Show the full transcript (2,230 words)
Every week, developers release new open-source projects on GitHub and this is our weekly GitHub project update video covering the best of them. Top trending open-source GitHub projects this week, you'll discover useful and trending developer tools, from design editors to AI agent orchestration. Without wasting time, let's get started. Before we jump into today's project updates, here's a quick announcement for everyone.
We've launched a brand new YouTube channel called AI Agents Studio, dedicated entirely to AI agent projects, tutorials and tools. So if you're interested in staying up-to-date with the latest AI agent open-source projects, learning how to build your own agents, or exploring cutting-edge agent frameworks, make sure to check it out. Subscribe now to get weekly videos, in-depth guides and real-time project breakdowns. The link is right there in the description. Don't miss it. Alright, let's get into today's video.
Project number one, Claude Code Plugins Directory, Anthropic's official marketplace of Claude Code plugins. This is Anthropic's official open-source directory of high-quality Claude Code plugins. It's split into internal plugins maintained by Anthropic, like front-end design, code review, code modernization, security and language server integrations, and external plugins from partners and the community, such as GitHub, Linear, Playwright, Firebase and Terraform. Each plugin can bundle slash commands, agent skills and MCP servers and you install any of them from Claude Code with slash plugin install name at Claude plugins official. You should trust a plugin before installing it, browse it and extend Claude Code.
Before we move to the next project, let's introduce something for the engineers and developers. This part is sponsored. JetBrains just launched a new IDE plugin. It is called JetBrains Air in IDEs. JetBrains builds the tools a lot of us already use. Air in IDEs is their agentic workspace inside those IDEs. You bring the coding agents you already use: Codex, Copilot, Claude, Cursor, Junie or other ACP-compatible agents, and run them from one place. Keep several projects open, run sessions in parallel, see where something needs your attention. When an agent finishes, review the diff with the IDE tools you already trust: navigation, refactoring, tests, debugging. You stay in control.
If you want to try it without burning your own subscription, JetBrains is offering Junie Lite free for a limited time. That is their efficient baseline agent for normal tasks, not their strongest model. It is enough to test the workflow. You need a JetBrains account for that. Cloud runs still need a JetBrains AI subscription. So try this plugin right now. Okay, let's move to the next project.
Project number two, the Scientific Agent Skills, Gen 63 science skills for your coding agent. Scientific Agent Skills is an open-source library of 163 skills that turn a coding agent into a research assistant across biology, chemistry, medicine and drug discovery. Each skill is a markdown file that teaches the agent how to query a specific resource, connecting to over 100 databases and tools like PubMed, UniProt and the PDB. It works with Claude Code, Codex and Cursor and you install just the skills you need. Install it and run your science.
Project number three, ruflo, Multi-agent swarm orchestration for Claude Code. Ruflo is an open-source multi-agent orchestration harness for Claude Code. It lets you deploy swarms of specialized agents across machines, teams and trust boundaries, layering self-learning memory, federated communication, vector RAG and enterprise policy controls. Hooks automatically route tasks, spawn agents and learn from what works, so you use Claude Code normally. A CLI manages swarms, agents and memory and a policy layer caps spend and concurrency, gates sensitive actions and audits every decision. Install it and orchestrate a swarm.
Project number four, NCC, Performance optimization system for AI coding agents. ECC is an open-source system for making coding agents better engineers, working across Claude Code, Codex, OpenCode, Cursor and more. It layers on skills, memory, per-stack rules, security and a research-first workflow, so the agent plans before it builds, verifies changes with tests, reviews its own work from a fresh context and remembers what matters. It ships commands like review peer, learn and consult, plus a security layer, Agent Shield. Install it as a plugin or via setup. Level up your agent.
Project number five, NIM XERJ, A search engine for AI agents, Elasticsearch-compatible. XERJ is an open-source search engine for AI agents. Pointed at a folder, one command indexes code, docs, logs and PDFs, so an agent asks questions instead of reading files into context. Auto index sniffs each file and makes an index per data set. Code with tree sitter symbols and line numbers, plus CSV, JSON, SQLite, DOCX and mail. It speaks the Elasticsearch API, so clients work and XERJ net wires it into Claude Code and Cursor. Install it and let agents search.
Project number six, text-to-cad, Agent skills that give your agent CAD powers. text-to-cad is an open-source library of agent skills that give your agent CAD superpowers. Ask it to model a part and it writes and edits parametric CAD in Python. Build 123D on OpenCasCade, exporting STEP, STL, 3MF, GLB and DXF. A browser viewer previews the result, plus robot files and G-code and stable at CAD references let you request precise edits. Other skills cover robot descriptions, inverse kinematics and implicit CAD. Install it into Claude Code or Codex and model from text.
Project number seven, Agent Substrate, Run agent workloads at scale on Kubernetes. Agent Substrate is an open-source system built on Kubernetes that runs agent-like workloads at higher density and lower latency. It maps idle actors onto a pool of worker pods, suspending and resuming them with state intact for multiplexing. It keeps the Kubernetes control plane off the request path, isolates workloads with gVisor and stays framework-agnostic across 80K, LangChain, Claude Code and MCP. It runs on any cluster through a kubectl plugin and it's early. Deploy it and scale your agents.
Project number eight, Minds, OpenShell, Secure, policy-governed sandbox runtime for AI agents. OpenShell is NVIDIA's open-source, secure-by-design runtime for autonomous AI agents. Agents like Claude Code, Codex and OpenClaw run unmodified inside kernel-level sandboxes, governed by declarative YAML policy. You declare what each agent can touch, and OpenShell enforces it. Agents never see credentials, it injects them only into requests bound for approved endpoints, blocking exfiltration. Policy changes are formally verified, so risky access waits for review. It runs on prem in Docker or Kubernetes, with SDKs. Deploy it and sandbox your agents.
Project number nine, FreeLLMAPI, Stack free LLM tiers behind one endpoint. FreeLLMAPI is an open-source, OpenAI-compatible proxy that stacks the free tiers of 34 LLM providers, 635 endpoints, billions of tokens monthly behind one flash V1 endpoint plus any custom OpenAI-compatible endpoint. It routes and fails over automatically and a signed catalog syncs new models and quotas. Provider keys are AES 256 GCM encrypted in SQLite and your app sees one unified token. Launchers keep credentials out of Claude Code and Codex configs. It's local first for personal use. Install it and point your app at it.
Project number ten, Orchard, Orchestrate Tart VMs across Apple Silicon Macs. Orchard is OpenAI's open-source orchestrator for running Tart virtual machines across a cluster of Apple Silicon devices. A controller and workers let you create, run and manage a fleet of macOS VMs through a CLI or a built-in REST API. A local dev mode starts a controller and worker. It now runs Linux workers through V2, so one cluster can mix architectures and operating systems. It's written in Go and installs with Homebrew. Set it up and orchestrate your VMs.
Project number eleven, Nush OpenPencil, AI-native, open-source Figma alternative you can program. OpenPencil is an open-source AI-native design editor, a local first, MIT-licensed Figma alternative. It reads and writes native dot fig files and its own dot pen format, so you copy and paste nodes between it and Figma and there's a built-in AI assistant. Everything is programmable, a headless CLI and specs to queries, lints and exports to PNG, SVG, PDF, HTML and more. An MCP server exposes it to Claude Code, Codex and Gemini and a Vue SDK builds custom editors. Install it and design locally.
Project number twelve, Zierier Studio, A calm, open-source project management workspace. Zierier Studio is an open-source calm project management workspace. It gives you a home dashboard, an inbox, a my tasks view and projects you can see as a board, list, table, calendar, timeline or files, plus members, teams, activity and settings. It's built with plain JavaScript ES modules and CSS bundled by Vite, with no UI framework and no runtime dependencies. It's a polished front end ready to wire up to Rest and Postgrest. Deploy it and run your projects.
Project number thirteen, Insomnia, Keep a MacBook awake, lid closed, for agents. Insomnia is an open-source macOS menu bar app, an eye icon that lives in your MacBook menu bar. It has no Dock icon and no main window. Set how long your Mac stays awake, choose what happens when the lid closes, and see when something could not be undone. It turns off the display and keyboard backlight, saving their brightness. And if Wi-Fi drops, it can join your iPhone's hotspot. It's careful with permissions, a narrow sudoer's rule, owner-only files and signed releases. It needs a recent Apple Silicon Mac and is experimental. Use a ventilated surface, not a closed bag. Install it and run overnight.
Project number fourteen, BRUTAL Framework, Open-source C# framework for games and engines. BRUTAL is a free, open-source C# framework for making games, or building your own engine on top. It's a framework, not a full engine, a custom interop layer exposes native APIs, Vulkan and low-level C++ libraries: SDL, Gilaive, Jolt and Box2D physics, Fmod audio, Imgabi, Gltf to C# with low allocations and minimal overhead. From Ahu and RocketWerks, it already powers Kitten Space Agency and RimWorld's Waylith engine. It's free for commercial games, no royalties. Clone it and build.
Project number fifteen, librecalculator, Free, private calculators for everything countable. LibreCalculator is a free, open-source collection of calculators for everything countable. The universe in numbers. It gathers 72 calculators across math, geometry, unit conversion, finance, cooking, health, construction, physics, dates and statistics, and each one explains its answer rather than just a number. It's fast and private with no sign-up and no ads. Use it in a browser at librecalculator.com or run the desktop app on Mac, Windows or Linux, with a quick panel and keyboard shortcut. Open it and calculate.
Project number sixteen, GPT-with-taste, Better judgment and taste for GPT in Codex. GPT-with-taste is an open-source behavioral recipe that gives GPT and Codex better judgment and taste with less ceremony. From one recipe it compiles system instructions plus per-message steering through hooks. The prompt sets the behavior and a reminder follows each of your messages. Instead of hedging and self-justifying, the model is pushed to choose sensible defaults, preserve supported claims, synthesize useful collaborator results and finish work with care. It's backed by research on defensive writing. Install it and give GPT taste.
Project number seventeen, CC-stretch, Short replies that expand phrase-by-phrase on click. CC-stretch is an open-source Claude Code plugin that turns replies into stretched text, short by default, with each phrase expanding in place when you click it. You read the skim version, then open just the parts you want more on, so a long answer stays scannable. A full mode opens everything. It works in Claude Code's full screen TUI and everywhere else, every phrase shows opened, so nothing is ever hidden. Install it from its marketplace and read less, expand on demand.
Project number eighteen, DEsk, A skeuomorphic Wayland desktop that's literally a desk. DEsk is an open-source skeuomorphic Wayland desktop for Linux, where the desktop is a literal desk. Your windows are sheets on a walnut desk. The dock is a drawer, your apps are filed in binders on a bookshelf and notifications arrive in a post box. Nothing is a symbol, take a sheet and the app opens on it. The room is ray-traced, lit by one desk lamp. It's written in C and an early beta. Log in and work at your desk.
Project number nineteen, Remory, Learned memory that survives context compaction. Remory is an open-source project, learning residual memory for context compaction from Fudan's Fusion Lab. When an agent compacts its context into a summary, detail is lost. Remory supplements that summary with learned soft memory from its history. A memory network reads the model's hidden states and stores residual memory in a local SQLite file, keeping more of what mattered. It runs alongside Codex compaction, ships Quen and GLM checkpoints and needs two 80-gigabyte GPUs. Clone it and compact with memory.
Project number twenty, agentplay, Replay coding-agent sessions as a map and timeline. Agentplay is a free, open-source tool that replays your coding agent sessions as an interactive map and a timeline. A final answer hides the story: failed tests, approaches tried, errors fixed along the way. So agentplay reads your Claude Code, Codex, Cursor and Copilot sessions and shows what the agent did, where it stalled and how it got there. It runs locally with no tracing setup and no runtime dependencies and nothing leaves your machine. Run it with UVX and replay a session. Thanks for watching. See you in the next update.
Video
Video
Video
Video
