Many Android and Google TVs ship with sponsored rows, analytics services and preloaded apps on the home screen. This project gives a coding assistant such as Claude Code, Codex or Gemini CLI a prompt and reference lists so it can clean up your TV for you.

It works through the Android debug bridge (ADB). You turn on USB debugging in the TV's developer options and allow your computer once. The assistant backs up the package list and settings, disables unwanted packages with pm disable-user, and installs a small native launcher as the default home. Nothing is uninstalled, so a restore script can undo each step.

Bloat differs by brand, so the assistant discovers what is on your TV and asks before touching anything you might use. Contributors have added tested lists for many models, one folder each in the examples directory.

Read more on interesting repos · GitHub

Many AI note-takers join your call as a bot and send audio to a server. DoodleNote works from your side instead. It captures your microphone and the call audio on your computer, so it works with any meeting app.

Transcription runs on the device. On Apple Silicon it uses a Swift engine built on FluidAudio, and on Windows it uses a packaged local speech engine. Notes can be generated with a downloaded local model or with your own provider key. Audio stays on your machine.

Cloud features are opt-in. Hosted Sync uploads the meeting content you choose, but not the recordings, and costs $10 per user per month. The local app is free and needs no account. The repo also holds a Next.js web workspace, an iPhone app still in development, and a read-only MCP server for AI tools.

Read more on interesting repos · GitHub

Shockwave is a notes app for people who want an AI agent working on their notes without wiring tools together themselves. Your notes stay as .md files in a folder you own. The agent reads them, keeps them in order, and can act on what's in them, such as updating a TODO file.

It has two parts. A desktop app (macOS, Windows, Linux) is what you write in. A small server you run on a cheap Linux VPS holds your notes, keys and chats, and runs the agent, so work continues when your laptop is closed. Syncing uses a GitHub repo and a personal access token, which the README says also handles sync conflicts automatically.

The README sets it apart from typical notes apps by using GitHub for sync instead of a paid service, and by building the agent in rather than asking you to run a separate coding agent. It also lists memory and self-improvement features, though it gives few details on how they work.

Read more on interesting repos · GitHub

Most people can't say exactly what they did last Tuesday, and writing standups or ticket updates by memory is slow and often wrong. Meridian runs in the background, captures screen activity, and rebuilds each day as a timeline you can scrub back through.

At the end of the day it summarises what you finished and what pulled you off plan. It also drafts a standup and a worklog update for your tickets. You review the draft and approve it before anything is posted to Jira.

Your activity is stored in one encrypted database on your own machine. Analysis goes through whichever AI provider you connect, and only what a summary needs is sent. The app is written in Rust with a Tauri tray, and the README names macOS on Apple Silicon and Windows 10/11 as the supported platforms.

Read more on interesting repos · GitHub

Meeting note tools usually send your audio to a cloud service. Meetily runs the whole pipeline on your computer instead: it captures audio, turns speech into text, and then asks a language model to write a summary.

It is one Tauri app. A Rust backend handles audio capture and transcription, and a Next.js frontend provides the interface. Transcription uses Whisper or NVIDIA's Parakeet models. For summaries, the README recommends Ollama running locally. You can also point it at Claude, Groq, OpenRouter, OpenAI, or any OpenAI-compatible endpoint, though those send text to an outside service.

The community edition is MIT licensed. The README also describes a separate paid product, Meetily PRO, built on a different codebase. It lists features such as custom summary templates, advanced exports and self-hosted team deployment. Speaker identification is listed as coming soon for PRO.

Read more on interesting repos · GitHub

Email piles up, and most of it is routine: newsletters, cold pitches, messages waiting on a reply. Inbox Zero connects to your mailbox and uses an AI model to handle that routine work.

You write rules in plain English about how to treat certain mail. The assistant applies them to sort messages, drafts replies in your style, and tracks which emails need an answer and which are waiting on someone else. It also builds briefings before meetings from your email and calendar, and saves attachments to Google Drive or OneDrive. You can chat with it from Slack or Telegram.

The README calls it an open source alternative to Fyxer, with more customization and security. You can use the hosted version or self-host with Docker. It is a Next.js and TypeScript app with a Postgres database. The GitHub license field shows NOASSERTION, so check the license file before reusing the code.

Read more on interesting repos · GitHub

FluidVoice lets you press a global hotkey, speak, and have the text typed into whatever app you are using. Speech recognition runs on your Mac, so your voice does not leave the machine unless you choose a cloud AI provider for cleanup.

You pick a speech model during onboarding. The choices include Nemotron, Parakeet, Cohere Transcribe, Apple Speech and Whisper, which trade off language coverage, download size and latency. Text goes in through macOS accessibility APIs, which is why the app asks for microphone and accessibility permissions.

On top of plain dictation there is optional AI enhancement for formatting and capitalization. One caveat: the local enhancement model, called Fluid Intelligence, is a separate runtime that the maintainers keep private. The app itself is GPLv3. The README presents it as a local alternative to paid dictation apps such as Wispr Flow, though that comparison comes from the repo description.

Read more on interesting repos · GitHub

Your Twitter/X history is hard to search and hard to hand to an AI agent. Birdclaw takes the archive zip you can download from X and loads it into SQLite on your own machine. Tweets, DMs, likes, bookmarks, profiles, media and follow edges all end up in the same tables.

Search uses SQLite's FTS5 full-text index, so local reads are fast and make no network calls. You can use it through a web app on localhost, a CLI with JSON output, or an optional MCP server that exposes cached tweet search and thread lookup as read-only tools.

Live data is opt-in. When you ask, it syncs your timeline or bookmarks through xurl or an existing bird install. The README says it is for people who want their own history without a cloud backend, and that it is not affiliated with X Corp.

Read more on interesting repos · GitHub

Tolaria helps you manage a large collection of notes. Each note is a markdown file with YAML frontmatter, and the whole folder (a vault) is a git repository. You get version history and can sync with any git remote.

The app works offline and needs no account or subscription. If you stop using it, your files still work in any editor. Types in Tolaria are just categories that help you find notes. There are no required fields and no validation.

Because the notes are plain files, AI agents can read and edit them directly. The README names setup paths for Claude Code, Codex CLI and Gemini CLI, and it provides an AGENTS file for agents to follow. The author says they use it with a vault of more than 10,000 notes.

Read more on interesting repos · GitHub

Making a slide deck by hand takes a lot of time, and hosted tools like Gamma, Canva and Beautiful AI tie you to their subscriptions and servers. Presenton is an open-source alternative you run yourself, either as a Docker container or as a desktop app for Mac, Windows and Linux.

You give it a prompt or a document. A language model writes the slides, an image provider adds pictures, and a template lays them out. You can then drag and edit slides in the browser and export the deck as PowerPoint (PPTX) or PDF.

You choose the models. The README lists Ollama, LM Studio, OpenAI, Gemini, Anthropic, Azure OpenAI, Amazon Bedrock and any OpenAI-compatible endpoint. Local models keep your data on your machine. It also offers an API and a built-in MCP server, so other programs and agents can generate decks too.

Read more on interesting repos · GitHub

Standups produce updates like "I started the ICE restart ticket" that someone then has to type into a board. This demo lets the board listen instead. People join a shared room from the browser, speak naturally, and the cards change on everyone's screen.

The Go server uses Pion WebRTC to receive each participant's audio and mixes it into one stream. It sends that stream to an OpenAI Realtime peer. The model is given instructions and a set of tools, and it calls those tools (function calling) to create, move, tag, edit or delete tickets. The server then broadcasts the new board state to the room.

It is a demo, not a product. There is no authentication, so anyone who can reach the URL can join. Background audio can also be misread as board updates, so the README suggests headphones. The instructions, tools and starting cards live in one file, kanban.go, which makes it easy to read and adapt.

Read more on interesting repos · GitHub

Clicky is an AI buddy that sits next to your cursor on a Mac. You hold a keyboard shortcut and speak. It captures your screen, sends your words and the screenshot to Claude, and reads the answer back to you.

Under the hood, audio streams to AssemblyAI for transcription, Claude replies over streaming, and ElevenLabs turns the reply into speech. Claude can add a tag with screen coordinates to its answer, and the app uses it to fly a blue cursor overlay to the right control, even across several monitors. A small Cloudflare Worker holds the API keys so they never ship inside the app.

This repo is the open-source snapshot under the MIT license. The author says new work is now private and the latest version is a separate download, so treat this as a base to fork and tinker with, not the current product.

Read more on interesting repos · GitHub

Omi is a personal memory tool. It listens to your conversations and watches your screen, turns the audio into text, and writes summaries and to-do items from what it captured. You can also chat with an AI that draws on everything you have seen and heard.

The repo is a full stack, not a single app. A Python backend (FastAPI, Firebase, Redis, Deepgram for speech-to-text, and LLMs) does the processing. Clients feed it audio: a Swift macOS app, a Windows app, a Flutter mobile app, and open hardware (an nRF-based wearable and an ESP32-S3 camera glasses kit) that connects over Bluetooth.

The README does not compare Omi to named alternatives. What stands out is that the clients, backend, firmware and hardware designs live in one open repo, along with an app system and SDKs for extending it.

Read more on interesting repos · GitHub

DeepTutor is a learning workspace built around large language models. You load in documents, books, folders or other sources, and the tutor answers from them, cites where answers came from, and keeps track of your progress across sessions.

It is a Python backend with a Next.js web interface. It builds knowledge bases with retrieval engines such as LightRAG and PageIndex, and connects to many model and search providers. The README also mentions a command-line interface that agents can use, and partners you can reach through chat apps such as Mattermost and WeChat.

The README excerpt is mostly a release log, so the details here come from those notes. There is an arXiv paper, and the project is under active development, with several releases in September 2026.

Read more on interesting repos · GitHub

Proof SDK is the code behind Proof, a hosted collaborative document editor made by Every. The SDK gives you the pieces: a markdown editor, a realtime collaboration server, a provenance model, and an HTTP bridge for agents.

Provenance means tracking where each piece of text came from. The editor supports comments, suggestions, and rewrite operations. The server stores documents in SQLite and exposes them through a set of HTTP routes.

The agent bridge is the distinctive part. An agent can read document state, see marks, post comments and suggestions, make edits, announce presence, and poll for events, all over plain HTTP. The README does not compare it to other editors, so this page doesn't either.

Read more on interesting repos · GitHub

VoiceInk is a Mac app for dictation. You press a shortcut, speak, and the text appears where you were typing. The README says transcription runs offline, so your voice data is not sent anywhere.

It runs local speech models: whisper.cpp for OpenAI's Whisper, FluidAudio for Parakeet, and others such as SenseVoice Small. The app adds the parts around the models: global shortcuts, push-to-talk, and a personal dictionary for names and jargon.

It also has modes that switch settings by app or URL, plus an optional AI assistant mode. The source is GPL v3, but the author is not accepting pull requests. The paid download adds automatic updates and priority support, and comes with a free trial. You can also build it yourself.

Read more on interesting repos · GitHub