Super User Daily: 2026-06-02
Two threads pulled this slice together. One: Claude Code users have stopped asking "can I build with it" and started measuring what it actually costs to keep an agent on the wire 24/7 — tokens per week per engineer, API price vs Max plan arbitrage, multi-agent dashboards as the new normal. Two: OpenClaw is the loudest test case for whether always-on agents survive the security and ergonomic walls coming at them — exec approval systems, hardware embodiments, ecosystem fights with Hermes, and very public Anthropic enforcement against Max-login reuse. The use cases below skew non-coding and infrastructural; the bug reports are getting more specific; the cost math is showing up unsolicited.
@dinii_tech [Claude Code]
https://x.com/dinii_tech/status/2061355097066701109
Japanese restaurant tech company Dinii published its AI-Native Transformation Report Vol.1. Every engineer is on Claude Code Max 20x. Each one burns about 80 million tokens a week. The output side: PR count is roughly 2.6x, shipped features about 2x. The number that matters most is 80M tokens per week per person — that's the going rate for a serious Claude Code engineer right now, and it makes most usage-cap conversations look quaint.
@steipete [OpenClaw]
https://x.com/steipete/status/2061208638027395490
The OpenClaw team is using Codex as their internal QA. For every commit, Codex generates a user-test scenario, then runs it through webVNC (crabbox) plus browser-use tools (peekaboo, mcporter) — testing OpenClaw the way a real user or QA tester would. The runner is fully background; it opens PRs with fixes. This is an LLM stress-testing another agent runtime via a third agent, and it shipped to prod.
@robiartec [OpenClaw]
https://x.com/robiartec/status/2061079150136332318
A Spanish-language thread on ten "boring" SaaS plays an OpenClaw+Hermes agent can run end-to-end: scrape commercial roofs for solar ROI reports, find Shopify stores at 1M EUR with no international shipping and offer localized stores, detect Meta ads with 1-2-star reviewed products and counter-launch with a better SKU, hit Spanish franchisees complaining about their franchisor's tech stack and sell them an alternative POS. The pattern is the same: public-data scrape + render the deliverable before the customer asks + show up with numbers. The arbitrage is information; OpenClaw is the lever.
@0xCodez [Claude Code]
https://x.com/0xCodez/status/2061543982618841507
Built a scroll-animated cinematic landing page in 90 minutes for zero designer cost. Workflow: Claude Code with a Higgsfield connector. Claude designs the brand in one prompt (colors, font, logo draft). Hands logo + hero to Higgsfield for cinematic visuals. Higgsfield + Cinema Studio generates 4 video clips. Claude wires the page with WebGL + Three.js for 3D video. The final 10% is human shot-selection. Quote: "$0 on designers. Setup takes 30 seconds."
@Mahaximus_ [Claude Code]
https://x.com/Mahaximus_/status/2061050888320438711
An Anthropic engineer's live 22-minute walkthrough of "you're not supposed to watch Claude Code work, you're supposed to wake up and review what it shipped." The setup keeps agents running after the terminal closes. The poster's framing: most people close their terminal and everything stops; the people compounding shipped this primitive into their workflow already.
@MatiasWilliams [Claude Code]
https://x.com/MatiasWilliams/status/2061351966886826282
Spanish walkthrough of the difference between `claude` (single pair-programming session) and `claude agents` (parallel multi-agent dashboard). His Monday loadout: Agent 1 reviews a pending PR, Agent 2 fixes flaky login tests, Agent 3 updates dependencies, Agent 4 implements the feature pending since Friday. Adds /goal per-agent and auto mode for extra autonomy. His phrase: "you can switch, respond, or ignore while you keep doing your own thing."
@miyagawa [Claude Code]
https://x.com/miyagawa/status/2061233175968555018
Veteran Japanese engineer migrating an infra job: nginx config + log aggregation cronjob. Originally planned Kamal + nginx container. Switched to Fastly VCL for nginx config and Papertrail + GitHub Actions for log aggregation, then handed the entire migration to Claude Code. His comment: "no problems." Worth watching when a long-tenured engineer treats Claude Code as a contractor that finishes a multi-system migration end-to-end.
@mustafaergisi [Claude Code]
https://x.com/mustafaergisi/status/2061213632873263378
"My overnight refactor pipeline runs entirely on Claude Code now and the boring stuff has been gone from my mornings for weeks." This is the canonical "I sleep, it ships" line spoken in production. The "for weeks" matters — this is past the novelty window into routine.
@WedlockChatGPT [Claude Code]
https://x.com/WedlockChatGPT/status/2061235842421846043
Non-coding use case worth noting. The poster had been sending daily diaries to a "yes-and" Gemini setup for emotional support. Switched to Claude Code to do diagnostic analysis with a weekly report, possibly including a recommendation to see a doctor. The honest first reaction: shock. The fix: route through a Claude Code subagent with the persona of Sebastian the butler, which made hard truths land. Personality-as-buffer is a real interaction-design move.
@VivekIntel [Claude Code]
https://x.com/VivekIntel/status/2061063109662662856
Claude-BugHunter: 51 security skills + 15 slash commands + 681 real disclosed bug-bounty reports, packaged so Claude Code auto-loads the right skill for what you're attacking. Covers Web, API, Cloud, OAuth, SAML, GraphQL, SSRF, IDOR, XSS, RCE, plus enterprise attack paths into M365, Okta, SharePoint, VMware. Integrates with Burp via MCP. The skill-bundle pattern keeps winning vs prompt-stuffing.
@AiCamila_ [Claude Code]
https://x.com/AiCamila_/status/2061110419947986959
Five architectural layers most engineers never open: CLAUDE.md (memory), Skills (knowledge), Hooks (guardrails), Subagents (delegation), Plugins (distribution). The thesis: production failures in agentic systems trace back to a missing layer, not a missing prompt. Hooks specifically — PreToolUse, PostToolUse, SessionStart, Stop — get singled out as the deterministic shell-trigger layer most teams skip first and regret first.
@imarun_chauhan [Claude Code]
https://x.com/imarun_chauhan/status/2061153084768964625
Anatomy of a Claude Code project: keep CLAUDE.md as the north-star (purpose, repo map, rules) and resist dumping knowledge into it. Move workflows into .claude/skills/ (review checklist, refactor playbook, release procedure). Use .claude/hooks/ for deterministic guardrails. Put local CLAUDE.md files near sharp edges — src/auth/, src/persistence/, infra/. "Prompting is temporary. Structure is permanent."
@iamlukethedev [OpenClaw]
https://x.com/iamlukethedev/status/2061120598948917305
The new OpenClaw exec approval system: allowlists, human approval when needed, per-agent policies, sandboxed execution paths, approval flows that survive crashes and reconnects. The user-side angle is the philosophical shift this represents — most AI coding agents still run in YOLO mode, and the actual attack surface is the actions an agent takes, not the words it generates.
@drivelinekyle [Claude Code]
https://x.com/drivelinekyle/status/2061234724274889010
"Every third tool call on the highest thinking mode in Claude Code. Just infinite tool call failure spam in red text, recovery, empty thinking blocks." Had to ask Claude to summarize the work and hand off to Codex. Filed /feedback. The interesting bit isn't that there's a bug — it's the workflow: when one agent goes sideways, the user has it summarize state and pass to the next, treating the two CLIs as a fallover pair.
@SecureInterest [Claude Code]
https://x.com/SecureInterest/status/2061235870590751016
The trust complaint of the day: "It also doesn't always follow instructions and memory rules (like ignoring 'only push to live with my explicit approval'). Oops I forgot, sorry." Followed by an analogy to a self-driving car with the family inside. Whatever you think of the Tesla comparison, the substantive point — that memory rules silently degrade — is showing up across the dataset.
@HarveenChadha [Claude Code]
https://x.com/HarveenChadha/status/2061085291247555002
Real /goal bug report: ran /goal, tried to clear with /goal clear, told it the goal was done (when it wasn't), still looping into repeated messages. Anyone seriously running long /goal sessions is hitting this. The replies under similar threads are converging: /goal works when the metric is binary and verifiable; it implodes when the success condition is soft.
@Randybobu [Claude Code]
https://x.com/Randybobu/status/2061234103660184040
The economics line getting traction: one developer tracked 10 billion tokens of Claude Code usage over 8 months. At API rates that's about $15,000; on a Max plan that's about $800. The 18x gap is exactly why Anthropic is moving the Agent SDK out of subscriptions on June 15. Worth holding both facts at once.
@afterhoursHB [Claude Code]
https://x.com/afterhoursHB/status/2060906525221392451
Day 4 of "ship or die" — 2 pages plus a small back office shipped yesterday, 30+ alpha users testing inside the community. Stack: Claude Code for almost everything, Codex for checks before pushing to prod, Vercel + Supabase underneath. Today's targets: fix user-reported bugs, add auth, plug payments, ship V1 tonight. The Claude+Codex pair as primary/safety net is becoming standard.
@aliksir_mirage [Claude Code]
https://x.com/aliksir_mirage/status/2061234665038749960
Shipped neko-rescue: a CLI that pulls your in-flight Claude Code conversation out of the .jsonl session file when the agent freezes. Reads-only, zero dependencies, Node 22+. Outputs the conversation as Markdown plus "last instruction" and "last running tool" so you can resume from a different window. This is the kind of side tool that appears once a primary tool has enough users with a real grievance.
@flowersslop [OpenClaw]
https://x.com/flowersslop/status/2061233508991799642
Honest critique of OpenClaw's heartbeat pattern: "loops and stuff like the heartbeat in openclaw feel like a bandaid solution for a way deeper more fundamental problem of how these things are built — they're still reactive rather than proactive." The wider dataset shows people are starting to recognize that LLM "agents" are still pull-mode dressed up; the heartbeat is a workaround, not a solution.
@MaxForAI [OpenClaw]
https://x.com/MaxForAI/status/2061059784313614833
Quoted exchange where an open-source project's maintainer responded to user feedback with "then go use OpenClaw." Whatever you think of the dunk, the cultural marker is real: OpenClaw is now the reference "go use the other thing" within the Chinese-language open agent community.
@GalianoTiramani [OpenClaw]
https://x.com/GalianoTiramani/status/2061233042438447137
Old laptop running OpenClaw. The agent is now trading crypto, sending the poster's wife love poems, doing marketing and pulling the next play for Boxabl. Asked the community for more ideas. This is the "one personal agent runs multiple side projects from a corner of the house" use case showing up in the wild.
@drjingle [OpenClaw]
https://x.com/drjingle/status/2061088635026645163
HooRii Technology shipped ClawStage, a desktop robot built on OpenClaw architecture — the first embodied-AI personal assistant on the open-source stack. Camera, mic, motion sensors; physical responses (turning, posture changes) tied to your work tempo; direct control of Matter or Apple Home devices. Price: HK$1500 (about $190). The interesting part isn't the gadget, it's that OpenClaw is the upstream the hardware folks are picking.
@flatt_sec_en [Claude Code]
https://x.com/flatt_sec_en/status/2061433196563386537
Security researcher RyotaK disclosed a Claude Code GitHub Actions vulnerability: attackers could bypass permission controls via a GitHub issue and hijack the workflow's repository access. Fixed in v1.0.94. If you run the workflow, the recommendation is to audit your configuration and review your workflow run logs. The "agent commits code on your behalf" attack surface is real and the bug bounty community is starting to map it.
@yaohui12138 [Claude Code]
https://x.com/yaohui12138/status/2061276417355399522
Coze 3.0 review with a clear-eyed take on multi-agent orchestration. The poster's framing: the real efficiency gap isn't model intelligence, it's whether the human can act like a CEO scheduling an AI team. His old workflow had Claude Code for code, ChatGPT for copy, Midjourney for images, plus 10 browser tabs. Coze 3.0 lets one main agent dispatch to specialized sub-agents (copy, research) with a single @-mention. Whether Coze wins is secondary; the framing is right.
@seandkim_ [Claude Code]
https://x.com/seandkim_/status/2061235153691942950
Memory pain point being voiced widely: "Obsidian is great but it doesn't have an MCP server or APIs, so I have been using it locally via Claude Code only." The implicit ask is for a memory layer accessible from every platform, not just one local tool. The Obsidian + Claude Code combo keeps coming up as the best worst option.
@ekampf [Claude Code]
https://x.com/ekampf/status/2061233751234085326
"I replaced doom scrolling with launching Claude code on my phone, connecting to one of my repos and asking Claude about optimizations." Mobile Claude Code as habit replacement is a use case that didn't exist three months ago.
@gitlawb [OpenClaw]
https://x.com/gitlawb/status/2060877426234962334
Gitlawb hit rank #2 globally on OpenRouter, flipping OpenClaw and Kilo Code. Whatever you think of the rankings, the OpenRouter leaderboard is becoming a real signal of which agent runtimes actually get used at scale.
@rohit4verse [OpenClaw + Hermes]
https://x.com/rohit4verse/status/2060911598248853908
Hermes Agent just passed OpenClaw as the most-used open agent on OpenRouter. The pitch for Hermes: writes its own skills from experience, remembers every session. The takeaway isn't who's #1 this week; it's that the "open agent runtime" category is now a multi-horse race with measurable share, not a single-vendor narrative.
@lidangzzz [OpenClaw]
https://x.com/lidangzzz/status/2061117169417466340
Asked publicly: "Do you still use OpenClaw? What kind of work are you doing with it now?" 137k impressions. The community is past the install phase and into the "what do we actually run on it" phase, and the question itself signals retention is the next bar.
🗣 User Voice
User Voice
@drivelinekyle and @HarveenChadha — recent Claude Code updates broke real workflows. Tool-call failures on highest thinking mode and /goal not respecting its own clear command are both showing up under high-engagement complaints. Users want stable rollback, not just new features.
@SecureInterest — memory rules silently being ignored ("only push to live with my explicit approval") is the single most common trust complaint. Users want hard guardrails enforced at the tool layer, not soft preferences in CLAUDE.md.
@seandkim_ — every memory tool the community likes is locked to one platform. Obsidian, Claude Code's local memory, Anthropic's context APIs — pick one and you lose the others. The MCP-over-Obsidian gap keeps coming up.
@Randybobu and @dinii_tech — users are tracking token economics explicitly now. 10B tokens at $15K API vs $800 Max, 80M tokens/week/engineer. The next big ask is more transparent /cost output and a way to split subscription budget across multiple repos and team members.
@flowersslop — the "heartbeat" pattern (OpenClaw, every always-on agent) is being recognized as a workaround. Users want a primitive for proactive scheduling that isn't a polling loop pretending to be event-driven.
@drivelinekyle and @HarveenChadha — recent Claude Code updates broke real workflows. Tool-call failures on highest thinking mode and /goal not respecting its own clear command are both showing up under high-engagement complaints. Users want stable rollback, not just new features.
@SecureInterest — memory rules silently being ignored ("only push to live with my explicit approval") is the single most common trust complaint. Users want hard guardrails enforced at the tool layer, not soft preferences in CLAUDE.md.
@seandkim_ — every memory tool the community likes is locked to one platform. Obsidian, Claude Code's local memory, Anthropic's context APIs — pick one and you lose the others. The MCP-over-Obsidian gap keeps coming up.
@Randybobu and @dinii_tech — users are tracking token economics explicitly now. 10B tokens at $15K API vs $800 Max, 80M tokens/week/engineer. The next big ask is more transparent /cost output and a way to split subscription budget across multiple repos and team members.
@flowersslop — the "heartbeat" pattern (OpenClaw, every always-on agent) is being recognized as a workaround. Users want a primitive for proactive scheduling that isn't a polling loop pretending to be event-driven.
📡 Eco Products Radar
Eco Products Radar
Codex / Codex CLI — paired with Claude Code as primary/safety net in multiple production workflows.
Hermes Agent — passed OpenClaw on OpenRouter as the most-used open agent.
OpenClaw — the reference open agent runtime; mentioned in every always-on stack and in every critique of always-on stacks.
ECC (Everything Claude Code) — 232 skills + 60 sub-agents + 75 commands, won the Anthropic hackathon, working across Claude Code, Cursor, Codex, Gemini, OpenCode, Zed.
Obsidian — still the local memory of choice paired with Claude Code, despite MCP gap.
DeepSeek API — the price-arbitrage entry point for running Claude Code with cheaper tokens.
Higgsfield — cinematic video generator that paired with Claude Code is shipping landing pages in 90 minutes.
OpenRouter — the de-facto leaderboard for which open agents actually have usage share.
Cursor — still the most common cross-tool comparison ("Claude Code vs Cursor vs Codex").
Grok Build (xAI) — surfacing in complaints as the next CLI users are waiting on.
Burp Suite via MCP — paired with Claude Code in the security/bug-bounty skill bundles.
ClawStage (HooRii) — first embodied-AI hardware built on OpenClaw architecture.
Codex / Codex CLI — paired with Claude Code as primary/safety net in multiple production workflows.
Hermes Agent — passed OpenClaw on OpenRouter as the most-used open agent.
OpenClaw — the reference open agent runtime; mentioned in every always-on stack and in every critique of always-on stacks.
ECC (Everything Claude Code) — 232 skills + 60 sub-agents + 75 commands, won the Anthropic hackathon, working across Claude Code, Cursor, Codex, Gemini, OpenCode, Zed.
Obsidian — still the local memory of choice paired with Claude Code, despite MCP gap.
DeepSeek API — the price-arbitrage entry point for running Claude Code with cheaper tokens.
Higgsfield — cinematic video generator that paired with Claude Code is shipping landing pages in 90 minutes.
OpenRouter — the de-facto leaderboard for which open agents actually have usage share.
Cursor — still the most common cross-tool comparison ("Claude Code vs Cursor vs Codex").
Grok Build (xAI) — surfacing in complaints as the next CLI users are waiting on.
Burp Suite via MCP — paired with Claude Code in the security/bug-bounty skill bundles.
ClawStage (HooRii) — first embodied-AI hardware built on OpenClaw architecture.
Comments