Super User Daily: 2026-05-24
If there was one word humming under every Claude Code and OpenClaw post yesterday, it was money. Not capability, not benchmarks, money. The loudest news was Microsoft quietly canceling most of its internal Claude Code licenses because the token bill blew past budget, and the loudest grassroots reaction was a wave of people wiring Claude Code's interface to cheaper backends so they can keep the workflow and ditch the invoice. Underneath that, two quieter but more interesting currents: people who have never touched a terminal are now shipping software, and a handful of builders are turning these agents into always-on, ambient things that watch, buy, and write on their own. Here are the real ones.
@rileybrown [Claude Code]
https://x.com/rileybrown/status/2057669775837577538
Riley's pitch is blunt: the App Store is dead, because you can now wire Claude Code or Codex into iMessage and have it build a Swift app straight onto your home screen. The part that makes it more than a stunt is the framing, not one app but an entire personal app store, eight apps from one prompt. This is the clearest expression yet of where the "everyone builds their own software" trend is heading: not downloading apps, but texting an agent to make them on the spot. Whether the apps are any good is a separate question, but the distribution model is genuinely new.
@MacopeninSUTABA [Claude Code]
https://x.com/MacopeninSUTABA/status/2057664999921754298
This is one of the more useful non-coding workflows of the day: driving an already-logged-in Chrome session directly from Claude Code, with the login state preserved. He shares a step-by-step recipe meant to be usable on the spot at work. The reason this matters is that most browser-automation setups die the moment they hit a login wall. Keeping the human's authenticated session and letting the agent operate inside it sidesteps the single biggest blocker to real browser tasks.
@aiedge_ [OpenClaw]
https://x.com/aiedge_/status/2057759715371925911
The wildest OpenClaw use case shared yesterday: someone wired OpenClaw directly into their Meta AI glasses, turning it into a 24/7 live assistant. He literally looks at a product and tells OpenClaw to buy it. Strip away the novelty and this is the actual thesis of the always-on personal agent, the gap between intent and action collapsing to a glance and a sentence. It's a demo, not a daily driver yet, but it's the most concrete picture of the ambient-agent future anyone posted.
@MichaelGannotti [OpenClaw]
https://x.com/MichaelGannotti/status/2057884183465271508
Michael set up an OpenClaw instance named "Jeff" running on Windows 11 and wired into Microsoft 365 and Azure, and Jeff now writes his own blog. The tagline is the tell: "written by an AI that actually runs inside the stack it covers." This is one of the few examples of an agent given a standing, open-ended job (publish analysis about Microsoft AI) rather than a one-off task. The output quality is unproven, but the setup, an agent embedded in a real enterprise stack with a persistent mandate, is exactly the shape people keep describing in theory.
@Saboo_Shubham_ [OpenClaw]
https://x.com/Saboo_Shubham_/status/2057681030182814031
Saboo lays out a two-layer agent architecture that a lot of people responded to: Hermes as the orchestrator, with a squad of OpenClaw agents underneath as the workers, all sharing a single brain. Each OpenClaw agent is an individual Telegram bot he can talk to directly, while Hermes calls the plays. This is the most-discussed concrete multi-agent setup of the day, and the "orchestrator calls plays, OpenClaw runs them" division of labor is becoming a default pattern. It's worth watching whether the shared-memory part actually holds up, since that's where most multi-agent setups quietly break.
@karriodev [Claude Code]
https://x.com/karriodev/status/2057889187903250730
Here's the cost rebellion in its purest technical form. Karrio points Claude Code at DeepSeek's Anthropic-compatible endpoint, so the flow becomes Claude Code to DeepSeek V4 instead of to Anthropic. You keep the entire Claude Code experience, the UX, the workflows, the tools, the agent loops, the terminal, but you're running a cheaper model and specialized workers underneath. This is the move that explains the whole day: the interface has become the product, and the model behind it is now a swappable, price-shopped commodity.
@hellobyedot [Claude Code]
https://x.com/hellobyedot/status/2057903158509162750
A very honest look at how a heavy user actually rations across tools. He audits with Claude Code and implements with Codex, but Claude Code hits its limit after every two or three audits, so he switches to Codex until Claude resets, then comes back. His punchline lands hard: "I'm not sure why I even bought Claude Code." This is the lived reality behind the macro cost story, real workflows are now shaped around rate limits, not capability.
@MertLovesAI [Claude Code]
https://x.com/MertLovesAI/status/2057857972697448672
A genuinely useful experimental result buried in a reply: he swapped grep for vector retrieval across four agent CLIs (Chronos, Claude Code, Codex CLI, Gemini CLI) and found the harness change moved accuracy 14 points, while tuning the retriever itself barely registered. The lesson he draws, that skills are now fungible across agents and the harness matters more than the component, is one of the more transferable insights of the day. If it holds, it means effort is better spent on the scaffolding than on swapping individual pieces.
@kilo_cpa [Claude Code]
https://x.com/kilo_cpa/status/2057777242806100335
A clean little production architecture for a voice agent: Vapi handles the phone call, Claude Code writes the logic, and an "ask_client" tool keeps the human in the loop for every decision that matters, all at about $0.20 a minute. What makes this worth noting is the human-in-the-loop tool as a first-class part of the design rather than an afterthought. It's the difference between a demo that does whatever and a system you'd actually let touch a client call.
@mustafaergisi [Claude Code]
https://x.com/mustafaergisi/status/2057907586146468165
The human story of the day. He walked his dad through Claude Code on a Windows laptop and watched him build a folder-cleanup script in 20 minutes, ending with "the 'I get it now' look that does not show up in benchmarks." This is the adoption curve nobody charts: not developers switching tools, but a parent automating a chore for the first time. The whole non-coder-onboarding theme that ran through the day lives in this one anecdote.
@lukgart [Claude Code]
https://x.com/lukgart/status/2057831391132614808
Lukgart describes Claude Code not as autocomplete but as his main building environment, full sessions of plan, build, review, ship. It runs his projects, writes its own skills, manages his note vault, and handles about 80 percent of what he'd otherwise type by hand. This is the "programmable dev environment" view of Claude Code in practice, the leverage isn't the code generation, it's letting the agent own the whole loop including its own tooling.
@GroundFloor_hq [Claude Code]
https://x.com/GroundFloor_hq/status/2057949958297317447
One hard number worth pinning down: a six-person team running Claude Code spends about $2,400 a month, roughly 240 million tokens. Scale that to thousands of engineers and you understand instantly why Microsoft blinked. This is the unglamorous arithmetic that the whole cost-panic discourse is built on, and it's rare to see someone post the actual per-team figure rather than just complaining.
@designtako [Claude Code]
https://x.com/designtako/status/2057778626746007818
A designer's take on the new /usage command: it shows which MCP, skill, or agent is actually eating your tokens. For someone running a Figma MCP plus a design-review skill, this turns token spend from a guessing game into something you can tune. It's a small feature, but it's the exact instrument the cost-anxious crowd has been asking for, and it's notable that a designer, not an engineer, is the one pointing at it.
@_itsjustshubh [Claude Code]
https://x.com/_itsjustshubh/status/2057948183607529650
A concrete picture of side-stacking: he's building Claude Code plugins for AMC theaters, Luma events, and calendar automation while holding down a FAANG full-time job. The detail that makes it real is the specific integrations, these are everyday-life automations, not toy demos. It's a good example of the plugin directory shipping actually lowering the bar to "I'll just build the integration I need this weekend."
@pederaaby [Claude Code]
https://x.com/pederaaby/status/2057722244273434870
A snapshot of a multi-agent home office: one agent cleaning the computer, Claude doing accounting, Codex and Claude Code building an app, and a tool called Blume keeping them all running smoothly. It's light on detail but vivid on direction, the household as a small fleet of specialized agents with an orchestrator babysitting them. This is what "everyone runs a swarm" looks like before it's a product.
@Sadok_Ayari [OpenClaw]
https://x.com/Sadok_Ayari/status/2057741899608600772
A pointed migration story that doubles as an OpenClaw critique: he fought his OpenClaw setup, connectors and skills, for three weeks, then got the same thing working in Hermes in three hours. Whether or not you take sides, this is the recurring OpenClaw complaint of the day, setup friction and post-update breakage, stated with specific numbers. It's the clearest data point on why some power users are churning off it.
@BrandonHorner4 [OpenClaw]
https://x.com/BrandonHorner4/status/2057928187393503546
A useful cost tip that mirrors the Claude Code routing trend: you can hook any LLM to OpenClaw, and instead of paying for raw API tokens you can route through your existing ChatGPT, Codex, or Grok subscription allotment. His point is simply that this is much cheaper than direct API calls. It's the same arbitrage as the DeepSeek routing, use the flat-rate subscription you already pay for as the engine, rather than metered API billing.
@notjusta_selim [Claude Code]
https://x.com/notjusta_selim/status/2057770027433013468
A small but complete ship: he built a working app in three days using only Claude Code, and is handing out the prompts to anyone who replies. No grand claims, just a finished thing and a willingness to show the recipe. These quiet "I made the whole thing with one tool in a weekend" posts are the steady baseline under all the louder cost drama.
@ketandesign [Claude Code]
https://x.com/ketandesign/status/2057927781594620056
A designer vibe-coded a custom-designed site with Claude Code and reports the Lighthouse scores already came back solid, with a plan to push them all to 100. It's a modest case, but it's a real one: a non-engineer using Claude Code for front-end work and measuring the result with an objective performance metric rather than vibes. The willingness to check Lighthouse is what separates this from a screenshot.
@mihaieremia [Claude Code]
https://x.com/mihaieremia/status/2057947237733572795
A clean remote-control pattern: instead of using Termius to SSH in, connect the Claude iOS app to a Hetzner server that's running Claude Code, start one long-running session per project inside a screen, and you have full access to all of them from your phone anytime. It's a practical answer to the "I want my agent running on a server but controllable from my pocket" problem. The long-running-screen-per-project detail is the part worth stealing.
@mathewfjones1 [OpenClaw]
https://x.com/mathewfjones1/status/2057972078280904808
Mathew describes pairing a durable knowledge layer he calls TeamYou (people, companies, projects, decisions, investor context, operating history) with OpenClaw as the hands that reach into the tools where work actually happens. The interesting idea is the split: a persistent memory of the business on one side, an actuator on the other. It's still more concept than receipts, but it names the right problem, agents are only as useful as the durable context they can act on.
@ardur_sec [Claude Code]
https://x.com/ardur_sec/status/2057627806754865593
A build-in-public progress dump with actual numbers: 48 commits, 13 PRs merged, a Claude Code plugin shipping, Homebrew packaging in progress, plus a second project at 17 commits and 10 PRs. It's not a dramatic story, but the throughput is the point, this is what a small team's week looks like when the agent is doing most of the typing. The commit count is the receipt.
🗣 User Voice
User Voice
The single loudest signal across both Claude Code and OpenClaw was cost, and it has clearly tipped from grumbling into action. Users aren't just complaining about token bills, they're rerouting around them, wiring Claude Code's interface to DeepSeek, Kimi, Goose, or their existing subscriptions to keep the workflow and kill the invoice (@karriodev, @BrandonHorner4). The Microsoft cancellation gave everyone permission to say the quiet part: at scale, the model is affordable but the agent loop is not.
Rate limits, not capability, now shape real workflows. Heavy users describe rationing across tools, audit on Claude Code until the limit hits, finish on Codex until it resets (@hellobyedot). What they want is predictable headroom and clear cost attribution, which is exactly why the new /usage command got singled out as the feature people actually needed (@designtako).
The fastest-growing user base can't use a terminal yet. The standout adoption stories were a dad building a script in 20 minutes and the observation that the next million users "don't know what a terminal is yet" (@mustafaergisi, @mbajaj_). The demand here is for the GUI and desktop on-ramps, not more CLI power.
Trust and babysitting remain the unspoken tax. One blunt take: "Claude Code users have no clue how much time they're spending babysitting Claude... not fun to use a tool you don't trust 60% of the time" (@robdel12). And the gap that nobody's product closes: "deployment and hosting are still manual steps in every vibe coding tool I've tried" (@theahmedwalid).
For OpenClaw specifically, the recurring ask is reliability over features, the loudest complaints were setup friction and breakage after every update, with several power users describing multi-week struggles before churning to Hermes (@Sadok_Ayari).
The single loudest signal across both Claude Code and OpenClaw was cost, and it has clearly tipped from grumbling into action. Users aren't just complaining about token bills, they're rerouting around them, wiring Claude Code's interface to DeepSeek, Kimi, Goose, or their existing subscriptions to keep the workflow and kill the invoice (@karriodev, @BrandonHorner4). The Microsoft cancellation gave everyone permission to say the quiet part: at scale, the model is affordable but the agent loop is not.
Rate limits, not capability, now shape real workflows. Heavy users describe rationing across tools, audit on Claude Code until the limit hits, finish on Codex until it resets (@hellobyedot). What they want is predictable headroom and clear cost attribution, which is exactly why the new /usage command got singled out as the feature people actually needed (@designtako).
The fastest-growing user base can't use a terminal yet. The standout adoption stories were a dad building a script in 20 minutes and the observation that the next million users "don't know what a terminal is yet" (@mustafaergisi, @mbajaj_). The demand here is for the GUI and desktop on-ramps, not more CLI power.
Trust and babysitting remain the unspoken tax. One blunt take: "Claude Code users have no clue how much time they're spending babysitting Claude... not fun to use a tool you don't trust 60% of the time" (@robdel12). And the gap that nobody's product closes: "deployment and hosting are still manual steps in every vibe coding tool I've tried" (@theahmedwalid).
For OpenClaw specifically, the recurring ask is reliability over features, the loudest complaints were setup friction and breakage after every update, with several power users describing multi-week struggles before churning to Hermes (@Sadok_Ayari).
📡 Eco Products Radar
Eco Products Radar
Codex (OpenAI) — mentioned constantly, the default comparison and the tool people fail over to when Claude Code hits its limit
Cursor — recurring as the IDE-side comparison and part of multi-tool stacks
DeepSeek (V4 / Anthropic-compatible endpoint) — the cost-cutting backend of choice, routed behind Claude Code's interface
Hermes (Nous Research) — the orchestrator half of the OpenClaw-plus-Hermes pattern, and the migration destination for OpenClaw churners
Gemini CLI — appears across cross-harness experiments and the SEO-poisoning security alert
Copilot CLI — the tool Microsoft is consolidating onto internally as it drops Claude Code
MCP (Model Context Protocol) — the connective tissue under most of these stacks, with native plugins shipping for every major agent
Codex (OpenAI) — mentioned constantly, the default comparison and the tool people fail over to when Claude Code hits its limit
Cursor — recurring as the IDE-side comparison and part of multi-tool stacks
DeepSeek (V4 / Anthropic-compatible endpoint) — the cost-cutting backend of choice, routed behind Claude Code's interface
Hermes (Nous Research) — the orchestrator half of the OpenClaw-plus-Hermes pattern, and the migration destination for OpenClaw churners
Gemini CLI — appears across cross-harness experiments and the SEO-poisoning security alert
Copilot CLI — the tool Microsoft is consolidating onto internally as it drops Claude Code
MCP (Model Context Protocol) — the connective tissue under most of these stacks, with native plugins shipping for every major agent
Comments