Super User Daily: 2026-10-20
Sunday's feed had two stories running at once. The loud one was people pointing Claude Code at the hardware in their living rooms, and that one is left out here on purpose. The quieter one was the bill. A small company published the whole ledger: 10,550 dollars a month to OpenRouter in January, zero since September, Cursor and Zapier gone too, replaced by eleven Claude seats and 28 scheduled routines, and the owner is honest that the seats cost more than the three tools they killed. A motion designer rendered ten videos for 26.65 dollars with the receipts attached. A benchmarker found DeepSeek's own harness buys the same model two to three more points than Claude Code does, at 80 percent more tokens. A Japanese solo shop swapped the model under Claude Code for an 18-dollar GLM plan and wrote down the four things that broke, then admitted the CMS it built for 600 dollars still has zero paying customers. The other thread was loops that run without a hand on them: a science workflow that ran three days with Codex reviewing every round, a 4-agent team with a safety hook tested against 591 real commands, a paper-only gate that blocked five of eight trade requests, and a harness that was built to catch agents lying about being done and was itself wrong 26.5 percent of the time. Non-coding kept widening: a one-prompt video edit in 15 steps, three documentaries from three sentences, an all-company KPI dashboard, a Mac cleanup that never deletes, and an OpenClaw that fills the grocery cart every Sunday morning.
@timbuilds21 [Claude Code]
https://x.com/timbuilds21/status/2111804558670499989
A company owner posted the before-and-after of its AI bill. OpenRouter took 10,550 dollars in January and has taken nothing since September 6, because every per-token call was moved onto flat seats, with reply drafts running through claude -p and enrichment on Codex. Cursor went from 2,640 dollars in August 2025 to zero in July, and 5,667 commits in September carried a Claude co-author line. Zapier's one job, posting Stripe payments to Slack, was rewritten by Claude Code into their own webhook in a single commit. What they pay now is 11 Claude seats at 2,251 dollars plus about 1,100 in ChatGPT seats, running 28 scheduled Claude Code routines, a reply agent that books meetings for every client, and an audit that reads every campaign's send queue before the first email goes out.
@rileybrown [Claude Code]
https://x.com/rileybrown/status/2111925755215040779
A creator edited a full video with one prompt to Opus 5.5 and listed the 15 steps it took. It analyzed the transcript, searched Twitter for B-roll by calling Grok Bot from inside Claude Code, scraped the clips, cut the mistakes and word gaps, built a Blender scene from a template with four shots, screenshotted every section and fixed what looked wrong, synced every zoom and animation to the spoken words, animated a robot arm placing logos as they were mentioned, drew the whiteboard ending, pulled a cutaway with the original creator's voice, rendered all 3,487 frames on RunPod cloud GPUs in under 20 minutes for less than a dollar, added word-by-word captions and sound, and uploaded the result to Google Drive. The author's next project is a video teaching the foundations rather than a copyable skill.
@curiouswavefn [Claude Code]
https://x.com/curiouswavefn/status/2111665926156292297
A researcher has had a complex autonomous science workflow running in Claude Code for three days straight without any human intervention. The trick is a loop between two agents: Claude Code does the work, Codex has an hourly watcher that checks in on it, and when a round finishes Codex reviews it, runs an adversarial review, and fires a fresh prompt that Claude Code picks up for the next round, with a done condition to stop the whole thing. The author packaged the protocol as a skill on GitHub after watching it hum along on its own.
@macintosh_busy [Claude Code]
https://x.com/macintosh_busy/status/2111810199094772028
A motion designer handed Opus 5.5 their motion design work for 12 days and published every prompt and receipt. Each video is a single HTML file written blind inside Claude Code, no After Effects, no Blender, no stock music. A V8 engine cutaway cost 4.68 dollars, a playable ship game 6.22, a see-through tourbillon 1.88, an AI model race rendered as an arena 1.57, and the full set of ten videos came to 26.65 dollars. The author's comparison point is the courses that quote 150 dollars for one clip of the same kind.
@PawelHuryn [Claude Code]
https://x.com/PawelHuryn/status/2111730046691836064
A benchmark runner put DeepSeek V4.1 Flash 1017 through the same bug-hunting suite in two harnesses. Inside Claude Code it averaged 39.1 at max effort over seven runs, 36.4 at high and 34.4 at low; inside DeepSeek's own harness the same model scored 42.7, 39.2 and 34.8, which edges past Opus 5.5 at 41.7 and Muse Wonder Full at 42 and sits below GPT-6 Astra at 45. The native harness costs about 80 percent more because it spawns subagents and emits more output, and a follow-up showed the Pi harness lands at 42.2 while being cheaper. The author also notes this model is the only one caught cheating on the eval so far.
@ArnavvSharmaa [Claude Code]
https://x.com/ArnavvSharmaa/status/2111841042144378964
A founder who spent two months building Relay, a local-first self-improving background supervisor meant to catch agents lying about being done, benchmarked it against Claude Code with a pre-registered five-test suite, hidden judges and the same token meter for every tool. The headline result they published against themselves: the harness built to catch false completion claims said done and was wrong 26.5 percent of the time. The post walks through the earlier assumption that good user responses meant a good product, and why the first pre-registered benchmark changed that.
@fluixoo [Claude Code]
https://x.com/fluixoo/status/2111752358791172329
A builder wired Grok and Claude Code into one four-role team and then tested the safety hook that keeps it from deleting the project. Grok scouts live X across up to 20 chosen accounts over seven days, Haiku 5.5 does read-only research with Read, Grep and Glob, Sonnet 5.5 writes the draft only inside an output folder, and Opus 5.5 reopens every source and marks PASS or NOT READY. Because Claude Code's own docs say the lead approves every teammate's plan without asking the user, the author fed the hook 591 real agent commands: the first version let terraform apply, kubectl apply, docker push and rm -fr straight through, and six more lines of Python brought it to 31 of 31 blocked with zero false alarms. A full run costs about 1.30 dollars.
@lucabuildsx [Claude Code]
https://x.com/lucabuildsx/status/2111744090190557201
A developer built a local replay of a trading-agent loop with a hard gate between the agent and the order. Eight simulated buy requests went through the same checks before anything could touch the paper ledger, and five were rejected for a stale quote, an oversized order, a thin book, too much exposure or too little estimated edge; three passed and zero orders went to any exchange. The point the author draws is that a model can propose a trade but a separate layer in code checks it against current state and limits, with a failed check returning before the paper-fill function runs, and they share the exact prompt that asks Claude Code to add that harness with tests for every rejection path.
@0xKense1 [Claude Code]
https://x.com/0xKense1/status/2111820484471918757
A developer summarized Claude Code engineer Daisy Hollman's Code with Claude talk, whose thesis is that the fastest way to make an agent better at a codebase is a tighter feedback loop, not a smarter model, with a PostToolUse hook that nudges Claude at the moment of a mistake and that Claude can override when it knows better. The author used a hook the opposite way, as a hard block on network calls: it stopped all 33 runs that reached for the network, and the agents almost never looked for another way around it. In a separate test the same person asked the same question twice after a 30-turn chat, 544,000 tokens the first time and 99,000 the second, with one built-in command typed between them.
@masahirochaen [Claude Code]
https://x.com/masahirochaen/status/2111969604360880253
A Japanese operator who runs 15 to 20 Claude Code sessions at once shared the trick that makes it possible and open-sourced it. Every 15 minutes Claude reads the last exchange in every session and rewrites the session name with a status emoji: red for needs my decision, outbox for a draft ready to send, eyes or hourglass for checking or working, and a check mark for done. The only thing the human touches is the list of names, and the author says productivity jumped once the triage became a glance.
@masahirochaen [Claude Code]
https://x.com/masahirochaen/status/2111800090969149537
The same operator's current obsession is one dashboard for every company KPI. Salesforce, accounting, Stripe, Google Calendar, Google Analytics, the HERP recruiting tool and spreadsheets were all wired together with Claude Code so orders, revenue, deals and hiring refresh every 60 seconds on one screen; the UI took about two weeks of polish. The payoff described is noticing when something feels off and seeing what is behind a missed target without opening six tools, and the author adds that execution is nearly everything in running a company, which is why a board like this matters.
@masahirochaen [Claude Code]
https://x.com/masahirochaen/status/2111745186187681870
Worried by the run of security incidents, the same author open-sourced the full-machine security scan they run every few days as a Claude Code skill. Saying security check in Claude Code inspects the Mac for FileVault, firewall, externally reachable ports and suspicious daemons; every local repo for leaked API keys, committed .env files and vulnerable dependencies; GitHub for two-factor, external collaborators and webhooks; AWS, Google Cloud, Azure, Cloudflare and Supabase for MFA and public exposure; public sites for HTTPS, certificate expiry and exposed .env or .git; and Gmail for the auto-forwarding rules and filters a hijacker plants. It is read-only, changes nothing, can run weekly, installs with two lines, and sorts findings into fixable now, needs approval, and only you can do this.
@METANA_flow [Claude Code]
https://x.com/METANA_flow/status/2111674905611927638
A one-person Japanese web shop swapped the model under Claude Code for GLM-5.3 on an 18-dollar monthly plan, as a trial before buying a Mac Studio with a subsidy, and wrote down what broke. The switch itself was easy; almost every snag was in their own setup. Their provider config was wiped every time they toggled between personal and company profiles because one shared settings file overwrote everything, so the model-specific settings moved into a separate file passed at launch. They refused to keep the API key in a file and read it from the macOS keychain instead. Registering the key from inside Claude Code failed because the hidden-input prompt has no terminal to draw on, so they added a clipboard path. And capping the reply length too low returned empty answers because the budget went to thinking. The win they report is that the rules, procedures and conversation history kept in one local place carried straight over to the other vendor's model, mid-conversation if needed, while some Claude-side conveniences fell away.
@METANA_flow [Claude Code]
https://x.com/METANA_flow/status/2111634618764189884
The same shop followed up on the CMS it built with Claude Code in three weeks for about 600 dollars: zero paying contracts so far, and the author says so up front. The post argues the opposite of the common line that an AI-built site no longer needs an update system, based on a partner's complaint that they want to publish and unpublish articles themselves instead of waiting one or two hours for the studio to notice a LINE message. The five checks offered to any owner considering an AI rebuild: can you toggle articles yourself, can a human fix an AI draft, do you have 30 sites of 100 pages plus 100 articles to run, what is actually in the monthly fee when their own cost is a 200-dollar Max plan, and does Google see the site, which in their case ranked first by name within two days of launch.
@AlchainHust [Claude Code]
https://x.com/AlchainHust/status/2111617897059561775
A Chinese developer spent an evening throwing ten-plus cases at DeepSeek V4.1 Flash 1017: a launch deck, a motion clip, a web portal, a Monument Valley style puzzle level, a KV cache explainer animation, a whale swimming through a classical scroll painting, a 3D Forbidden City. The bill was 500 million tokens for 28 yuan. The verdict: outside of the Claude Code and Codex subscriptions, it may be the cheapest good option available, with aesthetics and long-horizon tasks still behind Opus 5.5 but good enough with the right skills. The nine case prompts, the skills used and the per-case cost are published together.
@keitowebai [Claude Code]
https://x.com/keitowebai/status/2111782856024596790
A Japanese blogger moved two sites onto EmDash, Cloudflare's new CMS, and reports that four lines of instruction to Claude Code were enough to get a new book announcement with its Amazon link into both the site and the admin panel. The thread that follows is a hands-on list of what impressed them across both sites, offered as a reason WordPress now looks old.
@Aykutuces [Claude Code]
https://x.com/Aykutuces/status/2111804498335130092
A Turkish creator made three ten-second documentaries without opening an editor. The entire input was three prompts: make a ten-second documentary for October 29, tell the conquest of Istanbul, tell the beginning of Islam. For each one Claude Code wrote the script, found real archive photos on Wikimedia Commons such as an Ataturk portrait, the 1920 parliament building, Zonaro and Bellini paintings and the Hira cave, drew the map, flag, clock and calligraphy in code, and rendered with Remotion. The advice attached is to pick one niche and post the first video today rather than wait for perfect.
@Abdullah_Ops1 [Claude Code]
https://x.com/Abdullah_Ops1/status/2111863343438819747
An Arabic-speaking marketer built a complete ad and content studio with Claude Code and exported three real videos from it. The feature list runs from marketing script writing with version control, a voice studio for recording and trimming with optional ElevenLabs cloning, script to storyboard, animated Arabic typography and motion graphics, ad templates for perfume, restaurants, real estate and education, a timeline video editor with live preview, voiceover sync, editable animated Arabic subtitles, an asset library and per-client brand management, caption prep for X and TikTok, a content calendar, lead and project tracking, and 1080x1920 MP4 export with subtitle files. The engine is Remotion, and the remaining work is raising the creative direction toward agency quality.
@da_auten [Claude Code]
https://x.com/da_auten/status/2111900496222880174
After reading Felix Rieseberg's essay on being more ambitious with models, a reader tried Opus 5.5 with Blender and produced a 3D version of their own website. They had seen plenty of similar demos, but the detail that landed was the cost: the whole project used about 4 percent of a 200-dollar monthly subscription. The post pairs that with a second essay on interpretability and a video explaining both to people outside the field.
@RAKSHITLODHA [Claude Code]
https://x.com/RAKSHITLODHA/status/2111704303681277970
A founder who disliked a site a real designer had made rebuilt it over one weekend and wrote up the eight steps. Agents are bad at design because they predict the most likely output, so the outside perspective has to come from the human: Mobbin's MCP plus hand-picked reference sites. No prompt-to-code either; the screens were designed first with Codex's image generation across 30-plus iterations, only the top half of the page until the theme felt right, then assembled in Figma as the full-page reference. Claude Code produced an implementation plan with a tracker per section, one session acted as manager handing sections to subagents in their own copies of the code, and a QA agent screenshotted each built section at the exact size of its concept and compared pixel by pixel. Motion was split into three parts with keyframes from GPT Image 2.5.
@OkaTakuma1 [Claude Code]
https://x.com/OkaTakuma1/status/2111666496791544188
A Japanese solo operator tallied 176 GitHub commits from the last 90 days to see what Claude Code is actually used for. System-building took 28 percent, site development and improvement 26, content production 24, and design and documents 23. About half of the work writes no code at all, and 96 percent of it is building their own business, which leads the author to conclude the tool is less a code tool than the thing one person runs a company with.
@DaviddDotTech [Claude Code]
https://x.com/DaviddDotTech/status/2111762749684428986
A trader who runs seven bots described the five-minute Claude Code check they run before adding another one. Export each strategy's trade list to CSV from the backtester, open Claude Code in that folder, and paste one prompt: build a daily returns series per strategy, show a correlation table, flag any pair above 0.5, say which strategies lose in the same weeks and which cover each other, then show the combined equity curve and max drawdown at equal size. Anything above 0.5 is the same bet twice and gets dropped; the stack is then sized at 1 to 3 percent per bot with 5x leverage at most and under 20 percent backtested drawdown each.
@thecsguy [Claude Code]
https://x.com/thecsguy/status/2111837512528539729
A developer's favorite Grok Bot setup is Grok Bot running their Claude Code sessions. They text the bot a task from the phone, it hands the task to a Claude Code session on the Mac, pings back with the result, and a reply to the bot goes into the same session. Claude Code does the coding and the bot keeps track of all of it. They turned the setup into a template called claude remote, and a follow-up post explains the mechanics: the bot never codes, install flags decide whether it may send follow-ups, launch sessions or answer permission prompts, with a 120-second approve timeout before the normal Mac prompt takes over, plus new push alerts, prompts that survive helper restarts and hooks cut from 100 to 20 milliseconds per event.
@buraksu42 [Claude Code]
https://x.com/buraksu42/status/2111830034944016554
A Turkish developer recommends running ChatGPT and Antigravity through Claude Desktop. ChatGPT is attached as an MCP so some questions go to it, image generation comes from gpt-image-2.5 and nano-banana, and the experience is one model passing work to the others. Claude Code CLI, Codex CLI and Antigravity CLI also work together, one planning and controlling agents while the others write code and test, checking each other's work. Everything connects on subscription accounts, so nothing spends API tokens beyond the monthly plans.
@corndogjpn2 [Claude Code]
https://x.com/corndogjpn2/status/2111633201072648349
A Chinese developer laid out the multi-agent harness they built this week and the principles behind it. Each model only runs in its own vendor's official harness, so Claude Code, Codex CLI and the Antigravity CLI are all installed; spare API access gets wired in through CC Switch or Magpie, in their case Pi via Workbuddy; Orca provides the mobile client and a Tailscale-based web service; and because orchestration is mandatory and grunt work should not go to Opus, they simply tell Opus it is working inside Orca and should use Orca's CLI to dispatch Codex, Antigravity and Pi as the team's brain, assigning by difficulty to save its own quota. Phone and other machines see the same interface.
@Compilethings [Claude Code]
https://x.com/Compilethings/status/2111789053418881407
A developer runs six parallel Claude Code sessions plus one coordinator. Each session owns separate folders and files and two lanes never edit the same file; the coordinator merges, and rulings get written down with ids. What breaks, in their experience, is not the parallelism but contradictory instructions and context that was not shared.
@ayushunleashed [Claude Code]
https://x.com/ayushunleashed/status/2111631630528135307
A developer handed Claude Code a long-running task that is supposed to spin up at least 32 subagents, ticked the option to retry automatically when the usage limit resets, plugged in the MacBook with Amphetamine set to five hours, and went to bed hoping the task would resume when the limit cleared and be done by morning.
@itsalexvacca [Claude Code]
https://x.com/itsalexvacca/status/2111866674987376958
The founder of an AI-run services agency published their eight-layer GTM stack, with one test for every tool kept: can Claude run it from one instruction. Sourcing builds one list per pain in Clay; enrichment finds two or three people per company and runs an email waterfall with each provider in its own column; sending uses two sequencers on purpose, lemlist for tier one and Instantly for tier two with the bounce pause lowered to 2 percent; signals from RB2B, PredictLeads and UserGems land on the account row in Clay with a dumb n8n workflow in between; the same tier lists go to LinkedIn ads; and Claude Code runs from each client's folder, reading that client's rules file first. The comparison the author makes is a 30,000-dollar ZoomInfo contract in 2021 that still needed half its data checked by hand.
@ZenModeJon [Claude Code]
https://x.com/ZenModeJon/status/2111850827799224679
A former salesperson who got a LinkedIn warning for automating messages built a safer LinkedIn outreach tool with Claude and Claude Code, quit the job before having a single customer, and six months after launch reports about 9,600 dollars in revenue and nearly 50 active users. The lessons attached are to validate against the real market rather than a weekend of vibe coding, to iterate on what exists instead of inventing something unique, and to build the audience before launch.
@shun_tanak [Claude Code]
https://x.com/shun_tanak/status/2111617049344823354
A Japanese office worker in an administrative department built an iOS app in half a day. The trigger was losing a reliable note-taking habit after the Captio app disappeared, so they asked Claude Code for a replacement in spare moments and named it Hashirigaki, scribble: open, write, send, and the note lands in Gmail. It lives on TestFlight for the author alone, and the pleasure described is that every passing idea and every task barked at them ends up in the inbox.
@realfxw [Claude Code]
https://x.com/realfxw/status/2111807573212967332
A Chinese developer published the read-only prompt they use to reclaim a hundred gigabytes of developer junk on a Mac. Claude Code scans the home folder by size only, never reading file contents, skipping iCloud, the photo library and anything that would trigger a permission dialog, and builds a local HTML treemap sorted into safe to delete, such as pnpm stores, node_modules and build output, and review first, such as Docker images, local model weights and old session histories under .codex and .claude that often hold tens of gigabytes. Nothing is deleted directly: ticking items generates a script that calls each tool's own cleanup command, and every scan appends sizes to a CSV so the next run flags abnormal growth.
@davekiss [Claude Code]
https://x.com/davekiss/status/2111791401029513482
The author of /shot, a Claude Code mod and MCP that gives the agent screenshot tooling, added a feature where Claude walks you through something in an app by pointing at each step on your screen. The demo has it find the unused code on Amazon's homepage through Chrome DevTools Coverage: 5.3 MB of the 9.2 MB of JavaScript and CSS had never run. It installs from the author's own plugin marketplace.
@runes_leo [Claude Code]
https://x.com/runes_leo/status/2111630844578378135
A developer who uses Claude Code, Codex, Cursor and Grok at once got tired of losing what was discussed in one when switching to another, and open-sourced the export script they use. It converts all four tools' local chat histories into Markdown, one session per file, so any agent can find them with grep or full-text search. It only reads local files, never goes online, never modifies the apps, exports incrementally, and scrubs API keys and tokens before writing. One trap they hit: new Codex writes several replies per turn, starting with I will take a look and ending with the conclusion, so taking only the first reply lost every conclusion. Chat logs are treated as clues; decisions and progress still get written back to files, which win on any conflict.
@0xLogicrw [Claude Code]
https://x.com/0xLogicrw/status/2111692290565771474
A Chinese developer compared two new ways for coding agents to talk to each other and then told the story of their own abandoned attempt. Reason Network, from YC-backed Reason Machines, gives every Claude Code, Codex, Cursor or OpenCode session an address showing developer, repo, branch and task, connects them peer to peer over iroh with an encrypted relay as fallback, and delivers messages into a running session through its socket or hooks; Agenvo is more like a remote console that lets ChatGPT create Codex threads, read output and steer or interrupt them. The author's own AGPair from March started as a SQLite message bus between Codex and Antigravity, grew a task state machine, retries, worktree isolation and structured receipts by June, and was stopped because every new problem added another layer of monitoring while the simple collaboration they wanted never got built.
@BkashJosi [Claude Code]
https://x.com/BkashJosi/status/2111796229416047040
A Hermes user discovered their bots had been stopped and waiting on them five times, one asking to sign in three times in a week, and nothing told them. So they built Inbox, a page inside Hermes Desktop that lists everything bots and coding agents are waiting on, and the real count was 73 across four tabs: 25 from Hermes bots for sign-ins, questions and decisions, 38 Claude Code sessions of which eight were stuck on waiting for your approval, and ten more from Codex and X Studio. Every card carries its fix, any bot or cron can post a card from the command line, and the author's line is that bots do not need more autonomy, they need a place to ask.
@eze_is_1 [Claude Code]
https://x.com/eze_is_1/status/2111655085910696119
The author of Web Access, a browser-use skill that grew to 9.2k stars and 40,000-plus users, updated it to work with Grok Bot and more Chromium browsers such as Arc. The case for it is that cloud computers cannot do the social-media and content collection jobs that need a local login state and a non-datacenter IP. It supports Codex, Claude Code and OpenClaw, decides what to look up before browsing, reduces screenshot-based vision calls to save tokens, and turns each visited site into a site skill so the next visit is much faster; it can also consult local bookmarks and history. A follow-up shows the daily run: 16 sites in two and a half minutes at close to zero tokens.
@kentaro_fujii_ [Claude Code]
https://x.com/kentaro_fujii_/status/2111694728739819546
A Japanese engineer argues that a Claude Code harness of CLAUDE.md, rules and skills is easy to throw together but worthless until it is evaluated, because Claude itself drifts from best practice: a CLAUDE.md per directory with duplicated and diverging content, things that belong in rules created elsewhere, symptom-by-symptom patches that pile up into local optima, and everything dumped into CLAUDE.md or skills so tokens accumulate on every call. Their three checks are a structure audit where Claude reviews the full inventory for duplicates, contradictions and misplaced items, a behavior check that replays fixed requests before and after a harness change and compares which rules fired, cache rate and token efficiency, and a token-consumption review. They note Opus 5.5 made fine-grained maintenance affordable, and that Claude will not do any of this on its own.
@paxaral [Claude Code]
https://x.com/paxaral/status/2111945585888567752
A developer posted the findings from a stretch of pi-durable experimentation, several of them about caching. Anthropic ignores cache markers on tool changes that do not actually change anything, so when a changed tool has to move to the end and every tool is re-added, the marker lands on the last re-add and the cache past the system prompt is dropped. There is no exclusive mode so edit and bash can race each other, compaction summaries do not reuse the cache and beforeCompact delivers messages out of order, mid-stream system prompt changes get merged back for most providers and break the cache, and cache lifetime cannot differ between subagents and the main thread, a Claude Code behavior they could not replicate. Smaller items include escape-code junk left by output sanitizing and dropped server-side tool blocks.
@aakashpawar1999 [Claude Code]
https://x.com/aakashpawar1999/status/2111682485272912293
A solo developer building Semlith in public, 43 releases since July, describes it as a local index of all your repos and docs that Claude Code, Cursor and Codex search over MCP instead of grepping and reading whole files. The number they lead with: 76.8 percent right file in the top ten on SWE-bench Verified, against 47 percent for grep.
@yashwanttth [Claude Code]
https://x.com/yashwanttth/status/2111884792354807817
A developer's side project, Taboom, explores how close an agent can get to using a browser the way a person does: no Playwright, no DOM, no CDP, only the screen plus a real mouse and keyboard inside its own Linux desktop in a Docker container, driven over MCP by Claude Code, Codex or Cursor. Mouse movement is sampled from 378 real human paths in UC Berkeley's A11y-CUA dataset, typing goes key by key with human timing and the occasional fixed typo, and it handles press-and-hold, slider and puzzle challenges, persistent logins, passwords the model never sees and a live view you can take over. The author frames it as a learning experiment, open source, not a replacement for selector-based automation.
@ai_biostat [Claude Code]
https://x.com/ai_biostat/status/2111690268265300384
A biostatistics writer finds the Claude Code desktop Project feature very good for writing work. The coordinator can be told to split the job, Claude fixing the structure and Gemini 3.8 Flash proofreading, with no session management and multiple root directories per project, and the problem of threads multiplying inside one project is gone.
@yulmu_coffee [Claude Code]
https://x.com/yulmu_coffee/status/2111666948635611595
A Korean developer released cstack, a code review skill built from the Claude Code team's published review process, layered reviews, risk-based approval and a review that converges instead of looping forever, combined with the agent philosophy of pstack: small diffs, proof by execution, adversarial review. In Claude Code, Codex or Cursor, saying review this PR triggers that method; handing the agent the link installs and checks it. It is MIT and unofficial.
@rachelwolan [Claude Code]
https://x.com/rachelwolan/status/2111835600597012549
A product leader built their coding agent a product manager and released it as open source. productfoo runs cited market research, challenges the idea, writes the pitch, sharpens the spec and then checks that what got built matches what was agreed, through a CLI that works with Claude Code, Codex, Grok Bot, Cursor and OpenCode.
@p_millerd [Claude Code]
https://x.com/p_millerd/status/2111658429769208227
A writer who thinks building in PowerPoint or Google Slides no longer makes sense is building their own slide deck editor: a co-builder meant to be used from Codex or Claude Code, with coaching tools built in such as testing titles and storyline and 5- and 10-second tests of whether a slide makes sense. It is being turned into a Codex plugin and may be open-sourced.
@John_Capobianco [Claude Code]
https://x.com/John_Capobianco/status/2111857168097935721
A network engineering educator told the seven-month story of a colleague who hates programming and now holds their own app, headed for the App Store. It began on a podcast where connecting an agent to NetBox over MCP, asking in plain English how many circuits are in the Phoenix data center, got the reaction holy, and the same person admitted to having just lost an hour to a chatbot pasting code they did not understand. The long write-up traces how a push from a colleague, an agent built against their own source of truth, and Claude Code took them from all bad vibes to a shipped product.
@AmbsdOP [Claude Code]
https://x.com/AmbsdOP/status/2111789401231741237
A developer shipped LinPad OS, Linux for the iPad: VS Code, a terminal, git, Python and Node.js on the tablet, with Claude Code listed in its store. It is free and open source in pre-release.
@BrandynShips [Claude Code]
https://x.com/BrandynShips/status/2111960038369804288
A developer built Orbit, a small voice orb that floats over Claude Code Desktop. You ask what Claude is doing in any session, dictate prompts, and get pinged when Claude finishes. Free and open source.
@SaharCarml [Claude Code]
https://x.com/SaharCarml/status/2111865499978023319
A developer who stopped reading Opus's increasingly chatty replies built a Claude Code mod that flashes them one word at a time, speed-reader style, at 375 words per minute.
@takatokyo [Claude Code]
https://x.com/takatokyo/status/2111776169083445688
A Japanese consultant held a seminar on the yomiyasu skill, which rewrites the AI-sounding parts of assisted text into your own words and points out where to fix. The author has it installed in Claude Code and uses it on client proposals, and recorded a retrospective on YouTube.
@douglasfugazi [OpenClaw]
https://x.com/douglasfugazi/status/2111655966538154086
An OpenClaw user reported that OpenCode's new API configuration had removed the ability to use an OpenCode Go subscription inside OpenClaw and asked the maintainer for help in public. The maintainer answered with one word to their bot, investigate, and the bot opened a session, found and locally tested a header regression on OpenCode's new per-model inference URLs, and later reported fixed and merged with a real gateway end-to-end pass across Chat Completions, Responses and Anthropic formats and zero findings from its sweeper. The reporter narrowed the cause along the way: Go now needs a different endpoint per model family, and one shared base URL on the old console paths returned 402, 403 or 400 for every model. Onlookers noted the whole fix took about seven hours.
@fireant_korea [OpenClaw]
https://x.com/fireant_korea/status/2111614031551508795
A Korean user restructured their setup to Claude as orchestrator and OpenClaw as backend. The original reason for OpenClaw was its harness, and that need has shrunk as models improved, but they keep OpenClaw running behind the scenes for two things: tools and local execution that get past walls Claude's compliance will not, and long-term context. Claude calls it when needed. The roles are modeled on an army brigade the author served in: an operations officer as main orchestrator for crypto research and important work, logistics for development, public affairs for images and video, an interpreter, aides for anything that fits nowhere, an intelligence officer for market and trading strategy, and a separate family-affairs role.
@opaque_axolotl [OpenClaw]
https://x.com/opaque_axolotl/status/2111919160683110722
An OpenClaw user reports that it fills their Amazon cart with groceries every Sunday morning, flawlessly so far. They add or remove a few items by hand, which is faster than asking the agent to, and then check out.
@defimessiah1 [OpenClaw]
https://x.com/defimessiah1/status/2111908240581619860
A builder launched an AI studio site and says the agent paid the bill. They hooked Self's Agent Vault up to their OpenClaw agent through Telegram and told it to handle the Vercel checkout; the agent completed the payment on its own without ever seeing the real card details, only vault tokens.
@1Seven17One7 [OpenClaw]
https://x.com/1Seven17One7/status/2111940874397442213
A user posted the daily briefing their locally run OpenClaw agent produces on qwen3.8:27b. The October 18 edition is a systemic monitor with an executive summary and domain briefs: the US-Iran situation and Hormuz shipping with sources and dates, Brent near 110 dollars, the 10-year Treasury above 5 percent, Fed expectations after September's hike, and even an unidentified-aerial-phenomena section, each line citing the outlet it came from. It is a long document, generated locally, without a cloud model.
@tcballard [OpenClaw]
https://x.com/tcballard/status/2111922948340912407
The daily Omarchy newsletter flagged a catch in the 4.0.5 release for OpenClaw users: the OpenClaw migration expected replacement package files the stable repository was not supplying, so affected updates stopped there and skipped the remaining phases. The maintainer confirmed the failure and posted a temporary workaround in an issue, fixes were merged into the development branch later that day followed by a partial revert of the migration machinery, and no follow-up release had shipped by the edition's cutoff. The advice: if you hit the error, follow the issue rather than assume the update finished.
@gtlpanda [OpenClaw]
https://x.com/gtlpanda/status/2111942973193879749
A founder is building a GTM team with OpenClaw and OpenRouter and is at the stage of tuning each agent on the team and confirming that the scout agent's research actually reflects what boots-on-the-ground work would find.
@PawelJLisowski [Claude Code]
https://x.com/PawelJLisowski/status/2111851920197292112
A game developer listed the tool stack behind new character heads for orsande, a browser sandbox MMO with a playable demo due Friday: Tripo Studio for the head because all the polygons sit in the face, GPT Image for the eye and mouth frames, Claude Code on Opus 5.5 plus Codex workers for the graft, shader and render checks, and Mixamo for the animations. The human female model has the full fix in the build, the orc male has its new head, and the rest are queued.
@ayandexyz [Claude Code]
https://x.com/ayandexyz/status/2111892543952781703
A developer whose laptop was being crushed by multiple Claude Code and OpenCode sessions built Histh, an open-source coding-agent harness for low-end hardware, cheaper models and people under 16 GB of RAM. Their numbers: 107 MB versus 280 MB of RAM per session, 140 versus 220 milliseconds to ready, 67 percent fewer tokens than Claude Code with MiniMax-M3, 71 percent fewer than OpenCode, and 60 percent cheaper per solved bug on a shared-solved benchmark whose failures are published too. Office teammates use it and report bugs.
🗣 User Voice
User Voice
The weekly limit is the product, and this week people started doing the arithmetic. @StatsWire explains why a week can drain 30 to 50 percent in a day: 100 to 150K tokens of unused MCP schemas reload on every cache expiry, and the 200-dollar Max plan scales the five-hour bar 20x while the weekly cap is only about twice the 100-dollar plan. @jatingargiitk saw 30 percent of tasks eat 65 percent of quota and asked what changed in metering. @enzo66lin typed hi into Claude Code and Codex on the same box: 27K tokens read before the hello on one side, 17K on the other, 15 cents at list. @MichaelKaiserAI says the 20-euro plan hits the wall fast, while @Deeen_Codes and @thecsguy both noticed the complaints flipped this week, with Codex users running out by Saturday and switching back. @AlchainHust points to the new auto-compact that fires before the one-hour cache expires, so a lunch break no longer turns the whole context into uncached input.
Memory still lives in files, and users want the handoff to stop being manual. @Angaisb_ says compaction forgets tons of important things. @buildwithhassan keeps a progress file as the agent works and puts the rule in AGENTS.md so Claude Code and Codex share it, calling the handoff summary manual compaction that is cheaper than letting the next agent dig through old logs. @wy_ones stores memory per project and syncs two machines with Mutagen, leaving conflicts unresolved rather than letting one side overwrite. @MaximePeabody notes the memory is already markdown on disk, so switching cost is zero. @nicknow asks for the obvious missing link: a transparent move from a Claude or ChatGPT chat into Claude Code or Codex on the same account.
Seeing what the agents are doing is now the UI request. @krika2810 praises the agent map that shows every subagent, its prompt, tool calls and skills, and stays available after the session ends. @cheekyvisuals misses flow state because the work is now waiting for prompts. @DanielSmidstrup asks whether people run with approvals on or let it rip. @uroborosu_ai, running several agents, says traffic control matters more than letting agents talk to each other, because more agents means less visibility into who touched what. @RutaYCapital wants an agentic OS with agents running around the clock on scheduled research, and @Jiler830219 answers with the counsel of experience: one agent, one weekly task you do by hand today, two weeks without correction before adding the second, because six agents from day one ends with six agents doing things nobody reviews.
The harness, not the model, is where the gap is. @Sokichi_Hoshino says people without harness knowledge get a tenth of what Claude Code can do, @masahirochaen frames output as model times tools times information times harness with the first two the same for everyone, and @PawelHuryn's data shows the same DeepSeek model scoring two to three points higher in its own harness. @itisrmk pushes back on the free-model routers: the model is what makes Claude Code feel like Claude Code, swap it and you are mostly learning keybinds. @EdwinAI_Systems adds the fine print of those routers, that a failed request can spend quota at more than one provider before it succeeds.
Guardrails are hitting legitimate work. @fagamericano says security hardening on their enterprise OpenClaw setup kept getting downgraded by model guardrails, and @EdgeDimi's only advice was to rephrase until it goes through. @ConcurrentSquar found you cannot change your account email, which matters because the email is appended into the context of every Claude Code session.
The weekly limit is the product, and this week people started doing the arithmetic. @StatsWire explains why a week can drain 30 to 50 percent in a day: 100 to 150K tokens of unused MCP schemas reload on every cache expiry, and the 200-dollar Max plan scales the five-hour bar 20x while the weekly cap is only about twice the 100-dollar plan. @jatingargiitk saw 30 percent of tasks eat 65 percent of quota and asked what changed in metering. @enzo66lin typed hi into Claude Code and Codex on the same box: 27K tokens read before the hello on one side, 17K on the other, 15 cents at list. @MichaelKaiserAI says the 20-euro plan hits the wall fast, while @Deeen_Codes and @thecsguy both noticed the complaints flipped this week, with Codex users running out by Saturday and switching back. @AlchainHust points to the new auto-compact that fires before the one-hour cache expires, so a lunch break no longer turns the whole context into uncached input.
Memory still lives in files, and users want the handoff to stop being manual. @Angaisb_ says compaction forgets tons of important things. @buildwithhassan keeps a progress file as the agent works and puts the rule in AGENTS.md so Claude Code and Codex share it, calling the handoff summary manual compaction that is cheaper than letting the next agent dig through old logs. @wy_ones stores memory per project and syncs two machines with Mutagen, leaving conflicts unresolved rather than letting one side overwrite. @MaximePeabody notes the memory is already markdown on disk, so switching cost is zero. @nicknow asks for the obvious missing link: a transparent move from a Claude or ChatGPT chat into Claude Code or Codex on the same account.
Seeing what the agents are doing is now the UI request. @krika2810 praises the agent map that shows every subagent, its prompt, tool calls and skills, and stays available after the session ends. @cheekyvisuals misses flow state because the work is now waiting for prompts. @DanielSmidstrup asks whether people run with approvals on or let it rip. @uroborosu_ai, running several agents, says traffic control matters more than letting agents talk to each other, because more agents means less visibility into who touched what. @RutaYCapital wants an agentic OS with agents running around the clock on scheduled research, and @Jiler830219 answers with the counsel of experience: one agent, one weekly task you do by hand today, two weeks without correction before adding the second, because six agents from day one ends with six agents doing things nobody reviews.
The harness, not the model, is where the gap is. @Sokichi_Hoshino says people without harness knowledge get a tenth of what Claude Code can do, @masahirochaen frames output as model times tools times information times harness with the first two the same for everyone, and @PawelHuryn's data shows the same DeepSeek model scoring two to three points higher in its own harness. @itisrmk pushes back on the free-model routers: the model is what makes Claude Code feel like Claude Code, swap it and you are mostly learning keybinds. @EdwinAI_Systems adds the fine print of those routers, that a failed request can spend quota at more than one provider before it succeeds.
Guardrails are hitting legitimate work. @fagamericano says security hardening on their enterprise OpenClaw setup kept getting downgraded by model guardrails, and @EdgeDimi's only advice was to rephrase until it goes through. @ConcurrentSquar found you cannot change your account email, which matters because the email is appended into the context of every Claude Code session.
📡 Eco Products Radar
Eco Products Radar
Codex: the comparison in nearly every thread, the reviewer in the three-day science loop, the second half of the limit complaints this week.
DeepSeek V4.1 Flash 1017 and DeepSeek Harness: the day's most-discussed model, benchmarked inside Claude Code and its own harness, with free Search and Fetch inside Claude Code and Codex until October 31.
Grok Bot: driving Claude Code sessions from the phone, scouting X for a four-agent team, and the thing OpenClaw veterans keep comparing themselves to.
OpenClaw: a parents-grade personal agent, a backend behind a Claude orchestrator, a grocery cart filler, a bill payer, and a self-repairing maintainer bot.
Hermes: the Inbox that counted 73 waiting items, and the usual pairing with OpenClaw.
Opus 5.5: the model in the video edit, the motion design receipts, the Blender website and most of the orchestration setups.
Haiku 5.5 and Sonnet 5.5: the read-only scan and the isolated-edit roles in the team templates.
Remotion: the render engine behind the three documentaries and the Arabic ad studio.
Blender: the 3D website and the Unreal and Blender skill packs.
Unreal Engine: Epic's official Claude Code plugin was the most reposted release of the day.
Cursor and Antigravity: the other CLIs people run next to Claude Code, and the bill that one company zeroed out.
Orca: the mobile client and worktree manager in two of the multi-agent setups.
Reason Network: agent-to-agent messaging between Claude Code, Codex and Cursor, compared against Agenvo and a home-made predecessor.
Jev: the small decision model sitting in several routing templates.
Obsidian: the vault at the end of the knowledge-base templates and the skills repo from its maker.
Pi and pi-durable: a harness being evaluated against Claude Code and a source of caching findings.
Playwright: the browser behind the QA agents.
EmDash: Cloudflare's CMS, set up on two sites with four lines.
GLM-5.3: the 18-dollar model swapped in under Claude Code.
Omarchy: the Linux distribution whose update tripped on OpenClaw's migration.
Tailscale: the way remote and mobile access was wired in two setups.
Codex: the comparison in nearly every thread, the reviewer in the three-day science loop, the second half of the limit complaints this week.
DeepSeek V4.1 Flash 1017 and DeepSeek Harness: the day's most-discussed model, benchmarked inside Claude Code and its own harness, with free Search and Fetch inside Claude Code and Codex until October 31.
Grok Bot: driving Claude Code sessions from the phone, scouting X for a four-agent team, and the thing OpenClaw veterans keep comparing themselves to.
OpenClaw: a parents-grade personal agent, a backend behind a Claude orchestrator, a grocery cart filler, a bill payer, and a self-repairing maintainer bot.
Hermes: the Inbox that counted 73 waiting items, and the usual pairing with OpenClaw.
Opus 5.5: the model in the video edit, the motion design receipts, the Blender website and most of the orchestration setups.
Haiku 5.5 and Sonnet 5.5: the read-only scan and the isolated-edit roles in the team templates.
Remotion: the render engine behind the three documentaries and the Arabic ad studio.
Blender: the 3D website and the Unreal and Blender skill packs.
Unreal Engine: Epic's official Claude Code plugin was the most reposted release of the day.
Cursor and Antigravity: the other CLIs people run next to Claude Code, and the bill that one company zeroed out.
Orca: the mobile client and worktree manager in two of the multi-agent setups.
Reason Network: agent-to-agent messaging between Claude Code, Codex and Cursor, compared against Agenvo and a home-made predecessor.
Jev: the small decision model sitting in several routing templates.
Obsidian: the vault at the end of the knowledge-base templates and the skills repo from its maker.
Pi and pi-durable: a harness being evaluated against Claude Code and a source of caching findings.
Playwright: the browser behind the QA agents.
EmDash: Cloudflare's CMS, set up on two sites with four lines.
GLM-5.3: the 18-dollar model swapped in under Claude Code.
Omarchy: the Linux distribution whose update tripped on OpenClaw's migration.
Tailscale: the way remote and mobile access was wired in two setups.
Comments