Vercel's agent-desktop: open-source desktop control for any agent
Apache-2.0, cross-platform, MCP built in. Plus OpenAI's Nov 12 Cursor cutoff, Inkling 2 Small's imminent open weights, and fresh agent tooling.

Copy markdown
Desktop control, minus the screenshot tax
Vercel Labs open-sourced agent-desktop (Apache-2.0), a computer-use CLI that reads an app's accessibility tree and hands your agent JSON with a clickable ref per element — no pixel-matching and no model inside, so it drives Finder, native apps and more with whatever agent you already run. Works on macOS 13+, Windows 10+ and Linux; `npm i -g agent-desktop` to start, or `agent-desktop mcp` to expose every command as an MCP tool over stdio.
OpenAI's models leave Cursor on Nov 12
OpenAI's proposed shutoff is four days out: GPT-5.6 Sol, Terra and Luna drop out of Cursor's model picker, and no future OpenAI models ship under the contract (OpenAI cites Cursor's acquisition by SpaceX). To keep using them, bring your own OpenAI API key (billed at API rates, not your ChatGPT plan), run the separate Codex extension, or route through Bedrock, Azure, OpenRouter or the Vercel AI Gateway — none of which restores OpenAI models inside Cursor Tab, Auto or Background Agents.
Inkling 2 Small's open weights are days out
Thinking Machines' second model — a 594B-total / 17B-active MoE with vision — is live now on Tinker with closed weights, and the lab says full Apache-2.0 checkpoints land "next week" after red-teaming. It posts 86.3% on SWE-Bench Pro v2 (Hard) and 72.1% DeepSWE v1.1 with a 1M-token ceiling, so it's shaping up as a fine-tunable coding-and-vision base you'll be able to self-host, not just rent.
Parallel agents on one repo, no collisions
Worktrunk (open-source, by max-sixty) is a thin CLI over Git worktrees that gives each coding agent its own directory on the same repo — run Claude Code, Codex and others side by side without them clobbering each other's working tree or branch state.
An open Kubernetes layer for your GPUs
NVKE (Nvidia Kubernetes Engine) landed on Hacker News' front page at 124 points: an open-source, Kubernetes-native interface for managing Nvidia GPU clusters, pitched as a base layer for self-hosted model serving if you'd rather own your inference than rent GPU-as-a-Service.