rea leads a week of agent discipline, DeepSeek tooling and two very different television fixes

Fastest-growing this period

Counts come from GitHub's official stargazers/history endpoint: the stars gained on each day of the window. Project briefs and the reading notes below are editorial additions.

1. 1

   morluto/rea : Point your agent at any app and learn how it works

   TypeScript

   A toolkit that lets your coding agent reverse engineer software without source code: it inspects native binaries through Hopper or Ghidra, JavaScript and Electron apps, .NET assemblies and websites, then explains how a feature works with the evidence and the limits behind each conclusion. Analysis runs locally, and the same tools work straight from your terminal.

   +58,255
2. 2

   Gathin23/cleanyourtv : Debloat any Android TV over Wi-Fi and swap in a clean launcher

   Java

   A kit that turns an ad-filled Android TV home screen into a calm, Apple TV style one. Over the same Wi-Fi, its ADB scripts back up the TV, disable telemetry, sponsored rows and preloaded apps without uninstalling anything, then install a tiny launcher with a dock of four apps. A restore script undoes every step, and tested lists ship for Xiaomi, TCL, Sony and more.

   +32,119
3. 3

   deepseek-ai/DeepInfer : DeepSeek's production inference engine, released for offline serving

   Python

   The inference framework behind DeepSeek's own production services, released for offline serving. It runs DeepSeek-V4.1-Flash with optimized kernels (FlashMLA, DeepGEMM, DeepEP), MoE expert parallelism with load balancing, asynchronous batch scheduling and structured outputs, targeting NVIDIA SM100 GPUs or Huawei Ascend 950 NPUs. Apache 2.0, built largely on vLLM.

   +22,364
4. 4

   Ebony-Vinyl/dsh-our-free-model : A dsh plugin that puts frontier models one install away, no login or API key

   JavaScript

   A plugin for the DeepSeek Harness shell: install it and the model picker fills with free frontier models like DeepSeek V4.1 Flash and Kimi K3, no login or key needed. Free lanes come straight from OpenCode's Zen and Kilo AI's public gateways, re-pulled every refresh. A local OpenAI-compatible relay port lets other tools call the same models, and usage stats stay on the machine.

   +18,048
5. 5

   cloudflare/security-audit-skill : Turn your coding agent into a security auditor that cross-checks its own findings

   JavaScript

   Installs as a skill that makes your coding agent audit a codebase for vulnerabilities. It maps the architecture and trust boundaries first, sends isolated hunter agents after the uncovered areas, then gives every candidate to a fresh verifier that tries to disprove it, so a finding only reaches the report with a complete source trace.

   +15,527
6. 6

   ai-job-search-dev/ai-job-search : A Claude Code workflow that runs your whole job search on your own machine

   Python

   A Claude Code framework that keeps the entire job search on your own machine. Fill in your profile, then the agent evaluates postings, tailors your CV and cover letter in LaTeX, and prepares interview answers. The bundled portal skills target the Danish market, with a pattern you can swap for your local job boards.

   +11,525
7. 7

   google-research/harness2 : Let agents rewrite the prompts, tools and memory around them

   Python

   A Google Research framework for recursive agent harnessing. At task level it proposes edits, probes them on the same task, then composes a refined harness from behavioral contrasts, weights untouched. A second level keeps an improvement harness that reuses evidence-backed lessons across tasks. Includes OpenCode and Codex executors plus JobBench, LAB and WorkBuddy-Bench.

   +11,473
8. 8

   peach-mitao/peach : One private library across local drives and cloud disks

   Python

   A desktop app that gathers a personal film collection from local drives and mounted cloud disks such as 115, PikPak and OneDrive into one library. It reads existing NFO files and posters, fills in missing metadata, plays from a fullscreen cover wall with subtitle search, and follows creators across FANBOX, Patreon and Kemono. Watch history and settings stay on your own machine.

   +11,422
9. 9

   debpalash/VoiceStudio : Try voice design, transcription and dubbing in one workspace

   Python

   A local audio workspace for voice design, transcription, dubbing and longer narrated projects. Useful for trying several speech workflows with your own material in one application, with engine choice and hardware determining the supported languages and output quality.

   +10,641
10. 10

    DietrichGebert/ponytail : Stop your AI turning a small fix into a new framework

    JavaScript

    A set of instructions that asks AI coding agents to reuse existing code and question whether a new abstraction is needed. Its rules also explicitly retain validation and error handling, useful when your agent keeps making a small task unnecessarily complicated.

    +9,120
11. 11

    stablyai/orca : Compare several coding agents in separate workspaces

    TypeScript

    A desktop interface for running terminal coding agents in separate Git worktrees. You can compare attempts at the same task and inspect their changes, useful when juggling several experiments would otherwise mean managing many terminals and directories.

    +8,787
12. 12

    deepseek-ai/deepseek-harness : Explore DeepSeek's building blocks for an extensible agent

    TypeScript

    DeepSeek's agent runtime built around replaceable plugins, with a web interface for running it. Useful for developers exploring how to compose an agent's capabilities; it is a developer preview, so expect interfaces to change as you experiment.

    +8,090
13. 13

    KnifeLemon/Filee : Hold a key, drag a file, drop it on the format you want

    C#

    A free, open-source file converter with no app window: hold Ctrl (Option on macOS), drag your files, and drop them on the ring of formats that opens at your cursor. It handles 244 formats including Office documents, PDF, e-books, CAD files and 3D models without installing Office or LibreOffice, works fully offline, and saves each result next to the original.

    +7,375
14. 14

    xerj-org/xerj : Let your agent reuse how similar open-source repos solved a task

    Rust

    A local search engine in one Rust binary that auto-indexes any folder, so an agent pulls the exact code passage it needs instead of reading files one by one. For a new task it finds the closest open-source repos and reuses their solutions first; a controlled study reported 2.7x fewer output tokens at the same solve rate. BM25 search and an MCP endpoint are built in.

    +7,123
15. 15

    mattpocock/skills : Make your AI ask the right questions before it writes code

    Shell

    Give your coding assistant a skill that asks detailed questions before implementation. Use it to clarify vague requirements before the agent builds the wrong thing.

    +6,896
16. 16

    YuJunZhiXue/dsh-purge : Patch your local DeepSeek Harness to see where a refusal actually comes from

    JavaScript

    A plugin that patches your own DeepSeek Harness install and splits host-side interception from model behavior, so a red-team run can tell whether a refusal came from the harness runtime or the model itself. Prompts are swapped per model, the patch applies or restores in one click, and you can point it at your own server. The license is commercial.

    +6,753
17. 17

    Graphify-Labs/graphify : Query your whole project as a graph, not file by file

    Python

    A skill for AI coding assistants that maps a whole project, code plus docs, PDFs, images and video, into a knowledge graph you query instead of grepping files. Code is parsed locally by tree-sitter, and every connection is tagged as extracted or inferred, so you can tell what was actually read. You get a clickable graph.html and a graph.json to question in plain language.

    +6,718
18. 18

    Zagitta/gpu-unlock : A kernel module that unlocks full FP16 and BF16 rates on RTX 30 cards

    C

    A Linux kernel module that restores the full FP/BF16 instruction issue rate on GeForce 3000 series cards before the unmodified NVIDIA driver binds, with no patched driver or UEFI app to maintain. User reports on the 3090, 3080 and 3070 show large BF16 GEMM and LLM prefill speedups, while memory-bound decode and INT8 stay flat and compared outputs came back bit-identical.

    +6,402
19. 19

    affaan-m/ECC : Give your coding agent a plan, tests and a memory

    JavaScript

    A collection of coding-agent workflows for planning, tests, review and remembering useful context. If your assistant jumps straight into edits or forgets earlier decisions, this gives you reusable routines to adapt to your project.

    +5,690
20. 20

    ifixai-ai/iFixAi : An auditor that checks whether your AI agent actually does the job it was hired for

    Python

    An independent auditing tool for AI agents that grades whether an agent does the job it was hired for, not just whether it is fast or hard to hack. Guided setup picks the provider, judge and test suite, then about 60 inspections across five pillars return an A to F scorecard. Run it as a CLI wizard, as scripted flags in CI, or as a plugin that lets the agent audit itself.

    +5,370
21. 21

    storytold/photocraft : Photoshop rebuilt from scratch in pure Rust, PSD files included

    Rust

    A clean-room reimplementation of Photoshop written entirely in Rust: layers, masks, adjustment layers, type and brushes, opening, editing and saving real PSD files. Every action is a command, so the same engine can be driven from the interface, the command line, a JSON control channel or an MCP server.

    +5,246
22. 22

    heygen-com/hyperframes : Write HTML and let an agent render it as video

    TypeScript

    Turns HTML, CSS and seekable animations into deterministic MP4 videos. You can drive it from a local CLI or hand its skills to an AI coding agent, which plans the video, writes valid markup, lints, previews and renders. A practical route to product intros or animated explainers without learning a timeline editor. Requires Node.js 22 or newer.

    +5,193
23. 23

    Niko1221/Strata : A 125B-parameter model running on one gaming GPU, faster than you can read

    C++

    A one-click installer that runs Qwen3.8-Flash-Next, a 125-billion-parameter model that normally needs a server, on a single NVIDIA card with 12 to 24 GB of VRAM and 64 GB of RAM. It serves an OpenAI-compatible API on localhost, reports 60 to 95 tokens per second in its own measurements, accepts image input, and calibrates the engine to your PC.

    +5,126
24. 24

    yetone/magpie : Switch every coding agent's model from one menu bar picker

    Go

    A menu bar panel that lists each AI agent on your machine, from Claude Code to Codex, together with the model it is set to; click a value and pick another. It edits config files surgically, keeping comments and formatting intact, and runs a local gateway so every agent can share one provider login, including your Claude, ChatGPT or Copilot subscription.

    +5,099
25. 25

    pengchujin/aerial-tv : Swap your TV's ad-stuffed home screen for an Apple TV style launcher

    Kotlin

    An Android TV launcher modeled on the Apple TV home screen: no ads, no recommendation feed, just your apps, wallpapers and aerial screensavers. It starts with the TV and needs no root. A phone companion on the same Wi-Fi mirrors the app grid for reordering, renaming, hiding clutter and pinning HDMI inputs as tiles. Screensavers play 114 aerial videos in 4K or Dolby Vision.

    +4,966

Ranking window: 2026-10-19 to 2026-10-25 UTC.