CoScreen shut down on July 31, and its founders rebuilt it to outlive them.

Petal (Apache-2.0, Rust and TypeScript with a Tauri shell, 311 stars, five contributors, repo created August 27, 217 points on Show HN Wednesday) shares windows rather than screens. Each window someone shares shows up on your desktop as its own borderless native window that you can move, stack and dock next to your editor, not as a tile in a call grid. Several people can share at once, everyone gets a labeled cursor, anyone can draw on a shared window, and remote control injects a teammate's clicks and keys without stealing your focus, after a consent prompt each time by default. Rooms are persistent and named, and people without the app join from a browser. An opt-in AI chat gives anyone in the meeting a live Gemini session about a shared window, which can act on it only with the sharer's approval.

The no-sunset promise has a specific shape. The hosted instance runs on LiveKit, which two of the founders said in the thread they are paying for themselves. The self-hosting doc asks for three pieces: your own LiveKit server with TLS, the repo's stateless backend/ deployed as Vercel-style functions, and a desktop build with your backend's URL embedded at build time. That last step means building from source, which currently needs Xcode 26 on a Mac. A commenter asked whether the released app could point at a self-hosted backend without a rebuild; the answer in the thread came from a maintainer of Hopp, a competitor that does that. The founders also said Petal isn't yet as tuned for latency as CoScreen was. The README is plain about status: the macOS app is signed, notarized and in daily use, the Windows installer is unsigned and "incomplete and buggy," and docs/VALIDATION.md lists what has and hasn't been checked.

Reach for it if your team pairs over Zoom or Meet and keeps losing the editor in a grid tile. Delete the Tuple seat you kept only because CoScreen died. Tradeoffs: pre-1.0, with remote control called early in the README; Windows is rough and Linux doesn't exist yet; self-hosting means building and signing your own desktop client; and the free hosted service runs on the founders' own money.

Muse Spark Open (Muse Model License, released Thursday, 841 points and 421 comments on HN) is the open-weights version of Meta's Muse Spark: a mixture-of-experts model with 816 billion parameters, 42 billion active per token, a 1M-token context and text, image, video and audio input, trained for coding and long-horizon agent work. Meta shipped it with day-one support in SGLang, vLLM, TensorRT-LLM, Transformers, MLX, llama.cpp and Unsloth, an NVFP4 checkpoint about a third the size of BF16, and deployment guides for NVIDIA, AMD, Intel and Apple. Seven hosts serve it, including OpenRouter, Fireworks and Together. The launch headline is a desktop: two Mac Studios with M5 Ultra and 512 GB each split the model with MLX tensor parallelism, or four with 256 GB each, which Meta's guide says serve faster.

On hardware, Unsloth's dynamic GGUFs are the smallest route: its table puts the 2-bit quants at 250 to 316 GB of RAM plus VRAM, and BF16 at about 1.63 TB, and it recommends UD-IQ2_S as the minimum for agentic tasks. Meta's own Hugging Face repos are gated with manual approval; Unsloth's GGUF and mlx-community's quants are not, and Unsloth's had about 2,600 downloads by this morning. A commenter quoted the license's one restriction: a company with over $20 million in annual revenue that offers the model to others as an API must sign a separate agreement with Meta first, while internal use and features inside your own product are excluded. On quality, a commenter's Artificial Analysis chart put it at 39 on the index, level with DeepSeek V4.1 Flash and below Muse Spark 1.3's 45; another said it "feels dumber than GLM 5.3" but runs fast on Fireworks. Meta's post publishes no tokens-per-second for the Mac setup, which another commenter went looking for and didn't find.

Reach for it if you need a large agentic model you can keep running after a provider changes its terms, or you already have the hardware. Delete the assumption that a frontier-adjacent model always means a rented endpoint. Tradeoffs: the Mac route needs 512 GB machines commenters say you can't buy yet; at 2-bit you trade quality for fit; Meta's benchmark history is why several commenters said they'd wait for independent numbers; and the API restriction will matter to any provider that grows past $20 million.

Sill (Apache-2.0, Swift, 2,260 stars, repo created September 22, v0.5.1 on October 9, 174 points on HN overnight) streams one Mac window, or the whole desktop, to your iPhone or iPad, and sends your taps, trackpad gestures, keys and Apple Pencil back. Sill for Mac lives in the menu bar, captures with ScreenCaptureKit, encodes HEVC on the hardware encoder over TLS 1.3, and drops frames rather than queue them when the link lags. Devices connect over the same network, a USB cable, peer-to-peer Wi-Fi when there's no shared network, or your own VPN such as Tailscale when you're away. You pair each device once with a 12-digit code and every later connection is pinned to those keys. Both apps are free: the Mac app is a notarized download or brew install --cask saffsanity/sill/sill, and the iOS app is on the App Store.

The thread worked out what it replaces. Sidecar extends your Mac's display onto an iPad and starts from the Mac; Sill starts from the handheld, shows single windows, works on an iPhone and runs over a VPN, which commenters said Sidecar doesn't. One user said it let them drop a Sunshine and Moonlight setup that "breaks half the time." The README's "Good to know" paragraph lists the limits: one window at a time, no sound, text-only clipboard up to 256 KB, and the Mac has to be awake with Sill open. The virtual display, which keeps a covered window updating, uses a private macOS API, and the README says that's one reason the Mac app isn't in the Mac App Store.

Reach for it if you leave an agent running on your Mac and want to check on it, or answer it, from your phone. Delete the VNC client that only shows the whole desktop at phone size. Tradeoffs: one author wrote 696 of 697 commits; the virtual display depends on an API Apple can change; Sill can't wake a sleeping Mac; and an iPhone or iPad app older than 0.5 (2) can't connect to Sill for Mac 0.4.0 or later at home.

ARTEMIS (Apache-2.0, Python, 11,971 stars, 221 of them today on GitHub trending, repo created August 13, six contributors) turns natural-language instructions into actions on an Android device or emulator over ADB. ./start.sh or .\start.bat installs ADB, scrcpy, FFmpeg and the Python dependencies, opens a local web console with screen mirroring and replays, and offers to install its MCP server and a rules file into Claude Code, Codex, Cursor, Antigravity and others. From there an agent can build your APK, install it, walk the login flow and return screenshots and Logcat. There are two profiles: Flash, a single-model observe-and-act loop at three to five seconds a step, and Pro, a planner, operator and read-only checker at 15 to 40 seconds a step. A Python client puts the same runs inside pytest with typed results.

The headline is "99%+" on Google Research's AndroidWorld benchmark, and the README doesn't say which model produced it. The config does say what the visual locator needs. Its comment calls standard models, including GPT-4o and Claude 3.7, unable to place exact coordinates, so Flash's grounding "MUST use specialized Gemini ER" models, and the default is gemini-robotics-er-2-preview. Element indices from the accessibility tree come first and the visual path is the fallback for Canvas, Compose and Flutter screens, so how much that matters depends on your app. The first task installs an Accessibility helper on the phone, visible as a notification and a Settings entry; the README says it listens only on the device, and .env can block auto-install on personal phones.

Reach for it if you ship an Android app and want your coding agent to check its own changes on a device. Delete the Appium script you rewrite every time a layout changes. Tradeoffs: the benchmark claim has no model or method attached; custom-drawn UIs lean on a preview Gemini model; Pro takes 15 to 40 seconds a step; and the repo has no tagged releases, so you track main.

Scion (Apache-2.0, Go, 1,917 stars, nightly builds, 34 contributors, open-sourced in March) runs harnesses such as Claude Code, Codex, OpenCode and Antigravity as separate agents, each in a container with its own config, credentials and either a shared directory, a git worktree or a separate clone. scion server start opens an onboarding wizard that detects Docker, Podman or Apple Container; then scion init in a repo and scion start my-agent "task". Agents learn the scion CLI through skills and help text, so they can delegate to each other, message each other and notify you when work finishes or needs attention. scion attach drops you into an agent's tmux session, and you can follow along in a web chat or through Telegram, Discord, Slack, Google Chat or Teams. The same model scales from a local hub to Kubernetes or Cloud Run with roles and scoped secrets.

The README is candid about what it is: "a practical testbed for emerging patterns in agent collaboration," pre-1.0, with APIs and configuration that will keep changing. The disclaimer reads "not an officially supported Google product" and excludes it from Google's open-source bug bounty. The recommended install is the community homebrew-scion tap, and the CLI it installs pulls agent images from that tap's registry, ghcr.io/homebrew-scion, by default. One maintainer, ptone, wrote about 5,900 of its roughly 6,300 commits. The April InfoQ write-up drew 230 points on HN.

Reach for it if you already run several agents on one repo and are tired of them stepping on each other's files and tokens. Delete the tmux layout and the shell script that hands out worktrees. Tradeoffs: no stable release, only nightlies; the default install path trusts images a community tap publishes; it's built for exploring collaboration patterns, not running production work; and Google promises nothing about its future.

Verification notes: star, license, creation, contributor, commit and release data are from the GitHub API on 2026-10-16; commit totals use the per_page=1 Link-header last-page count, and trending counts are from today's github:trending scrape. HN points and comments are from the Algolia API. Petal's history, eight-week figure, hosting costs, latency answer and the Hopp reply are from HN item 50092470; its features, status and Windows caveats are from its README, and the self-hosting steps from docs/SELF_HOSTING.md. Muse Spark Open's specs, runtimes, hosts and Mac setups are from Meta's research post, developer blog and deployment guide; the quant sizes and minimum-quant advice are from Unsloth's guide; gating and download counts are from the Hugging Face API this morning. The license clause, Artificial Analysis score, GLM comparison and missing Mac speed are from HN item 50107379; the license text itself is gated and was not read directly. Sill's features, requirements, limits and private-API note are from its README; the Sidecar, VPN and Sunshine comments are from HN item 50112739. ARTEMIS's setup, profiles, helper and benchmark claim are from its README; the ER-model requirement is from config/artemis.jsonc and the provider list from .env.example. Scion's features, status, disclaimer and install path are from its README; the April thread is HN item 47675213. Pipeline health: today's pool held 71 rows, against the ~280 the compose prompt assumes. github:trending returned 12 rows across all five feeds, and Reddit returned zero again, 160 days since its last nonzero day on May 9. No editions exist for 10-14 or 10-15, and neither day has a log file.