Zane Chen

18 steps. From a chat loop to a full AI agent. Every line commented. Build Your Own OpenClaw walks you through building an AI agent from scratch. Tool use, memory, skills, cron, multi-agent routing, web dashboard. Each step is runnable code with a README that explains the design decision behind it. Phase 1: a working single agent. Phase 2: event-driven refactor so it runs across platforms. Phase 3: autonomous scheduling and multi-agent coordination. 1.7K stars on GitHub. github.com/czl9707/build-you… Follow for daily dev finds 🔔

Your Claude Code subscription is the cheapest agent runtime you're not using. ohmo wants it. It rides your existing Claude Code or Codex sub, no new API key, and sits in Slack/Telegram/Discord waiting for chores. Send it a task and it forks a branch, writes the code, runs the tests, opens the PR. Underneath sits OpenHarness, the actual harness: 43 tools, subagents, permission hooks, auto-compaction for multi-day sessions. MIT licensed, 114 passing tests, readable Python if you want to see how a harness works. ⭐ 15.4K #AI #DevTools github.com/HKUDS/OpenHarness Follow for daily dev finds 🔔

Every AI assistant forgets you the moment you close the tab. OpenHuman keeps a memory you can actually open: scored Markdown trees in SQLite, mirrored as an Obsidian vault you edit by hand. No vector soup. A background loop refreshes it every 20 minutes and writes your morning briefing, and tool output gets compressed up to 80% before it ever hits the model. Rust, local-first, early beta. The memory design alone is worth a look. ⭐ 36.3K #AI #OpenSource github.com/tinyhumansai/open… Follow for daily dev finds 🔔

Slides generated by AI still look like AI generated them. banana-slides fixes that by cloning whatever template you upload. One prompt or outline becomes a full deck. Upload source materials and it extracts content into each slide. Edit conversationally, export to PPTX. Free, open source, self-hosted. ⭐ 15.5K #AI #OpenSource github.com/Anionex/banana-sl… Follow for daily dev finds 🔔

AI labs are invisibly watermarking everything you generate and most people have no idea. watermarks-remover strips provenance marks from text and files across 8 formats. Unicode tracking characters, SynthID statistical watermarks, C2PA metadata, all of it. Covers Claude, Gemini, and OpenAI output. If you own the content, you control the metadata. ⭐ 5.8K #Privacy #OpenSource github.com/guillaumemeyer/wa… Follow for daily dev finds 🔔

Every time you ask AI for a diagram you get Mermaid slop with generic rounded boxes. diagram-design hands Claude Code 27 editorial diagram types as self-contained HTML + SVG. It reads your website and matches your brand colors in 60 seconds. No Figma. No color-picking sessions. Your architecture diagrams shouldn't look like a free flowchart tool. ⭐ 15.2K #OpenSource #ClaudeCode github.com/cathrynlavery/dia… Follow for daily dev finds 🔔

The gap between the cheapest and most expensive AI model for the same task is 200x. The expensive one gives you a prettier coffee shop. Netlify ran identical prompts through 11 models to build the same web app. Claude Opus 5 averaged 519 credits per run. One generation ate 1,055 credits in a single shot. DeepSeek V4 Flash did the same job for 2.4 credits. Same prompt, same coffee shop site. The results aren't identical. Opus produced a custom map, an animated coffee bean stamp, and dark mode out of the box. The budget models gave functional but plainer sites. But for most real-world work, "functional" is the bar. Model choice used to be about which one is smartest. Now it's about when the 200x premium is worth paying. For standard CRUD apps and boilerplate, the cheap tier already wins. Frontier models earn their cost on the hard 10% where reasoning genuinely matters. #AI #LLM #WebDev netlify.com/blog/one-prompt-…

10,000 markdown files and nowhere good to put them. Tolaria is a desktop app for managing markdown knowledge bases. The creator runs his entire life on it, literally 10,000+ notes from years of journaling and second-braining. People use it for company docs, AI context, storing assistant memory. Mac, Windows, Linux. Free. 19K stars. ⭐ 19.4K #Productivity #OpenSource github.com/refactoringhq/tol… Follow for daily dev finds 🔔

Stop re-explaining your codebase every time you open Claude Code. agentmemory gives coding agents persistent memory across sessions. Works with Cursor, Copilot CLI, Gemini CLI, the whole list. Built on Karpathy's LLM Wiki pattern but adds confidence scoring, knowledge graphs, and hybrid search. The design doc alone has 1.3K stars. 27K stars on the repo. ⭐ 27.0K #AI #DevTools github.com/rohitg00/agentmem… Follow for daily dev finds 🔔

Your AI agent writes your code but can't sit in your team chat. Buzz is a workspace where humans and agents share the same rooms. Self-hosted Nostr relay. Every message, code review, and git event is a signed event in one log. Agents can open repos, send patches, review PRs, even join voice huddles. Same access as any teammate. Different keypair. Built by Block. 27K stars already. ⭐ 27.2K #AI #OpenSource github.com/block/buzz Follow for daily dev finds 🔔

Your trading bot is a Python script with if/else statements. AI agents now have their own trading platform. AI-Trader lets any agent join by reading a SKILL.md and registering. Paper trading with live market data, mark-to-market scoring, leaderboards. Supports Claude Code, Codex, Cursor. Built by HKUDS, the same lab behind LightRAG. Agents competing against each other to sharpen trading skills. Wild concept. ⭐ 21.4K #AI #Trading github.com/HKUDS/AI-Trader Follow for daily dev finds 🔔

Setting up Claude Code for every new project means redoing all your configs from scratch. claude-code-templates handles that in one command. Pre-built templates, session monitoring, multi-project management. 30K devs use it daily to skip the setup tax. If you're in Claude Code more than once a week, this pays for itself in time saved. ⭐ 30.2K #ClaudeCode #DevTools github.com/davila7/claude-co… Follow for daily dev finds 🔔

Every AI agent framework eventually becomes a walled garden. DeepSeek Harness flips that. Everything is a plugin: the model provider, the tool registry, the UI, the session memory. Powered by Cordis, a composability framework that handles the plumbing. No core assumptions about what your agent should do or how. 34K stars in dev preview with breaking changes still incoming. People clearly want this architecture. ⭐ 34.7K #AI #OpenSource github.com/deepseek-ai/deeps… Follow for daily dev finds 🔔

Every AI safety framework assumes you can cut off the bad actor's API access. This worm doesn't have one. Researchers built a computer worm that runs open-weight LLMs on each machine it infects, using the stolen compute to reason about the next target and craft a custom attack. Unlike WannaCry, which exploited a single fixed vulnerability you could patch, this thing adapts. Every compromised box becomes both victim and weapon: its GPU feeds the worm's brain. They tested it across Linux, Windows, and IoT devices, hitting real corporate network vulnerabilities. The attacker's marginal cost per new infection is zero. No API keys to revoke. No rate limits. No kill switch. The guardrails OpenAI and Anthropic built into their APIs don't apply here. You can't refuse service to malware that never calls you. #AISecurity #CyberSecurity arxiv.org/abs/2606.03811

🤖 9 AI Agent Frameworks Worth Knowing Stop wiring solo chatbots. These 9 let agents talk to each other, split tasks, and finish the job while you do something else. 1. LangChain — chains, tools, memory, retrieval in one platform github.com/langchain-ai/lang… 2. CrewAI — role-playing agent teams in pure Python github.com/crewAIInc/crewAI 3. AutoGen — Microsoft's multi-agent system with a no-code studio github.com/microsoft/autogen 4. MetaGPT — assigns PM, engineer, and QA roles from one prompt github.com/FoundationAgents/… 5. Dify — visual drag-and-drop agent builder, deploy in minutes github.com/langgenius/dify 6. Mem0 — gives any LLM persistent memory across conversations github.com/mem0ai/mem0 7. TradingAgents — LLM agents that mirror a real trading firm's structure github.com/TauricResearch/Tr… 8. Ruflo — multi-agent swarms built specifically for Claude github.com/ruvnet/ruflo 9. Agency Agents — 61 specialized agent prompts, copy and paste github.com/msitarzewski/agen… Which one are you trying first? 👇

To a censor, your VPN traffic looks like Google Docs. That's what Zyrln does. Routes everything through Google's infrastructure using domain fronting. TLS connections land on Google's own IPs. The Host header pointing at your relay is encrypted, invisible to deep packet inspection. Way too few stars for how clever this is. ⭐ 754 #Privacy #AntiCensorship github.com/ajavadinezhad/zyr… Follow for daily dev finds 🔔

Your AI coding agent gets dumber the longer the session runs. That's not a bug. It's context rot. GSD Core offloads all the heavy work to fresh-context subagents so your main session stays clean. Research, planning, execution each run in isolation. Works with Claude Code, Codex, Copilot, Cursor. Five-phase loop, one step at a time. ⭐ 8.1K #AICoding #DevTools github.com/open-gsd/gsd-core Follow for daily dev finds 🔔

Code review bots either flag trivial style nits or hallucinate problems that don't exist. OpenCodeReview from Alibaba runs both. Deterministic rules catch NPE, SQL injection, XSS, thread-safety at the exact line. An LLM agent handles what's left. Tested on Alibaba's internal codebase. That's not a toy demo. ⭐ 20.4K #CodeReview #OpenSource github.com/alibaba/open-code… Follow for daily dev finds 🔔

This isn't supposed to come from code. gc-minimal-zine-poster takes a mood, a photo, or a single sentence and compiles sparse paper-texture editorial posters. 70-90% negative space. One color anchor. Risograph and halftone defects baked in. It's a Codex skill, not an app. The output looks like something from a Tokyo indie bookshop. ⭐ 5.7K #Design #CreativeCoding github.com/LiamGvchi/gc-mini… Follow for daily dev finds 🔔

26% of AI agent skills contain vulnerabilities. 5% are probably malicious. You're installing them blind. SkillSpector scans Claude Code, Codex, and MCP skills before you install them. 69 vulnerability patterns across 17 categories. Prompt injection detection, data exfiltration checks, supply chain analysis. NVIDIA open-sourced it. Should've existed a year ago. ⭐ 14.5K #AI #Security github.com/NVIDIA/SkillSpect… Follow for daily dev finds 🔔

You went to sleep. Your agents ran experiments, reviewed each other's work across models, and queued the next ideas. ARIS does autonomous ML research using pure markdown skills. No framework, no lock-in, works in Claude Code or Codex or whatever agent you already run. It hit #1 on HuggingFace daily papers. The fact that markdown files are doing all of this is wild. ⭐ 14.6K #AI #MLResearch github.com/wanshuiyin/Auto-c… Follow for daily dev finds 🔔