Weekly · open-source AI · week 42, 2026
aurelio-finance + 14 more trending open-source AI repos · week 42, 2026
This week's trending open-source AI on GitHub — aurelio-finance, hermes-official-plugins, mcd-coupon-strategist, and more: the newest, Hacker News talk, a subfield spotlight, and the fastest-rising, drawn out of the noise.
Newest this week
Open-source personal finance app with an AI financial advisor: track your net worth, investments, ETFs, cash and debts on your own computer.
Python · 2,437 stars already
View on GitHub →What we said about aurelio-finance
This one pulls all your money into a single dashboard you run locally — net worth, investments, ETFs, cash, debts — then layers an AI advisor on top to make sense of it. It's called aurelio-finance, and the self-hosted angle is what caught my eye. Most personal finance tools in this space, think Mint or the hosted trackers, want your bank logins living on someone else's server. Aurelio flips that: your financial picture stays on your own machine, which matters a lot when the data is this sensitive. Being Python-based means it's approachable if you want to extend the categorization logic or wire in your own data sources. The AI advisor piece is the part I'd test carefully — treat it as a conversation partner for spotting trends, not a licensed planner. This is for developers who already distrust cloud finance apps and don't mind a little setup in exchange for control. At 2,400-plus stars this early, clearly a lot of people were waiting for exactly this.
The official plugins for Hermes Agent, built and maintained by Nous Research: Home Assistant, Spotify, Claude Subscription DirectSDK and more, one install away from the Hermes plugin catalog.
Python · 206 stars already
View on GitHub →What we said about hermes-official-plugins
This is a plugin pack that lets the Hermes agent actually touch the world around you — dimming your smart lights, queuing a playlist, or routing requests through your existing Claude subscription instead of a raw API key. It comes from NousResearch/hermes-official-plugins, the first-party catalog for the Hermes Agent. What I like here is the focus on real integrations people already live inside. Home Assistant means your agent isn't trapped in a terminal; Spotify makes it ambient; and the Claude Subscription DirectSDK is the quietly clever one, letting you lean on a plan you're already paying for rather than metered tokens. Compared to wiring these up by hand through LangChain tools or MCP servers, the pull here is that they're maintained by the same team building the agent, so breakage risk drops. If you're already running Hermes, this is the obvious next install. If you're agent-curious, it's a reasonable reason to try the stack — just remember these are early, 206-stars early, so treat the smart-home bits as promising, not production.
An open-source Python project on GitHub.
Python · 174 stars already
View on GitHub →What we said about mcd-coupon-strategist
Picture a tool that scrapes McDonald's app coupons, cross-references them against menu prices, and tells you the cheapest path to whatever you're actually craving. That's the premise behind mcd-coupon-strategist, a Python project that's quietly pulled in 174 stars by solving a problem almost everyone has pretended not to care about. Under the hood it's really a small optimization engine — combinatorial math dressed up in fast food — and that's what makes it interesting beyond the novelty. If you've ever wanted a gentle introduction to constraint solving or coupon-stacking logic without wading into dense academic examples, this is a surprisingly readable codebase to learn from. Compared to the usual deal-finder browser extensions, the open-source angle means you can see exactly how the recommendations get made, and fork it for whatever regional menu or loyalty program you're dealing with. It won't change your life, but it's a tidy demonstration of applying real algorithms to an everyday annoyance. Worth a look if you like practical Python with a sense of humor.
Talk of Hacker News
Let your AI agents paint big arrows, boxes and text on your Mac screen. One CLI, click-through, gone by itself. Skill for Claude Code and Codex. MIT.
Swift · 415 points on HN
View on GitHub →What we said about big-arrow-on-the-screen
Picture handing your AI agent a highlighter for your entire screen—it can draw a giant arrow pointing at the button you keep missing, box off a region, or drop a line of text, all floating on top of everything. That's big-arrow-on-the-screen, a tiny Swift CLI from Franz Enzenhofer. The clever bit is the overlay is click-through and self-dismissing, so it never traps your cursor or lingers after the point is made.
Where this earns its keep is agent workflows. When Claude Code or Codex is walking you through a task, "the file is in the sidebar" is vague—an actual arrow on the sidebar is not. It ships as a skill for both, so your agent can gesture at real pixels instead of describing them.
Compared to screenshot-annotation tools or heavier accessibility overlays, this is deliberately minimal: one command, no GUI, native macOS. It's for developers building agent tooling, and anyone tired of translating text instructions into screen locations by hand. MIT licensed, so fork it freely.
Your personal AI agent in your own Cloudflare account. Chat, memory, tasks, notes and scheduled reminders, deployed with one command: npx create-talorys@latest. Free-tier friendly, single-user, no telemetry.
TypeScript · 323 points on HN
View on GitHub →What we said about talorys
This one spins up a private AI assistant that lives entirely inside your own Cloudflare account — chat, persistent memory, task tracking, notes, and scheduled reminders, all deployed with a single npx command. It's called Talorys, and what makes it interesting isn't the feature list so much as where those features run. Instead of handing your conversations and reminders to someone else's SaaS dashboard, everything sits on Workers, D1, and the rest of the Cloudflare stack you control. It's explicitly single-user, free-tier friendly, and ships with no telemetry, which tells you who it's for: developers who want a personal agent but don't want to be the product. Compared to something like a self-hosted Dify or a LangChain app you'd have to babysit on a VPS, this trades flexibility for a genuinely low-friction setup — one command and you're running. The tradeoff is you're locked into Cloudflare's primitives, and single-user means it won't scale to a team. But as a weekend "own your assistant" project, it's a clean starting point.
The Agentic-Native Operating System
Rust · 47 points on HN
View on GitHub →What we said about Veda
Picture an operating system built from the ground up to run AI agents as first-class citizens, not as apps bolted onto a design from the seventies. That's the pitch behind Veda, an ambitious Rust project calling itself the agentic-native OS. Instead of treating an agent as just another process fighting for file handles and memory, Veda wants scheduling, permissions, and resource management to understand that agents reason, call tools, and spawn sub-tasks on their own.
Now, I'll be honest — this is early, and "operating system" is doing a lot of heavy lifting here. It's closer to an opinionated runtime than something you'll boot on bare metal tomorrow. But the Rust foundation is the right call for memory safety when you're sandboxing code an agent decides to run.
Compared to frameworks like LangGraph or AutoGen that live inside Python, Veda is thinking lower in the stack — about isolation and control, not just orchestration. If you're a systems developer curious where agent infrastructure goes next, this one's worth watching.
Subfield spotlight
Open Source Computer Vision Library
C++ · 91,154 stars
View on GitHub →What we said about opencv
Here's a toolkit that lets your code actually see — detecting faces, tracking motion, reading shapes and edges out of raw pixels. OpenCV has been the quiet backbone of computer vision for over two decades, and it's still showing up in new projects every week. What's worth noticing is how it sits underneath the flashier deep-learning frameworks rather than competing with them. You'll often grab a model from PyTorch or ONNX, then lean on OpenCV for the unglamorous work around it — capturing frames, resizing, color conversion, drawing boxes, cleaning up output. The C++ core keeps it fast, but most folks reach for the Python bindings and never look back. If you're comparing it to something like scikit-image, OpenCV wins on real-time performance and hardware support; scikit-image wins on readability for research. For embedded work, robotics, or anything processing a live camera feed, this is still the default answer. It's not trendy, it's load-bearing — and that's exactly why ninety-one thousand stars feels earned rather than hyped.
List of Computer Science courses with video lectures.
83,657 stars
View on GitHub →What we said about cs-video-courses
This is a hand-curated directory pointing you to full university computer science courses that happen to be recorded and freely watchable online. It's called Developer-Y's cs-video-courses, and what makes it quietly useful is the breadth — you're looking at algorithms, operating systems, distributed systems, machine learning, compilers, even quantum computing, all organized by topic with links straight to the lectures. Think of it less as a course and more as a map of what's out there, pulled from places like MIT, Stanford, and Berkeley.
Now, why does a plain markdown list have over eighty thousand stars? Because the hard part of self-teaching isn't finding one video, it's knowing the right sequence and which professors actually explain things well. This saves you the rabbit hole. Compared to paid platforms or something like roadmap.sh, there's no path-holding here — it assumes you can self-direct. So it's best for folks who already know how to learn and just want signal over noise. Bookmark it, pick one subject, and actually finish it.
12 Weeks, 24 Lessons, AI for All!
Jupyter Notebook · 69,704 stars
View on GitHub →What we said about AI-For-Beginners
This is a structured curriculum that walks you through artificial intelligence from the ground up, spread across twelve weeks and twenty-four lessons. It's called AI-For-Beginners, and it comes out of Microsoft's education team. What I appreciate here is that it doesn't just dump theory on you — each lesson pairs concepts with hands-on Jupyter notebooks, so you're reading and running code in the same breath. It covers the historical stuff like symbolic AI and neural network fundamentals, then moves into computer vision and natural language processing, and crucially it supports both PyTorch and TensorFlow, so you're not locked into one framework.
Compared to something like fast.ai, which throws you straight into practical deep learning, this takes a slower, more academic path — better if you want to actually understand why things work before you build. It's aimed at students, career-switchers, and developers who've been meaning to fill in the gaps. Self-paced, free, with quizzes to check yourself. If you learn best with a syllabus instead of random tutorials, this is a genuinely solid starting point.
Fastest-rising
An open-source Lean project on GitHub.
Lean · +14,005 stars this week
View on GitHub →What we said about math
This one lets you write mathematics as code that a computer can actually verify, line by line, with no room for hand-waving. It's called openai/math, and it's built in Lean, a proof assistant that's been quietly winning over mathematicians and formal methods folks for years. What's interesting here isn't just another proof library — it's OpenAI putting weight behind formalized math, which hints at where they see models heading: reasoning you can check, not just trust. For a long time formalization lived in projects like mathlib, the community-driven backbone of Lean math. This feels like a signal that machine-generated proofs need a verifiable playground, and Lean gives you exactly that — if the proof compiles, it's correct. Who's this for? If you're into theorem proving, AI reasoning research, or you've ever wanted your math to come with a compiler error instead of a reviewer's doubt, start poking around. Fourteen thousand stars in a week tells you the appetite is real. Clone it, open it in VS Code with the Lean extension, and see how it feels to prove something that genuinely can't be wrong.
Qwen3.8-Flash-Next on any consumer hardware: one-click install for Windows / Linux. Strata inference engine, OpenAI/Anthropic API on localhost, optional image input.
C++ · +12,413 stars this week
View on GitHub →What we said about Strata
This one lets you run Qwen3.8-Flash-Next locally, serving it through an OpenAI- and Anthropic-compatible API on your own machine, with image input if you want it. Strata is the engine underneath, written in C++, and the pitch is a genuine one-click install on Windows or Linux—no wrestling with CUDA versions or Python environments that break every Tuesday.
What stands out to me is the drop-in API surface. If you've already built against OpenAI or Anthropic, you point your base URL at localhost and most of your code just works. That's a quieter feature than it sounds, because it means you can prototype against a cloud model, then swap to local for privacy or cost without rewriting anything.
Compared to something like Ollama or llama.cpp, Strata is leaning harder into the turnkey experience and that dual-API compatibility rather than raw flexibility. If you're a solo dev or a small team who just wants a local endpoint running in minutes, that trade-off makes sense. Twelve thousand stars in a week says plenty of people agree.
Makes your AI agent think like the laziest senior dev in the room. The best code is the code you never wrote.
JavaScript · +7,399 stars this week
View on GitHub →What we said about ponytail
This one intercepts your AI agent before it starts cranking out code and asks the question a tired senior engineer always asks first: do we actually need to build this? Ponytail wraps around your agent's planning loop and pushes it toward existing libraries, deletions, and config changes before it reaches for a blank file. It's a constraint layer, not a code generator — which is a refreshing inversion of where most of this tooling has been heading.
The practical payoff is smaller diffs and less of that confident sprawl where an agent writes three hundred lines to solve a problem a flag would've handled. Compared to the autonomous-everything frameworks, this feels almost contrarian. It's betting that restraint is the harder skill to automate.
Who's it for? Teams already drowning in AI-generated pull requests, or anyone who's watched an agent reinvent a well-maintained package from scratch. It's early, the JavaScript-first integration is narrow, and the heuristics will need tuning. But the philosophy underneath it is sound, and seven thousand stars suggest a lot of you feel the same fatigue.
Shipping hard
Orca is the ADE for working with a fleet of parallel agents. Run any coding agent with your own subscription. Available on desktop, mobile and remote runtime.
TypeScript · 2,764 commits this week
View on GitHub →What we said about orca
Picture running not one coding agent, but a whole squad of them at once, each chipping away at a different task while you watch from a single control panel. That's the idea behind Orca, from the Stably team — it calls itself an ADE, an agent development environment, and the framing is more useful than it sounds. Instead of babysitting a single assistant in your editor, you're orchestrating a fleet across desktop, mobile, and a remote runtime, using your own model subscription rather than paying through someone's markup. That bring-your-own-key approach matters for anyone wary of lock-in or surprise costs. Compared to inline tools like Cursor or Copilot, Orca leans toward parallelism and oversight — closer in spirit to something like Conductor or the various agent-swarm experiments floating around. It's aimed at developers who've outgrown the one-agent-at-a-time workflow and want to delegate several threads of work. With nearly 2,800 commits this week, it's moving fast, so expect rough edges — but the direction is genuinely interesting.
Every agent's model. One place. Codex on DeepSeek, Claude Code on Kimi, from the menu bar.
Go · 529 commits this week
View on GitHub →What we said about magpie
Here's a tool that lets you swap the model behind your coding agents without touching their config files. It's called magpie, from yetone, and it lives quietly in your menu bar as a routing layer between agents and whatever model you actually want them talking to. So you can run Codex against DeepSeek, or point Claude Code at Kimi, mixing the agent you like with the backend that fits your budget or your latency needs.
Why does that matter? Most agents assume you're staying inside one vendor's ecosystem, and switching usually means digging through environment variables and endpoint overrides. Magpie centralizes that decision into one menu, which is genuinely handy if you're cost-comparing providers or testing which model handles your codebase best.
It's written in Go, so it's a lightweight native binary rather than another Electron app eating your RAM. This is for developers juggling multiple agents who are tired of per-tool setup. With 529 commits this week, it's moving fast, so expect rough edges, but the core idea is clean and solves a real friction point.
Never stop coding. Free MIT AI gateway: one endpoint, 359 providers (150+ free), 1200+ models Kimi, Claude, GPT, Gemini, GLM, DeepSeek, MiniMax. Works with Claude Code, Codex, Cursor, OpenCode, Cline & Copilot. Quota-aware auto-fallback, RTK+Caveman compression saves 15-95% tokens, MCP/A2A, Desktop/PWA. Built by hundreds of contributors
TypeScript · 457 commits this week
View on GitHub →What we said about OmniRoute
This one gives you a single API endpoint that routes requests across hundreds of model providers, so when one hits a rate limit or runs dry, it quietly shifts you to another. It's called OmniRoute, and what makes it interesting isn't just the sheer count of models — it's the quota-aware fallback logic that keeps your coding assistant alive when a free tier taps out mid-session.
Compared to something like OpenRouter, OmniRoute leans harder into the free-provider angle and ships its own token compression, claiming meaningful savings on prompts you'd otherwise pay full freight for. It plugs straight into Claude Code, Cursor, Cline, Copilot and the rest, so you're not rewriting your workflow to adopt it.
Who's it for? Honestly, anyone burning through tokens on experiments, or teams who want a resilient routing layer they control rather than a single vendor's uptime. The MIT license and the desktop and PWA options are nice touches. Four hundred and fifty-seven commits this week tells you the pace is real — worth watching how the free providers hold up under load.