GitHub listed fifteen projects this Tuesday morning. The top entry is computer use built on the accessibility tree: Vercel Labs' agent-desktop lets agents click and type in desktop apps through references they can check, instead of guessing from screenshots. Worktrunk and Paperclip tackle the other half of the agent boom, keeping many parallel agents organised; colibri streams giant models from disk on machines without a GPU; Scrapling is a full scraping toolkit with a legal caveat; and nanochat remains a readable course in training a chat model. Curation order reflects practical interest, not GitHub rank. Trending is not a release date or an endorsement.
Picks are listed in curation order, not GitHub’s rank; each card shows the rank GitHub displayed when the page was checked. Stars measure attention, not security, quality or fitness for use, and no project here was installed or run. “Stars today” is GitHub’s own displayed figure for its daily window, not an audited rolling 24-hour count.
Pick 1 · AI agents / desktop automationGitHub rank #1 of 15
What it does. A command-line tool that lets AI agents operate desktop apps on macOS, Windows and Linux. It reads the operating system's accessibility tree and returns JSON with a reference for every element an agent can act on; the agent clicks or types on a reference, then takes a new snapshot to check the result, with no screenshot pixel matching. It has 60 commands (clicks, typing, windows, clipboard, notifications) and a C library for calling it in-process.
Caveat. An agent using it can do whatever your account can do in any open app, so only point it at tasks and windows you would trust it with, and treat text it reads on screen as untrusted input. The README warns that snapshots, clipboard reads and screenshots can contain private content. macOS needs Accessibility permission, and Linux support is partial (mouse input and window control are X11-only, notifications GNOME-only). It is young: v1.0.0 was 28 October. We did not install it.
Attention, not endorsement
GitHub showed
“2,275 stars today” as displayed for GitHub’s daily window; not an audited rolling 24-hour count
What it does. A CLI that makes Git worktrees as easy as branches. One command creates a worktree for a branch and moves into it (optionally starting an agent there), another removes the worktree and its branch, and a list view shows status. Extras include hooks on create and merge, a squash/rebase merge command, pull-request checkout, a unique dev-server port per worktree and copy-on-write sharing of build caches.
Caveat. Still pre-1.0, and releases can break setups: v0.81.0 stopped passing hook context as JSON on stdin, so existing hooks need rewriting. Changing directory needs shell integration, which edits your shell configuration. LLM commit messages and branch summaries use a model you configure, which sees your diffs, and shared build caches work only on APFS, btrfs and XFS. 'The most popular git worktree manager' is the maintainer's own line. We did not install it.
Attention, not endorsement
GitHub showed
“77 stars today” as displayed for GitHub’s daily window; not an audited rolling 24-hour count
What it does. A pure-C engine that runs very large mixture-of-experts models on ordinary computers. It keeps the always-used part of a model in RAM and reads the experts from disk when each token needs them, with one source file per model family. It runs on the CPU, on any GPU through Vulkan or on NVIDIA cards through CUDA, reads published GGUF files as they are, and serves OpenAI- and Anthropic-compatible APIs plus a browser dashboard.
Caveat. Streaming from disk is slow for the biggest models: the README's own table gives 0.05-0.1 tokens per second cold for GLM-5.2 on a 25 GB laptop, from a 429 GB download, and 0.9-1.6 tokens per second for the 284B DeepSeek V4 Flash. These are maintainer measurements we did not reproduce, and a fast NVMe drive matters. Model weights carry their own licences (Muse Spark Open is under Meta's licence agreement). The Windows starter offers to install Python for you. We did not run it.
Attention, not endorsement
GitHub showed
“675 stars today” as displayed for GitHub’s daily window; not an audited rolling 24-hour count
What it does. A Python web-scraping framework that combines fast HTTP requests, stealth browsers for JavaScript-heavy pages, CSS and XPath parsing with adaptive element tracking, and concurrent spiders with proxy rotation, retries, pause/resume and JSON, CSV or XML exports. It also has a CLI, an interactive shell and an MCP server so AI agents can browse pages and extract data.
Caveat. Much of its pitch is getting past anti-bot protection: the README says it bypasses Cloudflare Turnstile, and its sponsors sell residential proxies and CAPTCHA solving. Whether that is allowed depends on each site's terms and your local law; its own disclaimer says to respect terms of service, and robots.txt compliance is an optional flag. v0.5 has many breaking changes and a separate migration guide. We did not run it.
Attention, not endorsement
GitHub showed
“706 stars today” as displayed for GitHub’s daily window; not an audited rolling 24-hour count
What it does. A self-hosted dashboard for running teams of AI agents: a Node.js server and React UI where you set goals, give agents (Claude Code, Codex, Cursor, OpenCode, Hermes and others) roles in an org chart, assign tasks, approve work and track costs. Agents wake on a heartbeat for assigned work or schedules, and companies, agents and projects can have budgets that pause work at a limit.
Caveat. Its goal of agents running a business around the clock can spend real money, and budgets act on the spend reported to it. Anonymous usage telemetry is on by default (PAPERCLIP_TELEMETRY_DISABLED=1 or DO_NOT_TRACK=1 turns it off), and the quickstart starts in a trusted local loopback mode, so pick the authenticated mode before exposing it on a network. It needs Node.js 24.11 or newer. 'The app everyone uses' is the project's own claim. We did not run it.
Attention, not endorsement
GitHub showed
“412 stars today” as displayed for GitHub’s daily window; not an audited rolling 24-hour count
What it does. Andrej Karpathy's minimal, hackable harness for training a small chat model end to end on one GPU node: tokenizer, pretraining, fine-tuning, evaluation and inference. One script, runs/speedrun.sh, trains a model of GPT-2 capability, and you can then talk to it from a command-line chat.
Caveat. The reference run needs a rented 8xH100 GPU node; a CPU path exists for experimenting, not for reaching GPT-2 level. The result is, in the README's words, a bit like talking to a kindergartener, so it is for learning, not production use. There are no releases and no commits since 3 July 2026, and the cost and leaderboard figures are the maintainer's. We did not run it.
Attention, not endorsement
GitHub showed
“321 stars today” as displayed for GitHub’s daily window; not an audited rolling 24-hour count
Every repository GitHub listed, in GitHub’s order, with the figures it displayed when the page was checked (): 6 picked today, 8 covered in an earlier digest, 1 not covered. For projects covered before, growth compares the total stars at our first pick with today’s displayed total.
TypeScript · 1,803 stars today · 111,058 stars in total
Persistent Context Across Sessions for Every Agent – Captures everything your agent does during sessions, compresses it with AI, and injects relevant context back into future sessions. Works with Claude Code, OpenClaw, Codex, Gemini, Hermes, Copilot, OpenCode + More
Covered before — first picked 30 Oct 2026 at 105,064 stars; now 111,058: +5,994 stars (+5.7%) in 4 days, about 1,498 a day. Also picked 2 Nov 2026.
The BEST Web Scraping framework that handles everything from a single request to a full-scale crawl! Don't be shy, join here: https://discord.gg/EMgGbDceNQ and follow here for daily tips and tricks: https://x.com/Scrapling_dev
Training-free language model modification: Remove censorship, suppress slop, reduce sycophancy, and more
Not covered yet.
Following Our Picks
Every project picked in an earlier digest, biggest growth first. “Then” is the total at our first pick’s metadata check; “now” is GitHub’s API star count checked . Growth measures attention, not quality or safety.
All 15 listing entries were inspected. Exact primary public GitHub repository metadata and recent releases were checked for all 15 at about 07:00-07:01 UTC, with README files for the seven entries not featured before; all were active and unarchived. Every pick has a licence: worktrunk's API value is NOASSERTION, and its LICENSE file states MIT or Apache-2.0.
No repeats today. Eight entries featured before were left out without a material change: thedotmack/claude-mem (its v15.9.0 adds Devin support and more secret redaction, incremental after our 2 November privacy card; issue 5176 was still open when checked), hyprwm/Hyprland, FlashML-org/FreeToken and kornia/kornia (2 November), pingdotgg/t3code (nightly builds only) and mattpocock/skills (31 October), Donchitos/Claude-Code-Game-Studios (30 October) and shiyu-coder/Kronos (20 October; no commits since April).
Not picked: p-e-w/heretic, whose v2.0.0 (1 November, AGPL-3.0) modifies language models without training and whose default configuration removes their safety refusals; left out as an editorial choice, as with earlier jailbreak tooling.
No security-focused project was on today's list.
The preceding seven calendar dates, 27 October-2 November, were checked for local history; all seven editions exist.
The official listing was captured at 08:00:33 Europe/Paris. Literal stars-today labels are GitHub's displayed period figures; API total stars and README evidence were checked separately at the times shown per card. They are not an independently audited rolling 24-hour growth measurement.
Feature, speed, cost and adoption descriptions are the maintainers' own, including colibri's speed tables, nanochat's leaderboard and costs, and the popularity lines of Worktrunk and Paperclip. No candidate code, installer, plugin or model was installed, downloaded or run, and none of this is a security audit or endorsement.