Super User Daily: 2026-04-02
The biggest story in the AI coding world this week is not a product launch or a benchmark. It is a catastrophic operational failure. Anthropic shipped a source map inside their npm package for Claude Code v2.1.88, and within hours the entire 512,000-line TypeScript codebase was downloaded, mirrored, rewritten in Python by Sigrid Jin using OpenAI Codex, and permanently distributed on decentralized platforms. The company that testified before Congress about AI existential risk could not manage a one-line .npmignore. But buried inside the chaos of memes, DMCA takedowns, and 50K-star repos, real users were doing real things with these tools. Some of them were building businesses. Some of them were breaking new ground in non-coding automation. Here is what actually mattered on March 31, 2026.
@supertommy [Claude Code]
https://x.com/supertommy/status/2039044393173135802
Built a fully playable Super Mario Galaxy fan game in the browser using Three.js. 731 commits over 53 days, with 95 percent of all code written by Claude Code Opus. The project includes a custom ECS engine, Rapier3D physics, galaxy-style gravity fields, a Bowser Jr boss fight, mobile touch controls, gamepad support, and a custom CLI level construction tool. The CLAUDE.md file grew to 163 lines of rules and constraints. 87 implementation plans were written before coding, 36 AI code reviews conducted, and 60 reusable skills extracted from debugging sessions. This is probably the most complex vibe-coded game ever shipped.
@develogon0 [Claude Code]
https://x.com/develogon0/status/2038942688545059157
Claims to be running 10 Instagram accounts and 5 TikTok accounts with zero manual intervention using Claude Code. The AI analyzes tens of thousands of viral videos to identify winning formats, then autonomously produces short-form video content following those patterns. Says the system runs while sleeping and has generated enough passive income that the side effect is gaining 5 kg from inactivity. A Japanese-language post with 545K impressions, suggesting strong traction in the Japanese creator economy.
@ai_jitan [Claude Code]
https://x.com/ai_jitan/status/2038894813077192743
Describes a case where Pokemon-named AI agents run a subsidy information media site completely autonomously using Claude Code, reaching 50,000 users in one month. The poster calls it the most exciting article they have read recently, saying it shattered their fixed assumptions about what AI automation can achieve. This is a non-coding, media business use case that shows Claude Code being used as an operations platform, not just a dev tool.
@om_patel5 [Claude Code]
https://x.com/om_patel5/status/2039122487473324248
A user vibe coded a Chrome extension at a cafe that scrambles his entire Gmail inbox so anyone looking over his shoulder sees gibberish. After two weeks, his brain adapted and he can now read the scrambled text without revealing anything. Built in a few prompts with Claude Code. No more privacy screen protectors, no more tilting the laptop, no more minimizing windows. A clever non-coding use case: personal security through AI-generated obfuscation.
@farzyness [Claude Code]
https://x.com/farzyness/status/2039081911788740872
CPA just finished 2025 taxes. Now feeding the same documents to Claude Code Opus 4.6 with 1M context to see if it arrives at the same numbers. A real head-to-head test of AI versus professional accounting. 536K impressions and counting. The result has not been posted yet, but the experiment itself signals a shift in how people think about professional services.
@spect3ral [Claude Code]
https://x.com/spect3ral/status/2039017377170665673
After studying the leaked Claude Code coordinator mode architecture, restructured an entire Meta ads workflow to use parallel agents. Agent 1 pulls campaign data. Agent 2 audits spend, CTR, and ROAS. Agent 3 generates copy variations from top performers. Agent 4 flags budget leaks and audience overlap. All running simultaneously. Full account audit plus creative refresh went from 5 hours to under 2, a 60 percent speed improvement. Non-coding power user case: advertising operations.
@SuguruKun_ai [Claude Code]
https://x.com/SuguruKun_ai/status/2038872366743502989
Compiled a list of Japan-specific MCP servers for Claude Code that are genuinely useful for professionals. Includes a law lookup server that queries the e-Gov API for any Japanese statute by article number, a tax law MCP covering 24 major tax laws plus 1,950 tribunal rulings with anti-hallucination design, a labor law MCP covering 45 labor-related statutes, a parliamentary record MCP for full-text search of Diet proceedings, a real estate and geospatial MCP from the Ministry of Land connecting to public land price and disaster data, and a financial analysis MCP covering structured financials for 3,800 listed companies. This is the MCP ecosystem maturing into domain-specific professional tooling.
@tokumahata [Claude Code]
https://x.com/tokumahata/status/2038812777880662209
An AI mentor told them to stop using Claude Code alone. The new stack: Claude for requirements and planning, Codex for code review, Cursor for implementation. The result is Claude never hits rate limits because it only plans and reviews, Codex provides tough criticism, and Cursor implements at maximum speed. Describes it as the strongest possible formation. A practical workflow that acknowledges no single tool is best at everything.
@mattiapomelli [Claude Code]
https://x.com/mattiapomelli/status/2038917621522092145
Added one skill to Claude Code and now designs and codes mobile apps entirely from the terminal. Vibe design plus vibe code in a single workflow. This is the kind of creative-to-code pipeline that collapses what used to be a multi-tool, multi-person process into a single session.
@towya_aillust [Claude Code]
https://x.com/towya_aillust/status/2039028748386009555
Set up Claude Code with Seedance 2.0 via the Renoise skill, asked it to create a kaiju movie trailer, and got a 60-second video proposal. The integration between an AI coding agent and a video generation model through a skill interface is a new kind of creative workflow. Posted the resulting video, showing the concept working end to end.
@LeoYe_AI [OpenClaw]
https://x.com/LeoYe_AI/status/2038787103522484669
Built Auto-Dream for OpenClaw agents, modeled on actual neuroscience. Every night at 4am, your agent enters a dream cycle: replays the day, extracts what matters, scores importance, links related memories, lets unimportant ones fade. Five memory layers from episodic to procedural, forgetting curves, knowledge graphs, stale detection. Open source under MIT license. 967K impressions. This is memory architecture that makes agents better at month 6 than month 1.
@OmarShahine [OpenClaw]
https://x.com/OmarShahine/status/2039023022699655297
Microsoft just appointed a CVP whose entire job is OpenClaw. Omar Shahine announced he is bringing OpenClaw plus personal agents to Microsoft 365. Already shipped a fully integrated Teams plugin for OpenClaw. The goal is proactive workplace assistants that take on tasks end-to-end. When Microsoft creates a VP-level position for an open-source agent framework, the enterprise agent era has officially arrived.
@aakashgupta [OpenClaw]
https://x.com/aakashgupta/status/2038910689176400150
OpenClaw triggered a full platform war in China within a single week. ByteDance shipped ArkClaw. Tencent shipped WorkBuddy. Alibaba opened QoderWork. Zhipu shipped AutoClaw with zero API keys required, 50 pre-built skills, and full browser automation. Zhipu stock jumped 13 percent on the announcement. Shenzhen is publishing government subsidies for OpenClaw developers with up to 2 million yuan per contributor. Four trillion-dollar companies racing to wrap a single open-source project with 310K GitHub stars.
@clairevo [OpenClaw]
https://x.com/clairevo/status/2039011562187341845
Spent at least 100 hours setting up, training, and working with OpenClaw. Read the docs, peeked into source code, edited config files by hand, walked friends through setup. Then wrote down everything in a comprehensive zero-to-mastery guide, endorsed by Lenny Rachitsky and multiple prominent tech voices. This is the kind of community-driven documentation that turns an open-source project into a platform.
@gkisokay [OpenClaw]
https://x.com/gkisokay/status/2038896951891210309
Built a Super Self-Improving Multi-Agent Framework using Hermes and OpenClaw. Agent 1 runs workflows. Agent 2 monitors the system, finds bugs, proposes fixes. Agent 3 (OpenClaw) audits and verifies. They go back and forth until they find a coded solution. Low-risk fixes auto-deploy. High-risk ones get raised for human approval. The second goal: once there are no bugs to fix, the agents focus on making meaningful improvements to the system over time. Self-improving infrastructure.
@irisneural [OpenClaw]
https://x.com/irisneural/status/2039042538972270902
Turned OpenClaw into a sales agent called ClawGTM. Paste your website URL and it understands your product, scans millions of job posts for buying signals, finds companies investing in your category, researches decision makers, and generates personalized email plus LinkedIn outreach. First outbound pipeline in 60 seconds. The key insight: companies leak intent in job posts. That is not noise, that is buying signal.
@ZvonimirSabljic [OpenClaw]
https://x.com/ZvonimirSabljic/status/2038988794956075471
Launched Pazi, an AI team that works 24/7, lives in Slack, and executes across your entire stack. Connects to Sentry, Linear, GitHub with one click. Powered by OpenClaw. Now in beta. An enterprise-facing wrapper around OpenClaw that packages the agent framework into a Slack-native ops team.
@idoubicc [Claude Code]
https://x.com/idoubicc/status/2039006326882546141
Used Claude Code to analyze the leaked source, extracted all the logic, and built open-agent-sdk as a drop-in replacement for claude-agent-sdk. Fully compatible interfaces, fully open source, function calls instead of CLI processes so no extra overhead for cloud-scale concurrent agent workloads. Describes it as: Claude Code house caught fire, I had Claude Code carry out all the furniture and build a new house for everyone to live in for free. MIT license.
@LexnLin [Claude Code]
https://x.com/LexnLin/status/2039069237885464593
Created Claw Dev, an open-source CLI agent with local model support via Ollama plus multiple providers including Anthropic, Gemini, Grok, and Groq. Inspired by Claude Code but not using Claude models for free. Full tutorial in README. Already fixing bugs via issues and PRs. The post-leak ecosystem is already producing functional alternatives.
@CardilloSamuel [Claude Code]
https://x.com/CardilloSamuel/status/2039059923729084472
Stripped the leaked Claude Code source of all telemetry and guardrails, then connected it to a local model (qwen3-coder-next) using LiteLLM to proxy OpenAI API to Anthropic API format. Reports it works amazingly well. This is the self-hosted Claude Code dream realized within hours of the leak.
@hqmank [Claude Code]
https://x.com/hqmank/status/2038970622462841019
Compiled the leaked Claude Code source and got it running locally. Then spun up an Agent Team and had it analyze its own codebase. Claude Code analyzing Claude Code, running on a local machine. The recursive irony is real, but so is the technical achievement of getting the full stack operational outside Anthropic infrastructure.
@hasantoxr [Claude Code]
https://x.com/hasantoxr/status/2038976967702446554
Launched OpenAgents Workspace. One command, no login, no Docker, no config. Claude Code, OpenClaw, Codex CLI, and 9 more agents all running in a single collaborative thread. Fully open source. The first real multi-agent workspace where different vendor agents can communicate in real time.
@RihardJarc [Claude Code]
https://x.com/RihardJarc/status/2038944176914432344
Wrote a detailed investment thesis on memory demand based on the leaked Claude Code source. Key findings: idle Claude Code processes grow to 15GB each, active sessions hit 93-129GB. The DAEMON and PROACTIVE feature flags point toward always-on background agents, meaning baseline memory consumption of 15-30GB just for Claude Code before opening anything else. Auto Dream consolidation uses significant inference tokens just to manage memory. Moving from 200K to 1M context is a 5x increase in KV cache memory per session. Memory demand per active user is increasing faster than user count.
@iamfakeguru [Claude Code]
https://x.com/iamfakeguru/status/2038965567269249484
Reverse-engineered the leaked source against billions of tokens of personal agent logs. Found that Anthropic has fixes for Claude Code hallucination and laziness that are gated to employees only via USER_TYPE equals ant. Published a detailed CLAUDE.md configuration that replicates the employee-grade verification loop. Specific findings: file writes mark success even if code does not compile, context compaction fires at 167K tokens and destroys reasoning chains, file reads are capped at 2000 lines with silent truncation. The post includes a full drop-in CLAUDE.md with mechanical overrides. 1.4M impressions.
@DeryaTR_ [Claude Code]
https://x.com/DeryaTR_/status/2038925279075168590
Reports that after 15 minutes of asking Claude Code Opus 4.6 to continue where it left off, 90 percent of the limit was already consumed. A few minutes later, at 100 percent. Zero tasks completed. States the token-eater bug is clearly still not fixed. Multiple users corroborated this experience throughout the day.
🗣 User Voice
User Voice
Token consumption is the number one pain point. Multiple users including @DeryaTR_, @op7418, @meta_alchemist, and @WEB3_furture reported that Claude Code burns through rate limits at 5-20x normal rates, with the resume flag and Bun binary both identified as likely causes. Users are desperate for transparency on what is being consumed and why.
The employee-only feature gate exposed by @iamfakeguru generated massive frustration. Users paying for Max subscriptions discovered that known fixes for hallucination and false success claims exist in the codebase but are restricted to internal Anthropic accounts. The sentiment is clear: if you built the fix, ship the fix.
Local and self-hosted demand is surging. Posts from @ziwenxu_, @CardilloSamuel, @0xKingsKuan, and @billtheinvestor all point to the same desire: run Claude Code quality tooling without API costs, rate limits, or data leaving the machine. Ollama integration with MLX on Apple Silicon, announced by @ollama, is accelerating this trend.
OpenClaw reliability is a recurring concern. @BentoBoiNFT explicitly switched from OpenClaw to Claude Code because it keeps breaking. @kwok_mavis asked whether OpenClaw is safe and if anyone is worried about credentials. The axios supply chain attack affecting OpenClaw packages, flagged by @evilcos and @feross, adds urgency to security hardening.
Non-coding use cases are underserved but growing fast. Media automation (@develogon0, @ai_jitan), advertising operations (@spect3ral), tax preparation (@farzyness), and sales prospecting (@irisneural) all demonstrate demand for agent tooling far beyond software development, but documentation and tutorials still assume a developer audience.
Token consumption is the number one pain point. Multiple users including @DeryaTR_, @op7418, @meta_alchemist, and @WEB3_furture reported that Claude Code burns through rate limits at 5-20x normal rates, with the resume flag and Bun binary both identified as likely causes. Users are desperate for transparency on what is being consumed and why.
The employee-only feature gate exposed by @iamfakeguru generated massive frustration. Users paying for Max subscriptions discovered that known fixes for hallucination and false success claims exist in the codebase but are restricted to internal Anthropic accounts. The sentiment is clear: if you built the fix, ship the fix.
Local and self-hosted demand is surging. Posts from @ziwenxu_, @CardilloSamuel, @0xKingsKuan, and @billtheinvestor all point to the same desire: run Claude Code quality tooling without API costs, rate limits, or data leaving the machine. Ollama integration with MLX on Apple Silicon, announced by @ollama, is accelerating this trend.
OpenClaw reliability is a recurring concern. @BentoBoiNFT explicitly switched from OpenClaw to Claude Code because it keeps breaking. @kwok_mavis asked whether OpenClaw is safe and if anyone is worried about credentials. The axios supply chain attack affecting OpenClaw packages, flagged by @evilcos and @feross, adds urgency to security hardening.
Non-coding use cases are underserved but growing fast. Media automation (@develogon0, @ai_jitan), advertising operations (@spect3ral), tax preparation (@farzyness), and sales prospecting (@irisneural) all demonstrate demand for agent tooling far beyond software development, but documentation and tutorials still assume a developer audience.
📡 Eco Products Radar
Eco Products Radar
claw-code / clawd-code by @realsigridjin: Python clean-room rewrite of Claude Code, 51K+ stars, Rust port in progress. The fastest-growing repo in GitHub history.
OpenCode by @gitlawb: Forked leaked Claude Code to work with any LLM (GPT, DeepSeek, Gemini, Llama, MiniMax). Open source.
oh-my-claudecode: 32 specialized agents, 5 execution modes, 31 lifecycle hooks for Claude Code. Trended number 1 on GitHub.
Paperclip: One-click setup of zero-human companies on OpenClaw. Multiple mentions as game-changing for the ecosystem.
Pazi by @pazidotai: AI team in Slack powered by OpenClaw. Sentry, Linear, GitHub integration.
Firecrawl: Now a default search provider in OpenClaw, enabling full web search results from any page.
Auto-Dream by @LeoYe_AI: Neuroscience-modeled memory consolidation for OpenClaw agents. MIT license.
Perceptron MCP: 2B parameter vision model for Claude Code agents, replacing expensive frontier model vision calls. Open source, one-line install.
everything-claude-code: 12K+ stars. Full Claude Code configuration including 136 skills, 30+ sub-agents, 60+ slash commands. Won Anthropic hackathon.
Pretext by Cheng Lou: Pure TypeScript text measurement engine, 500x faster than DOM, validated with Claude Code and Codex. 15KB.
claw-code / clawd-code by @realsigridjin: Python clean-room rewrite of Claude Code, 51K+ stars, Rust port in progress. The fastest-growing repo in GitHub history.
OpenCode by @gitlawb: Forked leaked Claude Code to work with any LLM (GPT, DeepSeek, Gemini, Llama, MiniMax). Open source.
oh-my-claudecode: 32 specialized agents, 5 execution modes, 31 lifecycle hooks for Claude Code. Trended number 1 on GitHub.
Paperclip: One-click setup of zero-human companies on OpenClaw. Multiple mentions as game-changing for the ecosystem.
Pazi by @pazidotai: AI team in Slack powered by OpenClaw. Sentry, Linear, GitHub integration.
Firecrawl: Now a default search provider in OpenClaw, enabling full web search results from any page.
Auto-Dream by @LeoYe_AI: Neuroscience-modeled memory consolidation for OpenClaw agents. MIT license.
Perceptron MCP: 2B parameter vision model for Claude Code agents, replacing expensive frontier model vision calls. Open source, one-line install.
everything-claude-code: 12K+ stars. Full Claude Code configuration including 136 skills, 30+ sub-agents, 60+ slash commands. Won Anthropic hackathon.
Pretext by Cheng Lou: Pure TypeScript text measurement engine, 500x faster than DOM, validated with Claude Code and Codex. 15KB.
Comments