Weekly · open-source AI · week 45, 2026
Mika + 14 more trending open-source AI repos · week 45, 2026
This week's trending open-source AI on GitHub — Mika, AnyModelBot, DSH-Android, and more: the newest, Hacker News talk, a subfield spotlight, and the fastest-rising, drawn out of the noise.
Newest this week
An on-device AI agent for Android that operates apps and the web to get things done for you.
Kotlin · 4,432 stars already
View on GitHub →What we said about Mika
Imagine handing your phone a task in plain language and watching it actually tap through apps to finish it — booking, searching, filling forms — without shipping your screen off to some cloud server. That's the pitch behind Mika, an on-device AI agent for Android written in Kotlin. What makes this one worth a look is the "on-device" part. Most phone automation agents today lean on remote models and screenshot streaming, which raises obvious privacy questions when the agent can see your banking app or your messages. Mika keeps that loop local, which is a meaningfully different trust model for anyone wary of surrendering their screen.
Realistically, this is early territory — on-device agents trade raw capability for privacy, so don't expect it to out-reason a frontier model yet. But if you're an Android dev curious about accessibility-service-driven control, or someone building personal automation that never leaves the hardware, Mika is a clean codebase to study. Over four thousand stars already tells me the appetite for private, hands-free phone agents is real and growing.
Open-source, local-first alternative to Grok Bot, OpenAI dots and Claude Duet. Your own AI bots on any model: GPT, Claude, Gemini, Grok, DeepSeek, Kimi, GLM and 70+ more through one Modelflare key. Reach them in the app, WeChat or Feishu; they remember you, work in the background, run commands and use your computer. Pay per use.
TypeScript · 816 stars already
View on GitHub →What we said about AnyModelBot
Imagine spinning up your own personal AI assistant that isn't locked to a single provider, and then reaching it from wherever you already live—your app, WeChat, or Feishu. That's the pitch behind AnyModelBot, a local-first project that routes through one Modelflare key to hit GPT, Claude, Gemini, Grok, DeepSeek, Kimi, GLM, and dozens more.
What makes this interesting isn't the model buffet—it's the orientation. Instead of being tied into one vendor's walled bot experience like Grok Bot or Claude Duet, you own the bot, it remembers you across sessions, runs in the background, and can actually execute commands on your machine. That last part is where I'd move carefully; agents touching your computer deserve real sandboxing.
Who's this for? Tinkerers who want vendor flexibility and messaging-platform reach without stitching together five SDKs. The pay-per-use model keeps it honest for low-volume personal setups. At 816 stars out of the gate, there's clear appetite—just read the permissions before you hand it the keys.
An open-source Kotlin project on GitHub.
Kotlin · 723 stars already
View on GitHub →What we said about DSH-Android
This is a Kotlin project built for Android that's quietly picking up momentum on GitHub. It's called DSH-Android, from jackwener, and the honest truth is the repo keeps its description lean — so what you're looking at is a codebase that's letting the stars speak before the README does.
Here's why I'd still keep an eye on it: seven-hundred-plus stars in Kotlin's Android ecosystem usually means a working developer scratched a real itch and other people recognized the pattern. Clean Kotlin projects that trend this fast tend to be either a sharp utility or a reference architecture folks want to learn from.
If you're an Android dev who likes reading source to understand a technique rather than waiting for a polished write-up, this is worth cloning and poking at. Compared to the heavier, documentation-first frameworks out there, something like this rewards curiosity — you learn by tracing the code.
My advice? Star it, watch the commit cadence, and check whether the docs catch up to the interest. That's often the real signal.
Talk of Hacker News
An NES CPU compiled into the weights of a stock Qwen3.5 model
Python · 292 points on HN
View on GitHub →What we said about nesformer
Here's something that genuinely bent my brain this week: someone took a 6502 CPU — the chip inside the original Nintendo Entertainment System — and trained a language model to *become* it, executing instructions purely through next-token prediction. That project is dbreunig's nesformer, and it's running on a stock Qwen3.5 model with no custom architecture, just the weights fine-tuned to emulate processor behavior cycle by cycle.
Now, is this a practical emulator? Absolutely not. A real NES emulator is a few kilobytes of deterministic C and runs millions of times faster. But that's the point. This is a probe into what transformers can internalize about structured, stateful computation. If a model can reliably simulate deterministic logic from the outside, it tells us something about how these systems represent algorithms internally.
So who's this for? Interpretability folks, mechanistic-reasoning researchers, and anyone curious about the boundary between learned behavior and actual computation. It's less a tool and more a beautifully weird experiment — the kind that reframes questions rather than answering them.
A language model that lives in a boot sector
Python · 98 points on HN
View on GitHub →What we said about bootgpt
Imagine firing up a machine and, before any operating system loads, getting a tiny language model answering you straight from the boot sector. That's exactly the trick behind bootgpt, which crams a working model into the 512 bytes that live at the very start of a disk. Now, let's be honest about what this is — nobody's shipping a chatbot in half a kilobyte and expecting Shakespeare. It's a beautifully absurd constraint exercise, the kind of project that forces you to really understand quantization, token tables, and just how little a model can get away with. If you've been treating LLMs as these giant opaque black boxes, stripping one down to the metal like this is genuinely clarifying. Compared to the usual tiny-model demos running in a browser or on a Raspberry Pi, bootgpt goes a level deeper, skipping the OS entirely. Think of it less as a tool and more as a teaching artifact — one that rewards anyone curious about where software actually begins. Clone it, boot it in a VM, and go read the source.
A phone line over USB-C: LTE data, SMS and voice calls with the call audio on USB. A 4-layer KiCad board designed by Koppar from a one-page spec.
60 points on HN
View on GitHub →What we said about koppar-lte-modem
Picture plugging a small board into your laptop and getting a real cellular connection — LTE data, text messages, and actual voice calls with the audio routed straight over USB-C. That's koppar-lte-modem, a four-layer KiCad hardware design from Koppar that turns a one-page spec into an open phone line you can build.
What makes this interesting isn't just the functionality, it's the provenance. You get the full board files, so you can inspect the layout, order your own, or fork it for a custom carrier project. Compared to grabbing a USB LTE dongle off a shelf, this gives you something those sealed modules never do: the schematic, the routing, and the freedom to modify.
Who's it for? Hardware tinkerers, folks building remote telemetry, or anyone who wants a scriptable modem they actually understand. Voice-over-USB is the standout detail here — most cheap modems punt on call audio entirely. If you've ever wanted a phone stack you can read end to end, this is a genuinely fun starting point.
Subfield spotlight
User-friendly AI Interface (Supports Ollama, OpenAI API, ...)
Python · 157,508 stars
View on GitHub →What we said about open-webui
This is a self-hosted chat interface that sits on top of whatever model backend you've already got running, whether that's a local Ollama setup or a remote OpenAI-compatible endpoint. Open WebUI has quietly become the default front door for people who want a polished ChatGPT-style experience without handing their conversations to someone else's servers. What makes it stick isn't just the clean interface — it's the extras layered on top: document retrieval, multi-model switching, user accounts and role permissions, even a plugin system for custom tooling. That last part matters if you're a small team wanting shared access without rolling your own auth. Compared to lighter wrappers like LibreChat or just hitting Ollama's raw API, this leans toward the full-featured, batteries-included end — which is great until you realize it's also a bigger thing to maintain. If you're a developer prototyping locally, or an admin who needs to give a handful of non-technical colleagues a safe sandbox, it's genuinely hard to beat. Spin it up in Docker and you're running in minutes.
The agent engineering platform.
Python · 148,857 stars
View on GitHub →What we said about langchain
This is the toolkit most people reach for when they want to stitch language models into something that actually does work—calling tools, querying databases, chaining steps together. LangChain has been around long enough to watch the whole agent space grow up, and it's quietly rebuilt itself along the way. Early on it got a reputation for too many abstractions wrapping too little, and honestly that critique landed. But the recent direction is leaner, with LangGraph handling the stateful, multi-step orchestration that agents really need. Where something like the raw OpenAI SDK gives you a blank canvas, LangChain gives you batteries-included connectors to hundreds of models, vector stores, and APIs, so you're not reinventing the plumbing. If you're prototyping and want to swap providers without rewriting everything, it earns its place. If you want total control over every token, you might find it heavy. Nearly a hundred and fifty thousand stars means the docs, examples, and Stack Overflow answers are all there when you're stuck—which, with agents, you will be.
100+ AI Agents, Agent Skills and RAG Apps - Free and Open Source.
Python · 143,697 stars
View on GitHub →What we said about awesome-llm-apps
This is a sprawling collection of ready-to-run example projects that show you how to actually build AI agents and retrieval systems, not just talk about them. It's called awesome-llm-apps, and what sets it apart from the usual awesome-list that's just a pile of links, is that these are working apps with real code you can clone and tinker with today. You've got agents, agent skills, and a big catalog of RAG setups, all in Python, all open source.
Where this earns its keep is the learning path it quietly creates. If you've read a dozen blog posts about RAG but never shipped one, this gives you a concrete starting point — pick a project close to your use case, read how they wired the retrieval, swap in your own data. It's less a framework and more a cookbook, which means you're not locked into anyone's abstractions.
Great for developers moving from tutorials to their first real agent, and for teams who want reference implementations before committing to a stack.
Fastest-rising
The design language that makes your AI harness better at design.
Rust · +12,303 stars this week
View on GitHub →What we said about impeccable
Here's something for anyone who's watched an AI code generator spit out a layout that technically works but looks like it was assembled in the dark. This one, Impeccable by pbakaus, is a design language meant to sit inside your AI harness and give it actual taste — a structured vocabulary for spacing, hierarchy, and visual rhythm that a model can reason about instead of guessing at. What's interesting is the framing: it's not a component library or a Tailwind replacement, it's a layer that teaches your generation pipeline what good design decisions look like. Written in Rust, so it's fast enough to run inline without slowing your tooling down. If you've been pairing with Claude or Copilot on frontend work and constantly nudging it toward sane margins, this is aimed squarely at you. Compared to just stuffing style rules into a system prompt, a dedicated design grammar is more consistent and far easier to version. Twelve thousand stars this week tells you plenty of developers share that frustration. Worth a look if your AI's output needs a design eye.
A coding-agent skill for multi-phase security audits with independently verified, machine-readable findings
JavaScript · +10,647 stars this week
View on GitHub →What we said about security-audit-skill
This one hands your coding agent a structured way to run security audits, breaking the work into distinct phases and then independently verifying each finding before it reports back. It's called cloudflare/security-audit-skill, and the detail I keep coming back to is that machine-readable output. Instead of a wall of prose from an LLM that may or may not be hallucinating a vulnerability, you get findings that another tool can parse, and critically, a verification step that checks the claim actually holds. That matters because the biggest problem with AI security tooling right now isn't finding issues, it's the false positives that bury your team in noise. Compared to a plain prompt asking Claude or Copilot to "review this for security problems," the phased approach forces discipline and the independent verification cuts the junk. If you're a solo dev or a small team without a dedicated security reviewer, this is the kind of thing that fits right into your existing agent workflow. Ten thousand stars in a week tells you people have been waiting for exactly this.
Independent Auditing of AI Agents. Run by human or the agent itself, to answer the most crucial question in the AI Agent Economy. Is the agent doing what is supposed to do? With iFixAi you can have this answer in less than 120 seconds.
Python · +9,598 stars this week
View on GitHub →What we said about iFixAi
Here's a project that checks whether your AI agent is actually doing its job—not just running, but behaving the way you intended. iFixAi positions itself as an independent auditor for agents, and that framing is worth sitting with for a second. We've spent the last year obsessed with building agents, orchestrating them, chaining tools together—but verification has lagged way behind. This flips the lens: instead of trusting the agent's own logs, you run an external check that can answer, in about two minutes, "is this thing on the rails?" The clever bit is that it can be triggered by a human or by the agent itself, which opens the door to self-monitoring loops. Compared to generic observability tools like LangSmith or Arize, which lean on tracing and metrics, iFixAi is explicitly about intent versus behavior—a narrower, sharper question. It's early, Python-based, and clearly resonating given the star surge. If you're shipping agents into anything with real consequences, this is the kind of guardrail layer worth watching closely.
Shipping hard
Orca is the place for working with a fleet of agents (e.g. Grok Bot, Claude Code). Run any coding agent with your own subscription. Available on desktop, mobile and remote runtime.
TypeScript · 4,254 commits this week
View on GitHub →What we said about orca
Picture running not one coding agent, but a whole fleet of them—Claude Code, Grok, whatever you prefer—all orchestrated from a single control plane. That's the pitch behind Orca, from the folks at Stably. What makes this interesting isn't just that it wraps existing agents; it's that it lets you run them against your own subscription across desktop, mobile, and a remote runtime. So you could kick off a refactor from your laptop, check progress from your phone on the train, and let the heavy lifting happen on a server somewhere.
If you've been juggling separate tools for each agent, this is an attempt to unify that mess. It's closer in spirit to something like a multi-agent dashboard than a single-vendor IDE plugin, which means you're not locked into one model's ecosystem.
It's TypeScript, it's moving fast—over four thousand commits this week alone—so expect rough edges. But if you're an engineer already living in agent workflows and tired of context-switching between them, Orca is worth cloning and poking at this weekend.
The agent that grows with you
Python · 896 commits this week
View on GitHub →What we said about hermes-agent
Picture an AI assistant that actually remembers your preferences, your projects, and the way you like to work—then adapts instead of forgetting everything the moment a session ends. That's the pitch behind NousResearch's hermes-agent, a Python framework built around the idea of an agent that accumulates context over time rather than starting cold on every run. Where most agent stacks lean on brittle prompt chains or heavy orchestration layers, this one treats persistence and personalization as first-class features, which is a meaningful shift if you've ever gotten tired of re-explaining yourself to the same tool.
The 896 commits this week tells you it's in a fast, formative phase—so expect sharp edges, but also real momentum and responsive maintainers. I'd aim this at developers comfortable building on top of evolving APIs, folks experimenting with long-running assistants or local workflows. If you're weighing it against LangChain or AutoGPT-style setups, the difference is philosophical: less about chaining steps, more about an agent that genuinely learns your patterns. Worth cloning if memory is the piece you've been missing.
The open source coding agent.
TypeScript · 838 commits this week
View on GitHub →What we said about opencode
Imagine a coding assistant that actually lives in your terminal, reads your codebase, and makes edits you can review before they ship. That's the pitch behind opencode from anomalyco, a TypeScript-built agent that's racked up 838 commits in a single week — which tells you the maintainers aren't just maintaining, they're sprinting.
What makes this one worth a look is the open-source angle. Tools like Claude Code and Cursor are polished, but they're closed boxes tied to specific providers. opencode lets you bring your own model, inspect how the agent reasons, and extend the pieces you don't like. For teams with compliance concerns or anyone who just wants to see what's happening under the hood, that transparency matters.
It's aimed squarely at developers comfortable at the command line — not a drag-and-drop crowd. The trade-off is maturity: that commit velocity means fast-moving features but also churn, so pin your versions if you're putting it in a real workflow. Still, it's one of the more serious open attempts at a general coding agent right now.