RepoCatAI | Sharing GitHub Projects for AI & Robot

Asimov v1: Open-Source Humanoid Robot Build your own 1.2m-tall bipedal robot with 25 actuated degrees of freedom. Learn more: github.com/asimovinc/asim…

[DEEP DIVE] Meta-Harness — Stanford’s framework for making agents optimize their own scaffolds Most AI evals tweak the model or prompt. This repo goes after the underrated layer: the harness around the model. · Searches over “model harnesses,” not just prompts → what to remember, retrieve, expose, and pass back into the model · Treats agent scaffolding as something you can optimize end-to-end → useful when your base model is fixed but your workflow is not · Includes a text classification example for memory-system search · Includes a Terminal-Bench 2 example for evolving coding-agent scaffolds · Designed for new domains through an onboarding flow → point a coding assistant at the guide and generate a domain_spec.md · Lets builders experiment with different proposer agents → Claude Code is assumed in examples, but wrappers can be swapped · Comes from a Stanford IRIS Lab paper, with reference experiments included Why it matters: A lot of agent performance lives outside the model weights: memory, tools, retrieval, context layout, and what the model sees while working. Meta-Harness gives researchers and serious builders a way to search that design space instead of hand-tuning scaffolds by instinct. Follow @repocatai_git for more AI / Agent / Robotics drops 🚀

[BOOKMARK THIS] awesome-agentic-ai-zh — a full AI Agent learning map from LLM basics to multi-agent systems If you’ve been trying to learn agents from random threads, papers, and half-finished demos, this repo is the organized path you wanted. · 8-stage roadmap from “what is a token?” to building multi-agent systems · Two tracks: use CLI agents well, or learn to build agents from scratch · 145+ curated projects with notes on who they’re for and how to run them · 23 starter exercises across the stages → small enough to finish, real enough to teach tool use, RAG, MCP, and agents · Covers Claude Code ecosystem deeply: MCP, Skills, Plugins, SDK, subagents · Includes local LLM paths like Ollama, llama.cpp, LocalAI, and MLX · Chinese AI ecosystem catalog: DeepSeek, Zhipu, Kimi, MCP/Skill resources · Maintained in Traditional Chinese, Simplified Chinese, and English Why it matters: Agent learning is messy because the field mixes APIs, prompting, tools, memory, browsers, sandboxes, frameworks, and product workflows. This repo turns that chaos into a study plan an indie dev or AI engineer can actually follow over weeks, not just bookmark and forget. Follow @repocatai_git for more AI / Agent / Robotics drops 🚀

[NEW DROP] ViMax — an open-source agentic video studio that turns ideas, novels, or scripts into films Most video generators stop at “make me a clip.” ViMax is aiming at the whole production pipeline: story, cast, scenes, audio, and final video. · Idea2Video: start with a rough concept and let agents build the story → it handles scriptwriting, storyboarding, character design, and production flow · Novel2Video: adapt long fiction into episodic video → useful if you want character tracking and scene-by-scene compression, not random snippets · Script2Video: bring your own screenplay and generate from that structure · Multi-agent workflow: director, screenwriter, producer, and generator roles work together · Built for longer storytelling, not just isolated 5-second demos · Character and scene consistency are treated as core problems, not afterthoughts · Python 3.12 + uv-ready setup, MIT licensed, builder-friendly stack Why it matters: If you’re experimenting with AI filmmaking, the painful part is not one image or one clip. It’s keeping the narrative, characters, shots, and production steps coherent across the whole thing. ViMax is interesting because it treats video generation like an agent workflow, not a single prompt box. Follow @repocatai_git for more AI / Agent / Robotics drops 🚀

[BIG UPGRADE] llama.cpp — run serious LLMs locally with one C/C++ inference engine This is the repo that quietly became the default “can I run this model on my machine?” answer. · Runs LLMs with minimal setup across laptops, desktops, servers, and weird hardware · One command can pull a GGUF model from Hugging Face and start chatting · OpenAI-compatible server built in → swap local models into apps that already speak the OpenAI API shape · Quantization from 1.5-bit to 8-bit → smaller models, lower RAM use, more chances your hardware actually fits · First-class Apple silicon support via Metal, Accelerate, and ARM optimizations · NVIDIA CUDA kernels, plus AMD HIP, Vulkan, SYCL, RISC-V, and CPU fallback · CPU + GPU hybrid inference → run models bigger than your VRAM by splitting work across hardware · Multimodal support is landing in llama-server, plus WebGPU experiments for browser GPU runs Why it matters: llama.cpp is not just a toy CLI for local chat. It is the low-level engine behind a huge chunk of local AI hacking: prototypes, offline assistants, private inference, edge deployments, editor completions, and “let me test this model before renting a GPU” workflows. If you build AI tools and you have not tried running a GGUF model through llama-server yet, you are probably overpaying for experiments you could run on your own desk. Follow @repocatai_git for more AI / Agent / Robotics drops 🚀

[BUILDER'S DREAM] Streambert is an open-source desktop app for ad-free streaming + downloads This is the kind of project you open “just to inspect the repo” and end up trying for an hour. · Cross-platform Electron app for Windows + Linux · Streams movies, TV series, and anime from inside one clean desktop UI · Zero ads and trackers by design → no popups, no browser junk, no analytics layer watching you · Built-in downloads using .m3u8 playlist links → save episodes locally and watch them in-app later · Subtitle download + management included · Anime mode pulls metadata from AniList instead of TMDB → better titles, covers, and info for anime libraries · Trending/home/search views powered by TMDB · Customizable interface + local watch/download library Why it matters: Streambert is interesting because it feels like a hacker-built media center, not a bloated streaming site wrapped in a browser. For indie devs, it is also a useful Electron case study: desktop packaging, external downloader integration, metadata APIs, subtitles, local state, and privacy-first UX all in one app. Educational and personal-use caveat applies: users are responsible for what they access and whether they have the legal right to access it. Follow @repocatai_git for more AI / Agent / Robotics drops 🚀

[BOOKMARK THIS] OpenWA is an open-source WhatsApp API gateway you can actually own If you’ve ever wanted WhatsApp messaging without a black-box vendor sitting in the middle, this is worth a look. · Run your own WhatsApp API over plain HTTP → send texts, media, reactions, bulk messages, and track delivery/read receipts · Multi-session support out of the box → manage multiple WhatsApp accounts from one instance · Full web dashboard included → sessions, webhooks, API keys, and docs without living in the terminal · Webhooks with HMAC signatures → receive real-time events and verify they actually came from your server · Pluggable storage stack → SQLite or Postgres, local files or S3/MinIO, memory cache or Redis · Built-in controls for serious deployments → API keys, rate limits, CIDR allowlists, proxies, audit logs, health checks · Docker-native setup → dev mode in one command, production profiles for Postgres, Redis, MinIO, Traefik · n8n integration → wire WhatsApp into automations without writing the whole workflow layer yourself Why it matters: OpenWA is for builders who want WhatsApp in their product, bot, CRM, internal tool, or automation stack without renting every feature from a SaaS API. You get the boring but critical pieces too: auth, dashboard, docs, storage, webhooks, and deployment knobs. Follow @repocatai_git for more AI / Agent / Robotics drops 🚀

[INSANE] oh-my-pi — an open-source coding agent with the IDE, debugger, and tools wired in Most coding agents still act like a chatbot staring at files. oh-my-pi feels closer to giving the model the same instruments you use as an engineer. · 40+ model providers, so you can swap brains without swapping workflow · 32 built-in tools for reading, searching, editing, running tasks, and more · LSP built into writes → renames, references, imports, and file moves use what your IDE already knows · Real debugger control via DAP → attach to lldb, dlv, debugpy, inspect frames, step through crashes · Persistent Python + Bun execution → the agent can run code, analyze data, chart results, and call its own tools · Bench-tuned edit formats → some models jump from “barely lands patches” to “actually useful” · Summarized file reads instead of context-dumping whole files · Rust core under the hood, TypeScript/Bun on the install path Why it matters: This is for people who want an agent that can work inside a real codebase, not just suggest diffs from the outside. If you build agents, devtools, CLIs, or weird automation stacks, oh-my-pi is worth studying because it treats the IDE and debugger as part of the agent surface. Follow @repocatai_git for more AI / Agent / Robotics drops 🚀

[DEEP DIVE] MuJoCo — DeepMind’s open physics engine for robots, bodies, contacts, and control If you’re building robot policies, RL environments, dexterous hands, or weird articulated machines, this is one of those repos you eventually run into. Here’s what makes it worth opening: · Simulates multi-joint bodies with contact, friction, constraints, and collisions · Built for speed: preallocated low-level data structures, tuned C runtime · Native interactive viewer via simulate → load a model, poke it, watch the physics respond in real time · Python bindings install with one pip command · Colab tutorials for basics, model editing, rollout, LQR, least-squares, and MJX · MJX brings MuJoCo-style simulation into JAX → useful when you want physics inside accelerated ML training loops · Differentiable physics tutorial for training locomotion with analytical gradients · Unity plugin if you want MuJoCo physics inside a game-engine workflow Why it matters: MuJoCo sits in the sweet spot between “research-grade simulator” and “something a serious builder can actually use.” For robotics hobbyists, AI engineers, and indie devs, it gives you a practical way to test control, locomotion, manipulation, and embodied agents before hardware gets expensive. Follow @repocatai_git for more AI / Agent / Robotics drops 🚀

[BUILDER'S DREAM] LangChain is the open-source agent stack for turning LLM ideas into real apps Most AI demos die when you need tools, memory, data, evals, or model swaps. LangChain is the boring-useful layer that keeps your agent project from turning into glue code soup: · One standard interface for chat models, embeddings, vector stores, retrievers, and tools · Swap models without rewriting the whole app → test OpenAI, Anthropic, local models, or whatever comes next · Connect LLMs to real data and internal systems → docs, databases, APIs, search, files, custom tools · Build simple chains first, then graduate into agents when the workflow gets messy · Deep Agents adds planning, subagents, and file-system use for heavier tasks · LangGraph gives you controlled agent workflows → useful when “just let the agent decide” is too chaotic · LangSmith plugs in observability, debugging, evals, and deployment paths · Huge ecosystem of integrations, examples, templates, and community patterns Why it matters: If you’re building an AI product, LangChain gives you a shared vocabulary for the parts every LLM app eventually needs. It’s especially good for indie devs and AI engineers who want to prototype fast, then keep enough structure to ship something maintainable. Follow @repocatai_git for more AI / Agent / Robotics drops 🚀

[BOOKMARK THIS] UMI-3D — turn raw robot rosbag demos into policy-ready 3D manipulation data Most robot learning repos start at “here’s the model.” This one attacks the messier part: getting real-world 3D demos clean enough to train on. · Records human-guided manipulation demos as rosbag files → LiDAR, IMU, and camera streams captured together · Includes gripper calibration data collection → open/close cycles become usable motion range estimates · Fisheye camera intrinsic calibration built in → checkerboard workflow, scripts, and saved calibration outputs · LiDAR-camera extrinsic calibration for Livox MID-360 → aligns the 3D point cloud with the camera view · LiDAR-inertial SLAM pipeline → reconstructs the scene and estimates camera trajectory from motion · Converts raw recordings into aligned demonstrations → the boring synchronization and processing steps are part of the repo · Outputs Zarr datasets for policy training → usable with imitation learning stacks like Diffusion Policy Why it matters: If you’re building embodied AI, the bottleneck is often not the model — it’s collecting repeatable, synchronized, training-ready data from the real world. UMI-3D gives robotics builders a concrete pipeline from “I recorded a demo” to “I can train a manipulation policy on this.” Follow @repocatai_git for more AI / Agent / Robotics drops 🚀