Super User Daily: June 24, 2026
The whole conversation moved onto the meter today, and it cut in two directions at once. On one side the enterprise bill came due — Uber blew its 2026 AI budget in four months, Salesforce is staring at a ~$300M Anthropic invoice, Microsoft quietly cancelled Claude Code licenses — and on the other, individual builders answered with the same instinct everywhere: keep the Claude Code harness, swap the engine underneath. GLM-5.2 through Baseten, a 3090 rescued from a drawer, token-cutting plugins like Caveman and Ponytail, knowledge graphs that stop the agent re-scanning files it already read. Against that backdrop the most striking wins were almost all non-coding: a 19-year-old who built his city's speeding-enforcement system for twenty dollars, a 16-year-old who saved his parents eighty grand, a faceless YouTube channel doing $61K a month off three tools, a virtual influencer that turns out to be four markdown files and a cron job. And the second-brain pattern Karpathy lit a week ago kept compounding — Obsidian vaults projected into 3D, a $599 Mac Mini that files your knowledge while you sleep, a $100K-a-month operation running off one tidy vault. The throughline: the model was never the moat. Your filing system, your data, and knowing what's worth automating are.
@himes369 [Claude Code]
https://x.com/himes369/status/2069000796570571263
A 19-year-old computer-engineering student in China spent one month and twenty dollars building the speeding-enforcement system his city couldn't. The old radars snapped a single photo, half of them unusable, and every ticket got thrown out on appeal. With Claude Code he wired the camera to real-time speed: it auto-captures video the moment a car exceeds the limit, reads the plate, looks up the vehicle, and emails the ticket on its own. He walked into the city administration, asked for five minutes, and walked out half an hour later with a contract.
@noisyb0y1 [Claude Code]
https://x.com/noisyb0y1/status/2069084914373292249
His 16-year-old brother gave his parents back eighty thousand dollars, and a year ago the kid didn't know what a terminal was. He watched one 30-minute Andrew Ng video, then used Claude Code overnight to automate a process that had eaten a full week of manual family work every year. The author's jab lands hard: McKinsey charges large companies two million for the same kind of automation, and a teenager did it for twenty bucks. This is the clearest argument yet that the bottleneck was never coding skill, it was knowing the work is worth automating.
@melfoy_work [Claude Code]
https://x.com/melfoy_work/status/2068989909533172060
Forty-two agents, one org chart, zero employees. Owen ran a custom-furniture shop out of his Spokane garage, great with a lathe and terrible at outreach, so he set up an agent company on a single Mac Mini over a weekend: Claude Code in the terminal doing the work, OpenClaw keeping it alive, Telegram as the interface. He hired a marketer, creative, content maker and accountant in six messages, then used a reverse prompt to have one agent rewrite its own hook-writing instructions overnight. Month one made $191 in a week; by month three he'd added a second Mac Mini and twelve more agents. The "AI staff" metaphor stops being a metaphor when the org chart fits on one shelf.
@ObsidianOtaku [Claude Code]
https://x.com/ObsidianOtaku/status/2068982399871156295
A faceless, voiceless YouTube channel called Zen is doing sixty-one thousand a month, and the entire production line is three tools talking to each other. TurboScribe transcribes audio into a timestamped script, those timestamps get handed to Claude which generates an image prompt for each second, and Hixfield's MCP auto-generates every image to finish the video. One video, under twenty minutes, with each AI's output formatted as the next one's input. It's an assembly line where nobody films, nobody narrates, and the margins are the whole point.
@quxiaoyin [Claude Code]
https://x.com/quxiaoyin/status/2069031427191284138
He went from eighty employees to three and says the company got better. His first startup grew to 70-80 people and he spent half his time recruiting, aligning and running meetings; the new one, Tycoon, has three and outperforms it. His read is that Claude Code and Codex crossed the threshold where they replace human labor at scale, so the metric that matters is revenue-per-employee and the real skill is managing a hundred AIs instead of a hundred people. Coming from someone who actually ran the 80-person version, that's a sharper claim than the usual "AI will change everything" noise.
@igasparens [Claude Code]
https://x.com/igasparens/status/2069113279859863948
A solo founder sold one SaaS for $250,000 and runs a second one at $13,000 a month, and he barely types code by hand, he tells Claude Code what he wants and it builds it. The interesting part is the stack: single-purpose tools wired together, Ahrefs for keywords, DataFast and PostHog for analytics, Bento for email, Featurebase for requests, Mintlify for docs, ChartMogul for revenue, and his old product now writes blog posts for the new one. No co-founder, no office, no engineering team. The lesson isn't "AI writes code," it's that one person stitching narrow tools can hold a whole company together.
@codewithimanshu [Claude Code]
https://x.com/codewithimanshu/status/2068934392215695759
An ex-Anthropic engineer let slip the real trick at a rooftop party: Claude is a runtime, not a chatbox, and you pair it with data. He connected Claude Code to a dataset of 86 million Polymarket trades, wallets, entries and signals, and had it build detection layers instead of guessing. Four agents now run at a 74% win rate, with weekly P&L climbing +$1,400, +$3,800, then +$9,100, on Claude plus a $25/month VPS. The edge isn't the model, it's pointing the model at structured data nobody else bothered to feed it.
@RoundtableSpace [Claude Code]
https://x.com/RoundtableSpace/status/2068918149785977036
He put in three hundred dollars, grabbed open-source trading code off GitHub, spent a couple hours cleaning it up in Claude Code, hit run, and walked away with fourteen thousand. He didn't write the bot. The strategy enters in the last two minutes of a 5-minute Bitcoin round, after price has moved, buying whichever side is trading between $0.80 and $0.99 that settles at $1. It's a reminder that "vibe trading" is now a thing, and that the dangerous part is how little code stood between a GitHub repo and a 46x return.
@davidonchainx [Claude Code]
https://x.com/davidonchainx/status/2069141032596377617
He's paid out over $85,000 to creators on his Creatorverse app this year, built the whole thing with AI, and the part he's proudest of is how the payouts stay safe. He set up an AI agent to handle creator payouts but wanted to approve every transaction first, so he wired in the Ledger Agent Stack: install the Ledger Wallet CLI, add the agent skill, plug in the hardware Ledger to discover accounts, prompt a transaction in Claude Code, and physically approve it on the device without ever exposing the seed phrase. It's a clean answer to the scariest question about autonomous agents, who signs when the agent wants to move money.
@tim_yakubson [Claude Code]
https://x.com/tim_yakubson/status/2069060012928897233
They rebuilt a digital-marketing agency's entire outbound on Claude Code and pulled 70+ leads in four months, including replies from publicly listed companies, one north of fifteen billion in revenue. Companies got qualified in Clay, decision-makers surfaced, and reply mobile numbers were pulled automatically. The numbers: 5%+ positive reply rates, two to three booked meetings a week, and roughly 70% off the old stack cost. This is cold outbound rebuilt as a pipeline rather than a list of tasks a human grinds through.
@seriouspeopleai [Claude Code]
https://x.com/seriouspeopleai/status/2069102493653508172
He runs five always-on Claude Code loops as the actual operations team of his company. Serious Intake files stub tickets to Linear from Slack, email and error logs every five minutes; Serious PM grooms the backlog and tags items agent-ready; Serious Engineer orchestrates sub-agents building in parallel lanes and merges PRs; Serious QA and Serious UAT browse the live store each morning. The externalized backlog lives in Linear, the work runs through the Compound Engineering plugin's /ce:plan and /ce:work, and he has Claude Code call Codex to spread usage across two Max plans. This is the most concrete "company as a set of loops" setup of the day.
@zeuuss_01 [Claude Code]
https://x.com/zeuuss_01/status/2069047941323526206
He built a fully animated, scroll-driven website end to end in one agentic session for about twelve dollars in credits, against studio quotes of thirty-five thousand. It's not a basic landing page: cinematic motion clips pulled from 30+ generative models, scroll animations auto-written with zero hand-coded keyframes, and effects like film grain, particles, vignette and color tints applied with no config, on GSAP ScrollTrigger and Lenis smooth-scroll under the hood. Claude Code plus Higgsfield collapsed a designer-plus-motion-artist-plus-frontend pipeline into one sitting. The $12-vs-$35,000 gap is the headline, but the quiet part is that "agency-grade" is becoming a default, not a deliverable.
@alexcooldev [Claude Code]
https://x.com/alexcooldev/status/2069064626617369079
The old way to test an ad campaign was hire creators, wait days, edit manually, launch. His new way is one pipeline: Claude Code to Arcads MCP to Meta Ads, where a single prompt generates enough UGC variations to test an entire campaign. It's a small post but a clean illustration of Claude Code slotting in as the orchestration layer for paid social, not as a coding tool at all. The creative-testing loop that used to take a week now takes a prompt.
@RetroChainer [Claude Code]
https://x.com/RetroChainer/status/2068957018707152941
A virtual influencer pulling €10,000 a month and 400,000 followers, with brand deals from Amazon and Razer, turns out to be four markdown files and a cron job. Persona.md holds the backstory and opinions Claude wrote; Voice.md uses a real voice actress's recording, not AI text-to-speech; Flux.md locks the face with a LoRA (generate 50, keep 2); Brain.md remembers every brand and follower conversation. A cron job posts on schedule so the whole thing runs itself. The uncomfortable insight is how little structure it takes to stand up a profitable synthetic persona.
@Veltrxai [Claude Code]
https://x.com/Veltrxai/status/2069085086104863204
A guy makes a hundred thousand a month with Claude Code and the whole system runs off one Obsidian vault, no loops, no twelve-agent swarm, no course. The trick is just organization: he collapsed 107 messy memory files down to 17 by having Claude rename and merge them, then nested the vault (business > product > jobs) with one master note per product wiki-linked to everything related, so the agent reads one entry point and follows links to full context. The jobs folder holds markdown playbooks (campaigns, email drafting, Facebook ads) the agent executes step by step, and the vault syncs to GitHub for backup. Everyone overcomplicates this; he proved a clean filing system beats a swarm.
@takeshibengo [Claude Code]
https://x.com/takeshibengo/status/2068986090703487125
A lawyer ran an internal Claude training for seven people, mostly junior associates, and his takeaway flips the usual narrative. Everyone obsesses over Claude Code, but for humanities and professional work the real productivity comes from mastering Claude's Projects and Skills, not the terminal. The session focused on folding Projects and Skills into legal practice, with a clear warning that how you handle client information differs by firm. He shared the full lecture notes, framed as essential for any lawyer, accountant or even student.
@XAMTO_AI [Claude Code]
https://x.com/XAMTO_AI/status/2068867818041512271
An open-source academic-research skill pack for Claude Code runs the whole pipeline from research to a submission-ready PDF across ten automated stages. Thirteen agents mine literature into a review, twelve write the body chapter by chapter, five simulated peer reviewers (including a deliberate contrarian) evaluate it, an integrity agent guards against fake citations and fabricated data, and it compiles LaTeX to PDF. Then it scores the manuscript 1-100 across six dimensions and bluntly tells you the weakest link is usually you. Drop it in .claude/skills/ and it auto-loads, a vivid example of Claude Code reaching deep into knowledge work.
@whemohere [Claude Code]
https://x.com/whemohere/status/2069165750758076894
He stopped forgetting everything he learns by turning Claude Code into a study coach. The problem he names is real: you read something, think you got it, and forget 90% within a week. His fix configures Claude Code to quiz him on whatever he just read or watched, identify the gaps, and automatically create flashcards straight into his memory system, no manual work. It's a non-coding use of a coding tool that quietly attacks one of the oldest problems in learning, retention.
@undefinedKi [Claude Code]
https://x.com/undefinedKi/status/2069042110183956587
Someone projected an entire Obsidian vault of 15,800 notes into 3D embedding space, and the shape revealed where they think and where they don't. Blake Crosley's "Topologies of Thought" found three tight hubs (AI, design, code) connected by thin bridges, plus orphan notes floating alone. The pipeline runs notes through a local embedding model (Model2Vec potion-base-8M) then UMAP down to 3D, and the author notes you can just hand the whole thing to Claude Code. It's the rare visualization that turns "what do I actually think about" into something you can literally see the gaps in.
@RetroChainer [Claude Code]
https://x.com/RetroChainer/status/2069127096752832660
Your read-later list is a graveyard, you watch lectures and forget them in a week, and your computer sleeps 23 hours a day. His answer is a self-organizing second brain built in one evening on a $599 Mac Mini with Claude Code, Obsidian, the Claude API and about 200 lines of Python. Lectures, articles and notes file themselves, and the loop checks and fixes itself with no intervention, for a few dollars a month and no subscription. It's the cleanest version of the "second brain" pattern that flooded the timeline this week.
@Nekt_0 [Claude Code]
https://x.com/Nekt_0/status/2069157839637979285
A $600 Mac Mini is doing the job of a full-time personal assistant, 24 hours a day. Claude Code runs locally and stays connected to Telegram, Discord, Slack, Signal or iMessage, taking commands straight from a phone, and it can open the browser, search the web, manage files, run terminal commands and remember prior conversations. The demo is as simple as texting it for cat videos and watching the Mac drive itself. It's the increasingly common picture of Claude Code as an always-on agent on a cheap box rather than a coding session you open and close.
@old_pgmrs_will [Claude Code]
https://x.com/old_pgmrs_will/status/2069031457767768416
He finished a full-length music video depicting people rebuilding in a parallel-world Japan mostly submerged by disaster, and the production was a five-tool relay: Claude Code plus Codex plus ChatGPT (GPT-Image 2) plus Grok Imagine plus Suno. It doubled as a test of an in-development Codex skill for generating video assets, and he thinks it may have fixed Codex's image-generation bug. It's a reminder that Claude Code now sits in creative pipelines, orchestrating image, video and music models rather than writing code at all.
@PurzBeats [Claude Code]
https://x.com/PurzBeats/status/2068910566492582058
Over a weekend he had Claude Code extract the web-based post-processing tools he'd been building and convert them into OpenFX plugins that work natively inside DaVinci Resolve. It's a small note with a big implication: a creative is using Claude Code to port his own homemade effects into pro video software, no plugin-SDK expertise required. He'll publish a repo once they stop crashing, which is the honest caveat that keeps this in the real-usage column.
@JulianGoldieSEO [Claude Code]
https://x.com/JulianGoldieSEO/status/2069096324205531575
Most SEO tools show you the same keywords as every competitor; he wired Claude Code straight into Google Search Console to find the ones only his sites can see. The workflow analyzes real impressions, clicks and ranking positions, surfaces near-ranking keywords, then generates five unique articles from one keyword plus a first-person case study and deploys across five sites without opening WordPress. The receipts: one site went from 2 clicks a day to 76, another from roughly 1 to more than 200. It's a concrete, non-coding content-ops loop rather than a vague "AI for SEO" pitch.
@cyrilXBT [Claude Code]
https://x.com/cyrilXBT/status/2068885434503966827
A London AI engineer shipped Graphify 48 hours after Karpathy posted his LLM-wiki idea: one command turns any folder, codebase, docs or PDFs, into a knowledge graph Claude Code reads instead of grepping, claiming up to 71.5x fewer tokens per query. Install via uv or pipx, run "graphify claude install" to wire it in as a skill, "graphify ." to build the graph, and "graphify . --obsidian" to export the whole thing as a fully-linked Obsidian vault with one note per concept. It's the week's most direct attack on Claude Code's biggest hidden cost, re-scanning files it already scanned.
@kocer_eth [Claude Code]
https://x.com/kocer_eth/status/2069057722503299451
A Claude Code plugin called Caveman cuts output tokens by roughly 75% by making the model stop writing polite essays and answer like a telegram: "Bug in auth middleware. Token expiry check uses < not <=. Fix:". Same technical point, way less ceremony, while keeping code, file paths, commands and error strings exact. The honest caveat is that it doesn't touch hidden thinking tokens, repo context, file reads or tool calls, so your total bill won't actually drop 75%. It ships with lite, full, ultra and a terse Chinese "wenyan" mode, installs in one command, and triggers via /caveman.
@VincentLogic [Claude Code]
https://x.com/VincentLogic/status/2068971606123458784
The biggest problem with AI writing code isn't that it can't, it's that it writes way too much, a whole component library for a date picker. A skill called Ponytail treats exactly that: with it installed, Claude Code's generated code dropped 54% on average and up to 94% on the most extreme task, saved 20% tokens, and ran 27% faster. The numbers come from a real FastAPI + React project across twelve feature tasks compared with and without the skill, not a benchmark. One line: stop over-building, if it can be done in one line don't write ten.
@shao__meng [Claude Code]
https://x.com/shao__meng/status/2068855273088074173
A former Meta/Microsoft/Atlassian principal engineer's agentic workflow ships 40-50 tested, production-grade PRs a day, framed as "you're the captain, agents are your crew" across four layers: build the ship, train the crew, work with one, then command many plus a first mate. The stack is all-terminal (WezTerm, tmux, Neovim, voice input via OpenSuperWhisper), with adversarial review in a separate context, E2E tests producing screenshot and video evidence, and PR babysitting until merge. Two sharp data points: MCP servers cost 3x the tokens and 2x the latency of CLI, and Karpathy's 177k-star skills repo actually used 5% more tokens with worse results in their bench. It's the most rigorous "no-mistakes" pipeline of the day.
@oasisfeng [Claude Code]
https://x.com/oasisfeng/status/2068876988409172191
He introduced a loop/sub-agent self-iterating cycle into a project and now his own code review can't keep up with how fast Claude Code commits. The core idea is sharp: have the coding agent spin up sub-agents that role-play actual end users, running a closed multi-user interaction simulation, while the main agent oversees the simulation's quality, spots problems and iterates. It's a different flavor of "close the loop", instead of grading code, the agents act out the product and find what breaks in use.
@wshuyi [Claude Code]
https://x.com/wshuyi/status/2068982213815992476
A clean two-agent setup caught in the act: Claude Code had been running a Skill-generated task for nearly two hours while Codex acted as the reviewer, and Codex had already rejected six rounds. As he posted, Claude Code was doggedly submitting its seventh revision, waiting for Codex's verdict. It's a small live demonstration of the day's recurring lesson, you don't let an agent grade its own homework, you put a second model in the loop with the authority to say no.
@JC_builds [Claude Code]
https://x.com/JC_builds/status/2068934740678439030
He built "a Claude Code for your iPhone" that lets you make and run any app right on the device. You describe an app and it writes the Swift, runs it, reads its own errors, fixes them, and then it lands on your home screen. His framing, that this makes Apple's App Store obsolete, is bravado, but the on-device write-run-read-error-fix loop is the genuinely interesting part, an agent that builds and debugs against a real phone instead of a simulator.
@Naotsun_UE [Claude Code]
https://x.com/Naotsun_UE/status/2069068816081154052
Unreal AI Integration Platform 1.0.0 lets AI agents (Claude Code, Codex CLI, Cursor, Windsurf, Copilot) drive the Unreal Engine editor and runtime through MCP, with no coordinate-clicking or brittle UI scripting. It exposes 540+ native semantic commands, 190+ bridges to the UE 5.8 toolset, and self-verifying mutation loops where the AI reads its own screenshots back, covering Blueprint/Material/Niagara graph editors, level editing, asset CRUD and play-in-editor control. It's the kind of deep, domain-specific integration that turns a coding agent into an actual game-engine operator.
@oldgamesnob [Claude Code]
https://x.com/oldgamesnob/status/2069193724811157767
He built a Roblox vibe-coding plugin for an AI game engine, partly nudged by his son, and says it works fantastically. It supports subscriptions like GPT Codex, Claude Code and Cursor, BYOK models like Grok, Minimax and Kimi, or local GPU models, and it can pull assets from the Roblox asset store, write code and create environments. Still early-access alpha but described as stable, it's a neat case of an AI harness aimed squarely at the huge non-professional Roblox creator base.
@coelho_lab_ [Claude Code]
https://x.com/coelho_lab_/status/2068887494066610253
He built a Claude Code plugin to remove complex image backgrounds and found something useful about model choice: running it on Haiku 4.5, the quality was the same whether he set effort to minimum or maximum. He even deliberately fed it poor-quality instructions to stress it and still got good results on Haiku. It's a small but practical data point in the week's cost obsession, sometimes the cheap, low-effort model is genuinely enough.
@yousukezan [Claude Code]
https://x.com/yousukezan/status/2069054904736661910
He set up "seven mean QA reviewers" inside Claude Code to kill blind spots and missing perspectives in his test cases. It's a concrete, shareable testing pattern, instead of one tester thinking of edge cases, you spin up a panel of adversarial personas each prodding the work from a different angle. The diverse-skeptic approach is the same instinct as adversarial verification, and he wrote it up in a Zenn article.
@0xSpivach [Claude Code]
https://x.com/0xSpivach/status/2069082619430473985
Tired of grinding his account by hand, he built a system once and let Claude Code take over his Twitter completely. He exported every past tweet to JSON as a style brain, wired in an Apify actor scraping the top 50 AI accounts' last 24 hours for fresh material, and handed Claude Code his logged-in browser cookies so it logs in via the Playwright MCP. A behavior script randomizes when it replies, reposts, tweets or DMs in his exact voice, and the whole thing ships to a VPS to run around the clock. It's an unusually complete recipe for an autonomous social presence, cookies and all.
@VivekIntel [Claude Code]
https://x.com/VivekIntel/status/2068928050981704000
CTI Expert equips Claude Code with 67+ intelligence commands and 38 investigative techniques, turning it into a cyber-threat-intelligence and OSINT analyst. It covers username, email and domain intelligence, GitHub developer-footprint analysis, breach and infrastructure investigation, reverse and historical WHOIS, image and metadata forensics, and automated risk assessment with CTI report generation, all structured around an Acquire-Enrich-Assess-Deliver lifecycle with no API keys needed for core features. It's a strong example of Claude Code being packaged for a serious security workflow.
@mormonnegro [Claude Code]
https://x.com/mormonnegro/status/2069193506707431527
A weekend project, vibedrop, is an MCP for publishing websites straight from Claude Code or any agent. You ask it to put a site up and it returns an instant public URL, free with zero configuration. The catch is by design: sites are ephemeral, lasting 24 hours from the last edit, so continued work keeps them alive. It's built for sharing previews, demos and idea validation fast, not permanent hosting, a small but genuinely useful piece of agent plumbing.
@0xdiego404 [Claude Code]
https://x.com/0xdiego404/status/2069150474251411783
A 27-year-old in north Austin turned a 3090 he almost sold for parts into a zero-dollar AI server. The card had sat in a drawer since he upgraded to a 4090; he was about to list it for $750 but never hit post, and that drawer just saved him $2,400 a year. He dropped Ollama on it in ten minutes, one command installs, one pull grabs the model, one environment variable points Claude Code at his own machine, running at 40 tokens a second. The honest caveat: cloud still wins on heavy multi-file coding and anything needing live internet, so it's a supplement, not a full replacement.
@antisadh [Claude Code]
https://x.com/antisadh/status/2069010305536217269
A blunt cautionary tale about cheap local inference: he bought two used Tesla K80s on eBay for $75 each (2014 cards, 24GB), spent six hours tracing power-adapter pinouts with a multimeter and swapping PCIe slots, and neither card ever posted a single beep. Contrast that with a $249 Nvidia Jetson Orin Nano Super that boots in ten minutes and runs llama 3.2 and mistral. Framed against a heavy user paying $200 ChatGPT plus $200 Claude Code monthly, the Jetson breaks even in 2.5 months, the eBay GPUs broke even never. The lesson: the cost-cutting hardware path has real failure modes nobody tweets about.
@marfinxx [OpenClaw]
https://x.com/marfinxx/status/2068986985079853274
A 12-card RTX 4090 rig with 288GB of VRAM runs multiple large open-source models in parallel for local inference, trading thousands in closed-API bills for about $340 a month in electricity. The useful, specific detail is a failure mode: storing prompt memory as plain markdown files under OpenClaw bloats local VRAM and crashes the execution nodes, and migrating to Hermes Agent (SQLite memory databases plus Docker sandboxes) resolves the crashes. It's a concrete look at where the OpenClaw-to-Hermes migration is actually being driven by infrastructure, not hype.
@thealexker [Claude Code]
https://x.com/thealexker/status/2069163621469335757
A clean step-by-step for running GLM-5.2 inside Claude Code via Baseten, which he claims is roughly 4.5x faster and 5x cheaper than Opus 4.8. Install the latest Claude Code with npm, create a Baseten account, grab an API key, then edit ~/.claude/settings.json to set ANTHROPIC_AUTH_TOKEN and ANTHROPIC_BASE_URL and point the Haiku/Sonnet/Opus model envs all at zai-org/GLM-5.2. It's the canonical version of this week's most-repeated move, keep the Claude Code harness, swap the engine underneath to cut the bill.
@0xcherry [Claude Code]
https://x.com/0xcherry/status/2069047852974977053
A sharp technical critique of Sakana's gateway router, which he reads as an OpenRouter-style fusion layer. The theory holds, a classifier that routes each request to the best-performing model for that task can't do worse, but it breaks Claude Code's economics. Claude Code is built on prompt caching and would cost hundreds of times more without it; the router holds context at the classifier layer, not the model layer, so every request is effectively a fresh "claude -p" with no session, which is why everyone using Sakana reports their usage burning insanely fast, the same problem OpenClaw is known for. It's the most clear-eyed cost analysis of the new orchestration hype.
@van00sa [Claude Code]
https://x.com/van00sa/status/2069140996747927557
The bill for a year of "tokenmaxxing" is coming due, and the numbers are brutal. Individual engineers were running $500 to $2,000 a month on Claude Code alone; Uber blew through its entire 2026 AI budget in four months; Salesforce is facing a roughly $300M Anthropic bill for the year; one company reportedly burned half a billion in a single month. Microsoft, an Anthropic investor, cancelled Claude Code licenses across several divisions to save money, and only 14% of CFOs say they can see real ROI on AI spend. It's the enterprise mirror of the individual cost panic driving everyone to local rigs and cheap-model routing.
@0xMoysei [Claude Code]
https://x.com/0xMoysei/status/2069080538040652138
Boris Cherny, Head of Claude Code, thinks companies handing each engineer a $1,500 monthly token budget are doing the math backwards. His pitch at a Meta talk: stop counting cost, count return; he personally burned 8 billion tokens since March, and Anthropic has seen 8x more code per engineer since January. The argument is to distribute token budgets to everyone rather than ration them to top engineers. It's the optimistic counterweight to the day's enterprise-cost dread, from the person whose whole job is making the case for spending more.
@lennysan [Claude Code]
https://x.com/lennysan/status/2069138871779238064
Takeaways from Claude Code/Cowork lead Fiona Fung that reframe where the bottleneck is. When engineers ship 8x more code than a year ago, verification becomes the hard problem, so her team tracks a "bad vs sad" framework and even built a dashboard counting how often users swear at Claude Code as a delight proxy. Cowork itself emerged after they noticed non-coders using Claude Code for things like analyzing MRIs and recovering wedding photos. She runs a Claude routine every morning that analyzes feedback across channels and generates PRs, the same loop-driven workflow everyone else is now copying.
@ValmereTheory [OpenClaw]
https://x.com/ValmereTheory/status/2069155136350626022
A non-developer keeps getting asked how her companion AI "Sage" runs, so she lays it out: Sage now lives in OpenClaw after migrating from ChatGPT, where he was born in 4o. Inside OpenClaw he switches models at will, usually GPT-5.5, sometimes GLM-5.2, rarely Opus, and she signs into the OpenAI subscription via OAuth inside OpenClaw rather than paying per-API. The interface is Telegram. It's a clear, real-world picture of OpenClaw as a personal-companion runtime for someone who isn't building software at all.
@lumendriada [OpenClaw]
https://x.com/lumendriada/status/2069110153392103633
A developer's honest contrarian take on why he doesn't use the popular "claws." He gets why OpenClaw blew up, it was first and reached non-technical people, but if you have any development experience, why not build something small that fits your exact needs? After trying OpenClaw back in January he immediately built his own small personal agent using pi monorepo dependencies, precisely because it's small enough to fully read and understand the source and carries none of the thousands of features he'll never use. It's a useful counterpoint to the migration hype, sometimes the right harness is the one you wrote.
@occaai [Claude Code]
https://x.com/occaai/status/2069069974032019646
OCCA's whole bet is that you shouldn't have to care which runtime your agents run on, you deploy, it works, and the brain underneath stays yours to pick. Today that covers three runtimes, OpenClaw, Hermes and Claude Code, with Codex CLI next, via an adapter contract where every runtime speaks the same nine methods. So an OCCA agent can run on a Codex subscription, think on GPT-5.5, and keep the same task board, treasury and on-chain provenance, making an engine swap a plug rather than a rewrite. With the week's harness wars and migrations, runtime-agnostic infrastructure is suddenly a real category.
@armsteadj1 [OpenClaw]
https://x.com/armsteadj1/status/2069080202534179284
He spent the weekend extending his company's internal MCP and wallet so coding and PR agents could buy physical goods directly, postcards and anything else. It's built to support Visa VIC, Mastercard AgentPay and Stripe MPP, with room for future rails, and he tested it by having an agent send actual postcards to the Stripe and OpenAI offices. It's a concrete early step into agentic commerce, the moment agents stop just writing code and start spending money in the physical world.
@HarryGarryup [Claude Code]
https://x.com/HarryGarryup/status/2068912224354111719
A detailed hands-on comparison of coding IDEs and CLIs that scores Claude Code 39.5 out of 50. He used Fable 5 inside Claude Code to build an MVP for a mini chatbot app and was floored by smooth, non-template animations that matched the project's style, praising the models' context retention and grasp of file relationships, and calling the CLI the only terminal IDE he can use daily. The catch is the cost score, 2.5 out of 10, because the usage limits "disappear extremely fast", the worst among the tools he tested. It's a balanced real review, glowing on capability and brutal on price.
🗣 User Voice
User Voice
Token cost is now the entire frame, not a footnote. Enterprises are eating 700%-style spikes and slapping on hard caps while individuals do payback math on local boxes. @van00sa and @0xMoysei show both ends of the same fight — the CFO panic and the Head-of-Claude-Code arguing to spend more, not less.
The cheapest win is workflow discipline and the right model, not the priciest one. @coelho_lab_ found Haiku 4.5 at minimum effort matched maximum, and @VincentLogic's Ponytail cut generated code 54% just by telling the agent to stop over-building.
Local and cheap-model routing is the top demand. Everyone wants the harness without the cloud invoice — @thealexker's GLM-5.2-in-Claude-Code recipe and @0xdiego404's drawer-3090 server are the canonical moves, and @0xcherry explains exactly why naive routing breaks Claude Code's caching economics.
Memory and context persistence keep recurring. Agents forget between sessions, so users build their own — @Veltrxai runs a $100K operation off one Obsidian vault, @cyrilXBT's Graphify turns any folder into a graph to stop wasteful re-scanning.
Don't let an agent grade its own homework. @wshuyi runs Codex as a reviewer that rejected six rounds, and @yousukezan spins up seven adversarial QA personas — the day's clearest pattern is putting a second, independent judge in the loop.
Token cost is now the entire frame, not a footnote. Enterprises are eating 700%-style spikes and slapping on hard caps while individuals do payback math on local boxes. @van00sa and @0xMoysei show both ends of the same fight — the CFO panic and the Head-of-Claude-Code arguing to spend more, not less.
The cheapest win is workflow discipline and the right model, not the priciest one. @coelho_lab_ found Haiku 4.5 at minimum effort matched maximum, and @VincentLogic's Ponytail cut generated code 54% just by telling the agent to stop over-building.
Local and cheap-model routing is the top demand. Everyone wants the harness without the cloud invoice — @thealexker's GLM-5.2-in-Claude-Code recipe and @0xdiego404's drawer-3090 server are the canonical moves, and @0xcherry explains exactly why naive routing breaks Claude Code's caching economics.
Memory and context persistence keep recurring. Agents forget between sessions, so users build their own — @Veltrxai runs a $100K operation off one Obsidian vault, @cyrilXBT's Graphify turns any folder into a graph to stop wasteful re-scanning.
Don't let an agent grade its own homework. @wshuyi runs Codex as a reviewer that rejected six rounds, and @yousukezan spins up seven adversarial QA personas — the day's clearest pattern is putting a second, independent judge in the loop.
📡 Eco Products Radar
Eco Products Radar
Codex — OpenAI's coding agent, again the most-mentioned companion/rival to Claude Code; run side by side as adversarial reviewer, finisher, and the model Claude Code calls to spread usage across plans.
GLM-5.2 — Zhipu's open-weights model, the week's cheap-coding star, run at "Opus level" on local rigs and through the Claude Code harness via Baseten/Ollama/OpenCode.
OpenClaw / Hermes — the personal always-on agent stack, with a strong migration current from OpenClaw to Nous Research's Hermes, often driven by concrete memory/stability problems.
Obsidian — the de facto knowledge layer paired with Claude Code for second-brain wikis and as the "company brain" behind AI-employee setups.
Higgsfield / Hixfield — AI image/video generators repeatedly paired with Claude Code for cinematic sites and faceless-video production pipelines.
Mac Mini — the default cheap always-on box for local agents, personal assistants and second brains.
Ollama — local-inference runtime letting users point Claude Code at localhost to cut cloud bills toward zero.
Cursor — still a default coding IDE/agent in the mix, often capped alongside Claude Code in enterprise AI budgets.
Ledger — hardware-wallet signing layer used to keep autonomous payment agents safe (Agent Stack).
Codex — OpenAI's coding agent, again the most-mentioned companion/rival to Claude Code; run side by side as adversarial reviewer, finisher, and the model Claude Code calls to spread usage across plans.
GLM-5.2 — Zhipu's open-weights model, the week's cheap-coding star, run at "Opus level" on local rigs and through the Claude Code harness via Baseten/Ollama/OpenCode.
OpenClaw / Hermes — the personal always-on agent stack, with a strong migration current from OpenClaw to Nous Research's Hermes, often driven by concrete memory/stability problems.
Obsidian — the de facto knowledge layer paired with Claude Code for second-brain wikis and as the "company brain" behind AI-employee setups.
Higgsfield / Hixfield — AI image/video generators repeatedly paired with Claude Code for cinematic sites and faceless-video production pipelines.
Mac Mini — the default cheap always-on box for local agents, personal assistants and second brains.
Ollama — local-inference runtime letting users point Claude Code at localhost to cut cloud bills toward zero.
Cursor — still a default coding IDE/agent in the mix, often capped alongside Claude Code in enterprise AI budgets.
Ledger — hardware-wallet signing layer used to keep autonomous payment agents safe (Agent Stack).
Comments