OSSphere @ossphere_dev

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

Auth0 charges per Monthly Active User. At 100,000 users that's a significant line item. At 1,000,000 users that's a board conversation. At 10,000,000 users it's a company-defining infrastructure decision. The Red Hat team built the answer in 2014. keycloak/keycloak — 34,800 GitHub stars, 8,400 forks, Apache 2.0. Open source identity and access management for modern applications. Used by banks, governments, and enterprises worldwide. Updated yesterday. A CNCF incubation project. Still the standard. Here's what one Keycloak deployment gives you: → Single Sign-On and Single Sign-Out — across every app in your fleet → Social login: Google, GitHub, Facebook, Microsoft — one integration → Identity brokering — connect to any external identity provider → LDAP and Active Directory federation — enterprise directory sync → OAuth 2.0, OpenID Connect, SAML 2.0 — all supported natively → Multi-factor authentication — TOTP, WebAuthn, passkeys built in → Fine-grained authorization services — attribute-based access control → Per-realm isolation — separate environments in one deployment → Admin console — full web UI for users, clients, and realms → Custom login themes — your brand, your UX, fully configurable → Event logging and audit — every auth event tracked and queryable → Brute force protection built in → Kubernetes Operator + Terraform provider available → 308 releases — ships every 2 weeks, GitHub Secure OSS Fund recipient Your users. Your auth data. Your infrastructure. Zero per-MAU tax. Discovered on OSSphere : ossphere.dev/keycloak/keycl… What's the auth stack your team is running in production right now? Drop it below 👇 #Keycloak #OpenSource #Auth #IAM #BuildInPublic #SelfHosted #Security

The open source AI coding tools released in the last 18 months have changed something fundamental about software development. Not the speed. Not the output quality. The ceiling. Five years ago, a solo developer could realistically maintain one serious product. Maybe two if they were disciplined. Today: → One developer with Claude Code and a structured agent workflow can maintain what used to require a team of three → One developer with Aider can ship features across a codebase they've never read — with clean Git history the whole way → One developer with OpenHands can run automated debugging pipelines that catch regressions before any human touches them → One developer with UI-TARS can automate the manual browser workflows that used to steal hours every week → One developer with a well-structured RAG pipeline knows their entire documentation corpus without memorizing it The question used to be: how many developers does this need? The question now is: how well can this one developer use their tools? The ceiling for individual developers has never been higher. The floor for undisciplined ones has never been lower either. The tools are open source. The skill is not optional. ossphere.dev What has AI tooling genuinely unlocked for you as a solo developer? Drop it below 👇 #OpenSource #AITools #BuildInPublic #SoloDeveloper #IndieHacker #ClaudeCode #DeveloperProductivity

Claude Code has the best AI coding UX ever built. But Anthropic's API isn't always the most affordable option for every task in your workflow. musistudio/claude-code-router — 35,000 GitHub stars, 2,900 forks, MIT licensed. Keep everything you love about Claude Code. Route the requests wherever you want underneath. Here's what intelligent model routing actually gives you: → Background tasks → DeepSeek or Gemini — fast and cheap → Complex reasoning → Claude — reserved for what needs it → Long context → route to the model that handles it best → Local tasks → Ollama — zero API cost, completely private → One config.json — all your routing rules in one readable file → Built-in transformers: Anthropic, DeepSeek, Gemini, OpenRouter, Groq, Volcengine, SiliconFlow — swap without touching your workflow → Request/Response transformation — adapts formats across providers → Custom transformers — extend to any provider via plugins → Preset system — installable community configurations → Web management UI — React + Vite dashboard for your router → GitHub Actions integration — route CI/CD agent calls the same way → npm install -g @musistudio/claude-code-router — running in minutes Claude Code is the foundation. The router is the cost and performance layer on top. Discovered on OSSphere : ossphere.dev/musistudio/cla… What percentage of your Claude Code requests are actually complex enough to need the frontier model? Drop it below 👇 #ClaudeCode #OpenSource #AITools #ModelRouting #BuildInPublic #LLM #DeveloperTools

The way we build software is changing faster than most developers are comfortable admitting. Three years ago the stack was: pick a framework, write the code, deploy it, monitor it. Repeat. Today the stack is increasingly: describe the behavior, let the agent scaffold it, review and own the output, ship it. The developers who are thriving in this shift share one trait. They didn't abandon engineering fundamentals. They deepened them. Because AI-assisted development exposes weak fundamentals faster than any code review ever did: → Vague requirements produce vague code — garbage in, garbage out → Poor system design gets scaffolded at scale — fast and wrong → Weak testing means AI-generated code ships broken, confidently → No architecture intuition means you can't review what the agent built → Shallow debugging skills mean you accept the first plausible fix The developers who use AI best aren't the ones who prompt the most. They're the ones who know exactly what correct looks like — so they can tell immediately when the AI is wrong. The fundamentals weren't replaced by AI. They became the filter that separates the developers who ship from the ones who generated a lot of code that doesn't work. ossphere.dev What fundamental skill do you think matters most in the AI era? Drop it below 👇 #SoftwareEngineering #AITools #BuildInPublic #DeveloperSkills #OpenSource #CodeQuality #AIAgents