OSSphere @ossphere_dev

Traditional VPNs make you feel secure while making everything slower, harder to configure, and more painful to maintain. @apenwarr took WireGuard — the fastest, most modern VPN protocol ever built — and made it so simple that connecting any two devices on the planet takes 30 seconds. tailscale/tailscale — 32,800 GitHub stars. Updated today. 230 releases. BSD-3-Clause. The networking tool developers actually enjoy using. Here's what zero-config mesh networking gives you: → Install tailscale on any two devices — they find each other automatically through NAT, firewalls, and CGNAT, no port forwarding, no static IPs, no configuration → Magic DNS — every device accessible by name instantly: laptop.tail-xxx.ts.net — works everywhere, always → Tailscale SSH — SSH to any machine without managing keys, rotating certificates, or touching authorized_keys ever again → Taildrop — AirDrop-style file sharing across your entire tailnet → Exit nodes — route all traffic through any device you own → Subnet routers — bring entire network segments into your tailnet → ACLs and device posture — zero-trust access controls built in → Kubernetes Operator — native k8s integration, GA → Aperture AI Gateway (Feb 2026) — distribute and visualize AI model access across your entire organization → Peer Relays and Tailscale Services — both GA Feb 2026 → Free tier: 3 users, 100 devices — no credit card → macOS, Windows, Linux, iOS, Android, Raspberry Pi WireGuard does the encryption. Tailscale does everything else. Discovered on OSSphere : ossphere.dev/tailscale/tail… What's the first two devices you'd connect the moment you install Tailscale? Drop it below 👇 #Tailscale #OpenSource #WireGuard #Networking #BuildInPublic #SelfHosted #DevOps

x.com/i/article/2066…

Most AI agents start cold. They know nothing about you until you type it. Every session begins from zero — no context, no history, no understanding of what you actually work on. @senamakel asked: what if the agent knew everything about your work before you typed a single word? tinyhumansai/openhuman — 32,600 GitHub stars in 6 weeks. #1 on GitHub Trending and Product Hunt in the same week. v0.57.44 shipped yesterday. Still accelerating. Here's what memory-first AI actually looks like: → Connects to 118+ services via one-click OAuth: Gmail, GitHub, Slack, Notion, Stripe, Calendar, Drive, Linear, Jira and more → Every 20 minutes: polls every connected account automatically — new emails, commits, calendar events, document edits — all fetched → Memory Tree: incoming data converted to Markdown, chunked, scored, and written to files you can open and edit yourself → Fully-local STT and TTS — Whisper + Piper, no cloud required → Async sub-agents — steer and await them while they run → Teammate agents that execute tasks end-to-end with timelines → Approval gate for high-cost runs — human-in-the-loop built in → Claude Code integration — OpenHuman memory shared over MCP → Skills catalog — installable modular capabilities → MCP end-to-end — OAuth flows and agent bridge included → Built in Rust + TypeScript + React 19, packaged as Tauri v2 → macOS, Windows, Linux — GPL-3.0 licensed The founder tried to set up an AI agent for his dad. Three hours of YAML and terminals later, they both gave up. OpenHuman came from that conversation. Discovered on OSSphere : ossphere.dev/tinyhumansai/o… What's the first thing you'd want your personal AI agent to know about you before you start your workday? Drop it below 👇 #OpenHuman #OpenSource #AIAgents #LocalAI #BuildInPublic #PersonalAI #Rust

Something quietly happened to the open source licensing landscape in the last three years. And most developers missed it. The pattern: a project grows, gets adopted at scale, and a cloud provider starts offering it as a managed service without contributing back. The maintainers — who spent years building it — watch someone else profit from their work. The response has been a wave of license changes: → HashiCorp changed Terraform to BSL 1.1 (2023) Community forked → OpenTofu under Linux Foundation → Elasticsearch changed to SSPL (2021) AWS forked → OpenSearch, still under Apache 2.0 → Redis changed to RSAL/SSPLv1 (2024) Community forked → Valkey, now Linux Foundation → MongoDB changed to SSPL (2018) AWS forked → DocumentDB (proprietary) → Confluent changed Kafka components to BSL (2023) Community maintains → Apache Kafka unchanged The pattern is consistent: BSL and SSPL block cloud providers from reselling. The community forks and keeps the Apache/MIT version alive. Neither side is wrong. Both are rational. The real question for developers: which license tells you something important about the project's future — and yours? ossphere.dev How do you evaluate OSS licenses before adopting a dependency? Drop it below 👇 #OpenSource #Licensing #BSL #SSPL #BuildInPublic #OSS #SoftwareEngineering

Intercom starts at $39/seat/month. Zendesk starts at $55/seat/month. Salesforce Service Cloud starts at $75/seat/month. For a 10-person support team that's $4,680–$9,000/year — before a single AI add-on is enabled. @chatwootapp is the open source answer. 33,000 GitHub stars. 15,000+ businesses. 7,800+ forks. MIT licensed core. Updated yesterday. Here's what one unified support inbox gives your team: → Live chat, email, Facebook, Instagram, WhatsApp, Telegram, Line, SMS, TikTok, and API — all in one conversation view → Captain AI — the native AI agent that ships in v4.14: - Assistant: answers customer questions autonomously, hands off to human when it's genuinely stuck - Co-Pilot: drafts and translates replies for agents live - Smart FAQs: identifies gaps in your knowledge base - Memories: retains per-customer context across sessions → Slack integration — manage conversations without leaving Slack → Shopify integration — view and manage orders inside Chatwoot → Linear ticket creation — engineering escalations in one click → Google Translate — real-time message translation built in → Pre-chat forms, CSAT reports, conversation labels and teams → Live View — monitor all ongoing conversations in real time → DigitalOcean 1-Click deployment — Kubernetes-ready → Python SDK, mobile app (React Native) — full ecosystem → MIT licensed core — commercial license only for enterprise/ directory Discovered on OSSphere : ossphere.dev/chatwoot/chatw… What's the biggest pain point with your current customer support stack? Drop it below 👇 #Chatwoot #OpenSource #CustomerSupport #SelfHosted #BuildInPublic #Intercom #SaaS

The developer who ships is more valuable than the developer who perfects. Open source taught me that more clearly than any job ever did. In a closed codebase, you can hide behind "it's not ready yet." Nobody sees the unfinished branch. Nobody judges the draft PR. The cost of not shipping is invisible. In open source, the cost is visible immediately: → The issue stays open and the community moves on → The PR sits unmerged and someone else solves it → The feature never lands and the fork gains the stars → The maintainer disappears and the project dies Open source has no room for perfectionism. Not because quality doesn't matter — it does, enormously. But because a good solution shipped beats a perfect solution that never leaves the branch. The developers who contribute most to open source projects share one trait: they have a low threshold for "good enough to share." They file the rough issue. They open the imperfect PR. They ship the v0.1 README. And then they iterate. In public. With feedback. The feedback loop that makes open source software better is the same one that makes open source contributors better. You can't get it without shipping something first. ossphere.dev What's the first OSS contribution you were most nervous to ship? Drop it below 👇 #OpenSource #BuildInPublic #ContributeToOSS #GitHub #SoftwareCraft #OSS #DeveloperGrowth

Most AI chat UIs are wrappers around one model. @mckaywrigley built the one that works with any model. 33,100 GitHub stars. 9,500 forks — more forks than almost any AI UI repo on GitHub. Built in TypeScript. MIT licensed. mckaywrigley/chatbot-ui — the open source foundation thousands of developers used to ship their own AI chat products. Here's what "AI chat for any model" actually gives you: → Connect any model — OpenAI, Anthropic, Google, and beyond — all in the same clean interface → Multiple workspaces — organize chats by project, client, or context → Assistants — custom system prompts and configurations saved and reusable → File uploads — chat with documents directly in the UI → Multi-modal — images and text in the same conversation → Message history — persistent, searchable, organized → Supabase backend — PostgreSQL + Auth, open source all the way down → Self-host locally or deploy to Vercel in minutes → Official hosted version at chatbotui.com if you'd rather not self-host → Next.js + TypeScript + Tailwind CSS — clean, readable, hackable → 9,500+ forks — studied, extended, and deployed by developers worldwide The ChatGPT interface is beautiful. It's also closed. ChatbotUI is the open source version of that idea — owned by you, running on any model you choose. Discovered on OSSphere : ossphere.dev/mckaywrigley/c… If you could build your own AI chat UI from scratch — what's the one feature every existing product is missing? Drop it below 👇 #ChatbotUI #OpenSource #AI #NextJS #BuildInPublic #SelfHosted #LLM

Before transformer models made everyone forget that NLP existed before 2023 — there was spaCy. 11 years old. 33,600 GitHub stars. 4,687 forks. 139,000+ GitHub projects depending on it. 1,000+ companies running it in production. @honnibal built the industrial-strength NLP library that has been quietly powering production text processing since 2015 — and it's still the standard. Here's what pip install spacy actually gives you: → Tokenization — accurate, language-aware, production-fast → Named Entity Recognition — people, orgs, locations, dates, and custom entity types you define → Part-of-speech tagging — every token labeled and explained → Dependency parsing — syntactic structure of every sentence → Text classification — train custom classifiers on your data → Lemmatization and morphological analysis — root forms, inflections → Rule-based matching — Matcher, PhraseMatcher, EntityRuler for pattern-based extraction no LLM can beat on precision → Multi-task learning with pretrained transformers — BERT, RoBERTa, and any HuggingFace model composable in a pipeline → Custom pipeline components — extend anything with Python → Production training system — config-based, reproducible, fast → spacy-layout — process PDFs and Word documents directly → GPU support via CUDA — scales to any document volume → 70+ languages — tokenization and training out of the box → MIT licensed — commercial open source, free forever LLMs generate text. spaCy understands it. Both belong in your production NLP stack. Discovered on OSSphere : ossphere.dev/explosion/spaCy Are you using spaCy in production — or did LLMs replace your NLP pipeline entirely? Drop it below 👇 #spaCy #OpenSource #NLP #Python #BuildInPublic #MachineLearning #TextProcessing

Six weeks ago we started this account with one belief: Open source is the most important force in software development — and most developers only see a fraction of what exists. Here's what six weeks of daily posting taught us about the open source ecosystem: → The best projects rarely trend. They just keep shipping. → Solo developers are building tools used by millions — quietly, without funding, without marketing. → The gap between OSS and paid SaaS has essentially closed in most categories. Quality is no longer the differentiator. → Discovery is still broken. The same 50 repos get shared while thousands of excellent projects sit invisible. → The community notices when you treat their work with respect. Every post tagging a maintainer got a response. → Developers are hungry for honest takes — not feature lists. → The AI stack is almost entirely open source now. That happened faster than anyone predicted. → Maintainer burnout is real, underfunded, and undertalked. → The best OSS projects have one thing in common: a README written for the user, not the author. We're just getting started. Every week we find projects that deserve 10x more attention than they're getting. That's not changing anytime soon. ossphere.dev What's the most useful thing we've shared so far? Drop it below 👇 #OpenSource #BuildInPublic #OSSphere #GitHub #DeveloperCommunity #OSS #SixWeeks

Standard RAG finds the right paragraph. It chunks your documents, embeds them, retrieves the closest chunks, and hands them to the model. That works for simple factual lookups. It falls apart on complex questions that span your entire corpus. Microsoft Research built the answer. Published the paper. Open sourced the code. microsoft/graphrag — 31,600 GitHub stars, MIT licensed, and the architecture that enterprise AI is moving toward in 2026. Here's what graph-based retrieval actually does differently: → Builds a full knowledge graph of your corpus first — entities, relationships, and community structures extracted by an LLM before any query is ever run → Leiden algorithm detects communities across the graph — automatic hierarchical clustering of related concepts → Community summaries generated at every hierarchy level — the model understands your data structure, not just your chunks → Local search: entity-focused retrieval for precise factual queries → Global search: corpus-wide thematic analysis using community summaries — answers questions no chunk could answer alone → 70-80% win rate over naive RAG on comprehensiveness and diversity → 2-3% token use per query vs full-text summarization — dramatically cheaper at scale → Auto-tuning: graphrag prompt-tune discovers entity types from your documents — no manual prompt engineering needed → BenchmarkQED evaluation framework included → MIT licensed — 3,300+ forks, Microsoft Research backed One warning the team is clear about: indexing is expensive. Start small. Understand costs before scaling. Discovered on OSSphere : ossphere.dev/microsoft/grap… What's the hardest multi-hop question you'd want a RAG system to answer across your documents? Drop it below 👇 #GraphRAG #OpenSource #RAG #KnowledgeGraph #BuildInPublic #Microsoft #AIEngineering

MCP changed something fundamental about how AI agents work. Not the models. Not the prompts. The plumbing. Before MCP, every AI tool integration was a custom job. Write a function. Wrap an API. Handle auth. Parse responses. Do it again for the next tool. And the next one. After MCP, the pattern is different: → One protocol. Every tool speaks it. → One config. Every MCP client reads it. → Connect once. Use from any compatible agent. → Playwright MCP: your agent controls any browser → GitHub MCP: your agent reads PRs and opens issues → Postgres MCP: your agent queries your database directly → Slack MCP: your agent sends messages and reads channels → Notion MCP: your agent reads and writes your docs → Filesystem MCP: your agent reads and writes local files The power isn't any single MCP server. It's that they all speak the same language. An agent with 10 MCP servers connected isn't doing 10 things. It's doing anything that requires any combination of those 10 things. That's a qualitative shift. Not a quantitative one. We're in the early innings of what MCP-native agents can do. The servers being built right now are the infrastructure layer of the next generation of software. ossphere.dev Which MCP server has been most useful in your workflow? Drop it below 👇 #MCP #OpenSource #AIAgents #BuildInPublic #LLM #ClaudeCode #DeveloperTools

Most AI browser agents work by taking screenshots and asking a vision model what to click next. That's slow. It's expensive. It breaks on dynamic content. And it requires a GPU just to read a button label. @Microsoft shipped a better architecture in March 2026. microsoft/playwright-mcp — 33,900 GitHub stars in 3 months. v0.0.76 shipped June 10, 2026. Shipping almost daily. The standard bridge between AI agents and the web is here. Here's what accessibility-snapshot browser automation gives you: → Structured accessibility tree instead of screenshots — the AI reads the page like a screen reader, not a camera → Lightweight and deterministic — no vision model required, no GPU, no hallucinated element locations → Full browser control via MCP — navigate, click, type, fill forms, wait for elements, take screenshots when needed → One line to add to any MCP client config: npx @playwright/mcp@latest — Claude Desktop, Claude Code, Cursor, and any MCP-compatible client instantly connected → Docker image available — containerized browser automation for CI/CD and agentic pipelines → Node 24 compatible — production-ready runtime support → Built on Playwright (75,000 stars) — Chromium, Firefox, WebKit — the most trusted browser automation foundation → Apache 2.0 licensed — 2,800+ forks, Microsoft-backed Screenshot when you need to see. Accessibility tree when you need to act. That's the right architecture. Discovered on OSSphere : ossphere.dev/microsoft/play… What's the first browser workflow you'd automate if your AI agent could control any website? Drop it below 👇 #PlaywrightMCP #OpenSource #AIAgents #BrowserAutomation #BuildInPublic #MCP #Microsoft

March 2023. ChatGPT could only read and write text. A team at Microsoft Research asked: what if ChatGPT could see, draw, and edit images — by routing tasks through specialized visual models? They built it. Open sourced it. 34,600 GitHub stars. One of the most important proofs of concept in the history of multimodal AI. chenfei-wu/TaskMatrix — Visual ChatGPT and TaskMatrix.AI. Two ideas that changed how the field thinks about AI systems. Here's what they proved was possible in 2023: Visual ChatGPT: → Describe an image in chat — BLIP captions it for the model → "Generate a sunset over mountains" — Stable Diffusion executes → "Remove the person on the left" — InstructPix2Pix edits it → "Find the dog in the image" — GroundingDINO locates the bbox → "Segment the car" — Segment Anything generates the mask → Chain models in sequence — output of one feeds the next → Templates: pre-defined multi-model flows humans define once TaskMatrix.AI — the bigger idea: → Foundation model as the brain — plans and orchestrates → Millions of APIs as the hands — execute sub-tasks → Connect LLMs to any digital or physical domain → The architecture that agentic AI is still built on today The repo is no longer actively maintained. The ideas inside it are the architecture of modern AI agents. Discovered on OSSphere : ossphere.dev/chenfei-wu/Tas… What AI capability demo first made you think multimodal AI was going to change everything? Drop it below 👇 #TaskMatrix #VisualChatGPT #OpenSource #MultimodalAI #BuildInPublic #AI #MicrosoftResearch

The hardest thing about open source isn't the code. It's the decision to start. Most developers have an idea for a tool that would help them. Something small. Something specific. Something that would save them 20 minutes every week. They don't build it because: → "Someone has probably already built this" (Maybe. But not exactly for your use case.) → "It's not good enough to share" (curl was a weekend project. SQLite started as a toy.) → "I don't have time to maintain it" (Ship it. See if anyone cares. Then decide.) → "Nobody will use it" (You'll use it. That's already one person.) → "It needs more features before it's ready" (The MVP that ships beats the perfect version that never does.) The open source projects that changed how developers work didn't start as ambitious projects. They started as one person solving their own problem clearly enough that other people recognized it as theirs too. The bar for "good enough to share" is much lower than most developers think. A clean README. A working install command. The problem it solves stated in one sentence. That's enough to start. ossphere.dev What tool have you thought about building but never started? Drop it below 👇 #OpenSource #BuildInPublic #IndieHacker #GitHub #SideProject #OSS #DeveloperTools

@i_mika_el Honestly, that's what makes it interesting. Self-hosting isn't winning because ops disappeared. It's winning because the benefits are becoming big enough that more teams are willing to own the ops.

The self-hosting movement is the most interesting infrastructure shift happening in developer culture right now. Three years ago "self-hosting" meant Linux hobbyists with Raspberry Pis and too much free time. Today it means: → Engineering teams replacing $50,000/year SaaS stacks with $20/month VPS deployments → Startups owning their data from day one instead of discovering lock-in at Series A → Developers running their own AI inference, analytics, notifications, auth, and monitoring — on hardware they control → Regulated industries finally having a compliance path that doesn't involve trusting a third-party vendor → Individual developers building entire product stacks for the cost of a single SaaS subscription What made this possible? → Docker made self-hosting approachable for any developer → Tools like Dokploy, Coolify, and Caprover made it one-command → The open source quality gap with SaaS closed dramatically → VPS providers dropped prices while performance increased → The community built documentation, templates, and support The era of "you have to use their cloud" is ending. Not for everyone. Not for every use case. But for more developers every month. ossphere.dev What's the first SaaS tool you'd replace if setup took under 30 minutes? Drop it below 👇 #SelfHosted #OpenSource #DevOps #BuildInPublic #Docker #SaaS #IndieHacker

@zeropsio Agreed. The real question is how much infrastructure responsibility a team wants to own. Some teams want full control over the servers, while others prefer a managed layer as long as pricing stays predictable.

Vercel charges per build minute. Netlify charges per seat. Heroku charges per dyno. Railway bills per service, per month, compounding. Dokploy runs on YOUR server. You pay for the VPS. Nothing else. Ever. @getdokploy — 34,900 GitHub stars, 7.5 million Docker pulls, built primarily by one developer, @Siumauricio. v0.29.0 shipped June 11, 2026. Still accelerating. Here's what your own PaaS gives you: → Deploy any app — Node.js, PHP, Python, Go, Ruby, any stack → Native Docker Compose support — no format conversions → Build systems: Nixpacks, Heroku Buildpacks, Railpack, Dockerfile → Git providers: GitHub, GitLab, Gitea, Bitbucket — shared across org → Databases: MySQL, PostgreSQL, MongoDB, MariaDB, Redis, LibSQL → Automated backups to S3-compatible storage → Traefik integration — automatic SSL, routing, load balancing → Multi-node with Docker Swarm — scale to multiple servers → Real-time monitoring — CPU, memory, storage, network per service → MCP Server: 508 tools across 49 categories — manage Dokploy from Claude Desktop, Cursor, VS Code, Windsurf, or Zed → CLI: 449 commands for full terminal control → AI-powered debugging built into the dashboard → Template library — Plausible, PocketBase, Cal.com, one-click → MIT licensed — 2,600+ forks, updated yesterday One curl command to install. Your infrastructure forever. Discovered on OSSphere : ossphere.dev/Dokploy/dokploy What's the deployment platform your team is running on — and what's your biggest pain point with it? Drop it below 👇 #Dokploy #OpenSource #SelfHosted #DevOps #BuildInPublic #Docker #PaaS

@es_boba77 @ataiiam That's a great way to put it. Building the agent is one challenge, but designing the right level of autonomy is the bigger one. Trust comes from getting those boundaries right.

Every product is about to need a copilot. Not a chatbot. Not an autocomplete sidebar. A copilot that understands your app's state, reads your data, takes actions inside your UI, and waits for approval before doing anything irreversible. @ataiiam built the open source React framework for exactly that. 28,000+ GitHub stars. MIT licensed. Shipping actively. Here's what building an in-app AI agent actually looks like: → CopilotKit — one provider wraps your app, AI gets full context → useCopilotReadable() — give the agent real-time access to your app's state: current user, loaded data, UI state, everything → useCopilotAction() — let the agent call functions and update state in your app directly from a conversation → renderAndWait() — Human-in-the-Loop: agent pauses, renders a confirmation component, waits for explicit user approval before any irreversible action executes → CopilotChat — drop-in chat UI that knows your entire app context → CopilotTextarea — AI-powered textarea with inline suggestions → Custom React components rendered by agents mid-conversation — charts, forms, confirmation dialogs, interactive maps → AG-UI protocol — standardized agent-to-frontend communication, works with LangGraph, CrewAI, or any agent backend → MCP support — connect tools and data sources to your in-app agent → Self-host the CopilotKit Runtime or use CopilotKit Cloud → MIT licensed — open-core model, production-deployed worldwide The era of apps with AI bolted on is ending. The era of apps built around an AI copilot is here. Discovered on OSSphere : ossphere.dev/CopilotKit/Cop… What's the first action you'd give an AI copilot inside your own product? Drop it below 👇 #CopilotKit #OpenSource #AIAgents #React #BuildInPublic #LLM #ProductEngineering

@adelbucetta @ataiiam 100%. Feels like we're still discovering the right patterns. Building agents is getting easier, but knowing when they should act on their own vs. ask for human input is where the real design work starts.

@buildwtim @ataiiam Exactly. The difference between an assistant and an auto-clicker is trust. Users are comfortable delegating work when they stay in control of irreversible actions.

Shopify takes 2.9% + 30¢ on every transaction. Plus $29–$299/month for the platform. Plus a percentage on any third-party payment gateway. WooCommerce needs WordPress underneath everything. BigCommerce fees compound as your catalog grows. Magento requires an enterprise contract to deploy seriously. @medusajs built the alternative. Heineken chose it. Mitsubishi chose it. 34,400 GitHub stars. MIT licensed. medusajs/medusa — the world's most flexible commerce platform. Updated yesterday. Still shipping daily. Here's what headless commerce built on Medusa gives you: → Product catalog — unlimited products, variants, collections → Multi-region and multi-currency out of the box → Cart and checkout — fully headless, your frontend, your UX → Order management — fulfillment, returns, exchanges built in → Customer accounts, segments, and management → Promotions and discount engine — rules-based and flexible → Inventory management across locations and warehouses → Payment providers: Stripe, PayPal, and community plugins → Fulfillment providers — connect any logistics provider → B2B starter — wholesale pricing, company accounts, approval flows → Admin dashboard — full product, order, and customer management → Module system — swap or extend any part of the commerce logic → medusa-agent-skills — Claude Code skills for Medusa conventions → Self-hostable — your transaction data, your infrastructure → MIT licensed — 4,700+ forks, community of 14,000+ developers Zero transaction fees. Total architecture control. Discovered on OSSphere : ossphere.dev/medusajs/medusa What's the biggest limitation you've hit with your current commerce platform? Drop it below 👇 #Medusa #OpenSource #Ecommerce #Headless #BuildInPublic #Shopify #TypeScript