. Organization features include managed MCP server distribution.
- •Gateway sign-in fix Version 2.1.266 stops the CLAUDE_CODE_USE_GATEWAY variable from forcing cloud sign-in when custom API keys are in use.
- •Max effort caps Version 2.1.267 adds a maxEffortLevel setting across providers and restores fresh system prompt rendering on session resume.
- •Gateway pricing sync Version 2.1.268 cleans runner sessions and fixes HTTP 400 errors on Anthropic-compatible endpoints.
- •Inline output limits New settings cap bash and task output at 128K characters and add a /skill-doctor command to audit unused skills.
- •Stability patches Routine fixes improve reliability across the CLI and integrated services during long sessions.
- •Plugin directory support The --plugin-dir flag loads multiple folders and caps saved tool results at 1 GB.
- •Gateway regression resolved The standalone CLAUDE_CODE_USE_GATEWAY variable is ignored again when API keys are supplied.
- •Prompt cache diagnostics maxEffortLevel and --system-prompt-snapshot off help teams iterate prompts without cache invalidation.
- •Permission and certificate fixes Rules containing parentheses and Bedrock certificate errors now resolve correctly.
- •Fullscreen diff panel The /diff toggle shows uncommitted edits side-by-side and reports cache-miss reasons in /cost.
- •VS Code MCP dialogs Custom output styles and direct add/remove actions for MCP servers appear inside the IDE.
- •Subagent prompt files Large prompts can be appended via --append-subagent-system-prompt-file for complex tasks.
- •General reliability updates Bug fixes address interactive sessions and background agent stability.
- •Inactive session archiving Sessions older than 14 days archive automatically with corrected sidebar reload behavior.
- •Disk and telemetry caps Tool results are limited to 1 GB and gateway telemetry fields expand for desktop sessions.
- •Proxy configuration restored Custom headers work again without forced cloud gateway sign-in.
- •Managed MCP servers Organizations push HTTP/SSE servers centrally and run headless sessions with --permission-prompts none.
- •Uncommitted edit viewer /diff tracks changes in real time while /cost explains prompt cache behavior.
- •Skill pruning tool /skill-doctor identifies unused skills and surfaces context costs before runs.
- •Minor stability release Version 2.1.263 corrects issues in interactive and background agent flows.
- •Headless permission flags Unattended environments deny prompts automatically and recognize GitLab merge requests.
- •Cache miss reporting Fullscreen diff and prompt-cache diagnostics help teams tune 1M-context models.
- •Policy status checks Organization policy errors appear in /status alongside 128K output caps and MCP management.
OpenAI Codex
OpenAI Codex delivered eighteen updates that add GPT-6-Astra access, worktree checkouts, and Vim mode improvements. Releases also cover Windows daemon support and inline question handling. Python SDK changes include new reasoning effort levels and turn management.
- •GPT-6-Astra model access Codex CLI 0.154.0 surfaces GPT-6-Astra in the picker and Amazon Bedrock catalogs.
- •Python SDK release Version 0.154.0 adds max and ultra reasoning effort plus ExternalMessage support for sync and async calls.
- •Deprecated entry removal The codex mcp-server CLI command is removed in favor of standard server workflows.
- •MCP token refresh Coordinated OAuth refresh and blocks on pre-trust workspace helpers improve session security.
- •Rich-text copy command /copy preserves markdown when pasting into rich-text apps and copies individual session fields.
- •Vim replace mode R mode with undo and dot-repeat works in the terminal UI alongside better escape handling.
- •Windows background daemon Sessions connect to a shared Codex server with lifecycle commands and auto updates.
- •Inline question answers Questions are handled without losing the active composer draft during ongoing tasks.
- •Worktree session isolation worktree creates separate git checkouts that can be browsed and resumed directly.
- •Bedrock catalog addition GPT-6-Astra appears for Mantle and Runtime routes with updated migration notes.
- •Draft search in Vim Version 0.152.0 adds Vim-mode search, rate-limit banners, and configurable thread timeouts.
- •Node REPL policy handling Guardian reviews now respect policies supplied through model metadata.
- •iOS attachment support ChatGPT for iOS 1.2026.237 adds video and photo attachments plus Priority view.
- •Vim undo and plugin commands Version 0.153.0 brings undo/redo, marketplace CLI commands, and session reconnection.
- •API-only Astra config GPT-6-Astra can be set via API without changing the default model picker.
- •Fast tier label correction The tier now correctly states 2x speed instead of 1.5x.
- •Bedrock route fixes Astra is listed for global and US routes with corrected async clarification text.
- •Picker visibility fix Version 0.153.4 ensures Astra appears correctly and sets as default when unconfigured.
What this means for you
For Vibe Builders: Cursor Origin and Cloud Agent features let you host code and run multi-agent projects without managing GitHub or servers. Start from a prompt, publish to Vercel, and keep execution inside your own machines. Claude Code and Codex updates add reliable terminal agents that handle long tasks while you focus on shipping.
For Non-techies: SMB owners gain practical agent tools that monitor Slack or PRs and publish updates automatically. Cursor self-hosted options keep data inside company networks while Codex and Claude Code fixes reduce interruptions during daily work. The result is fewer manual steps for routine coding and deployment tasks.
For Developers: Production teams can now evaluate Cursor self-hosted pools and Claude Code headless flags against existing cloud stacks. Codex worktree and GPT-6-Astra support plus MCP server management give concrete options for isolated sessions and model routing. Watch the new maxEffortLevel and cache diagnostics before rolling into CI pipelines.
What to watch next
Track Cursor agent subscription adoption and Origin hosting uptake. Monitor Claude Code MCP server rollout in organizations. Check Codex worktree usage and GPT-6-Astra performance on Bedrock routes.
Harsh’s take
Cursor concentrated on full agent orchestration and hosting while the other two platforms focused on fixes and incremental model access. The pattern shows one vendor pushing end-to-end workflows and the others stabilizing existing surfaces. Builders should test Cursor Origin on a small internal project this week to measure boot time and self-hosted security before committing team workflows.
by Harsh Desai
More AI news
- Daily RoundupOpenAI Agents on Vercel, Gemini Windows app, and new agent tools for builders
Vendors expanded agent hosting, model access, and inference options while new trending models and hardware integrations appeared across platforms.
- Daily RoundupVercel CLI and v0 updates, Google AI tools, Apple iPhone AI features (agent memory and integrations)
Vercel expanded agent tooling and deployment security while Google and Apple rolled out practical AI features for everyday tasks and new hardware.
- Daily RoundupMercury 2.5 and Nex-N2.5 on OpenRouter, Meta Muse agent, and Vercel speed gains for agents
Model releases and agent tools expand options for builders while enterprise deals and routing improvements reduce friction for teams shipping AI products today.