Maya: Claude Code 2.1.273 added something people running remote sessions have wanted. You can now fork a Remote Control session that started through the Claude app and keep the new branch running as a background session on your own computer.

James: Right, and the key detail is where that fork lives. It runs locally, so you get a second branch without pushing the new work off to a separate cloud host.

Maya: There's more for teams. It adds five optional request headers for anyone routing Claude Code through an LLM gateway, things like the request class and agent type. Operators opt in by setting the gateway hint headers variable to one, and it exposes request context the gateway can use operationally.

James: And the Model Context Protocol side gets clearer failure alerts. When a server disconnects and automatic reconnection gives up, Claude Code now posts a notification pointing you to the mcp command. There are also permission fixes, including one where a subshell could hide a dangerous remove command while bypass mode was active.

Maya: Shifting from the cloud to the desk, two builders wrote up how to turn a Mac Mini into an always-available Codex workstation. On September twelfth, Adam Stewart published a checklist for leaving Codex running on a Mac Mini, giving you a persistent workspace you reach from your phone without moving every task to a cloud computer.

James: The checklist is grounded. You update the Mac, install the ChatGPT desktop app, open Codex, add one project folder, then enable control of the Mac in Connections and scan a QR code with your phone. Stewart keeps screen lock on while turning automatic logout off, because those are different actions, and the Mac has to stay awake with the app open.

Maya: He also suggests testing before you rely on it away from home. Give Codex one specific job with approvals on, test the connection over mobile data while the Mac is locked, then test a restart before travelling.

James: And Peter Yang offered a simple division of labor: local scheduled tasks stay in Codex, cloud tasks move to Grok Bot, so each job has an obvious destination. Brown did warn the setup consumes many tokens, so it isn't free to leave running.

Maya: On the infrastructure side, OpenAI announced first-class sandbox integrations with nine providers on September tenth. The list is Blaxel, Cloudflare, Daytona, DigitalOcean, E2B, Modal, Oracle Cloud, Runloop, and Vercel, and you can also bring your own sandbox instead of picking one of those.

James: What that really does is separate the agent or harness from the environment where its work runs. The announcement names three dimensions for choosing an environment. First is compute, so CPU, GPU, and memory options, which lets you match the sandbox to the workload instead of treating every task like the same machine shape.

Maya: The second is where it runs, either fully managed environments or deployments inside your own virtual private cloud. And the third covers file and secret storage, which matters when a workload needs source files, build artifacts, or credentials inside its sandbox.

James: It's worth being precise, though. This is an integration release, not a parity claim. OpenAI doesn't say every named provider supports every topology, so choosing resources and checking a provider's implementation still falls to the builder.

Maya: Next, Cline Desktop. Early users described it on September fourteenth as a standalone coding-agent app that's free and open source, with several free models available at the start and options to connect other providers or credentials later.

James: The part I like is that it can import conversations from Codex, Claude Code, and OpenCode, so you continue an existing session instead of re-explaining the whole project. Users also described switching models per task, which makes the model a project-level choice rather than something you lock in at the start.

Maya: It also supports multiple sessions across projects that can spawn subagents, a built-in difference viewer, checkpoints to return to an earlier state, and conversation forks to test alternate approaches. And there's recurring work, with schedules for finding bugs, scanning for security problems, and updating documentation.

James: One tester connected an existing Codex subscription, built a Three.js driving game, switched models while refining the controls and camera, and used checkpoints to undo changes.

Maya: Last one, and it's for the video folks. Hypit is an open-source video system built for coding agents like Codex and Claude Code. Posts on September fourteenth and fifteenth describe giving an agent a reference video and directing the recreation with natural-language instructions instead of a traditional editing timeline.

James: And the output is an editable workflow, not just a finished file. One demo described the host, dialogue, subtitles, supporting footage, and effects as parts tied to the video's written structure. So changing a host, chart, background, or language targets the related component instead of rebuilding the whole production.

Maya: That structure also supports reuse and batches of related versions, and Hypit Studio adds a review step. You inspect mock assets and previews, then keep directing Codex or Claude Code before compiling the final video. It's free with a bring-your-own-key setup, so you connect the outside services you choose and stay responsible for those costs.

James: Hypit doesn't replace the underlying generation services, it just turns a reference production into a project whose components stay inspectable and revisable.

Maya: That's your build from Pivot Build. Go make something.