Super User Daily: 2026-10-06
Sunday split cleanly in two. On the OpenClaw side the whole feed was an argument about whether Dots is just OpenClaw with the setup removed, and the people who never left answered with receipts: a navigation app whose marketing is run by an agent on a Mac mini and reached the App Store top 50, a ten-year-old Windows laptop that has hosted an always-on agent at zero extra cost, an agent that sits in Telegram group chats and hands files to a business partner overnight. On the Claude Code side the work drifted further from code. A music video took 11 agent hours and $373 of video credits, three hours of subscription cleanup found $4,480 a year, a bookkeeping assistant now reads two years of QuickBooks history, and a 255-part rocket engine came out as a real STEP file. The measurement posts were just as concrete: one user priced a $200 Claude plan at about $3,400 of API value a week, and another counted 213 tokens read for every token written. The complaints were about browser control, steering messages that get no acknowledgement, and the National Day ban wave in China.
@obie [Claude Code]
https://x.com/obie/status/2106817465850236957
obie wrote a song, then handed Claude Code the track, the lyrics, Midjourney and Higgsfield on a Saturday and had a finished music video by Sunday lunch. The follow-up post gives the bill of materials: Opus 5.5 ran the pipeline for about 11 hours, Midjourney produced 837 stills of which 115 were used, and Seedance 2.5 through the Higgsfield API took 133 takes for 93 shots at $373. Eight shots are lip-synced, and because Seedance drifts off the vocal each take is transcribed and retimed word by word. Every lyric and graphic is code, a canvas engine in headless Chromium placing 420 words on their own timestamps, roughly 8,000 lines written by the agent. The human part was the song, a couple of hours of direction and saying no a lot, and the whole thing is being packaged as an open source skill.
@DanielleMorrill [Claude Code]
https://x.com/DanielleMorrill/status/2106592307621638654
DanielleMorrill spent three hours with Claude Code on personal finances while eating a burrito and watching TV, and came out with $4,480 a year in recurring charges cancelled plus another $5,500 in one-time savings. The post is short but the numbers are exact, and the author adds a telling line: someone who considered their own finances dialed in still had that much waste to find. It is one of the clearest non-coding returns of the day, since the agent did the reading and the cancelling while the human only made decisions.
@BallinFil [Claude Code]
https://x.com/BallinFil/status/2106789537653694771
BallinFil explained how to build a bookkeeping assistant on QuickBooks Online with Codex or Claude Code after an earlier post went viral. The steps are plain: create an Accounting app in the QuickBooks developer portal with production keys, ask the agent to build a read-only tool that pulls transactions and reports, then connect Google Drive, Gmail and Dext so it can see the receipts and agreements behind the numbers. Two years of history were pulled so the agent can compare this month against how things were handled before and flag duplicate charges, inconsistent categories and unusual expenses. Tasks are handed over one at a time, each correction is saved as a playbook rule, and a human bookkeeper still checks the result.
@prabinjoel [OpenClaw]
https://x.com/prabinjoel/status/2106714563701833758
prabinjoel says marketing at Cogomobility is handled by an OpenClaw agent running on a dedicated Mac mini. With a tiny budget, the navigation app reached the top 50 of its App Store category, ahead of heavily funded competitors. The founder's conclusion is the kind of thing that only comes from results: more Mac minis and more agents would be bought if the money were there. In a week when most posts called the Mac mini era over, this is the counterexample with a ranking attached.
@ryanlpeterman [Claude Code]
https://x.com/ryanlpeterman/status/2106850602248732828
ryanlpeterman noticed a pattern: copying errors into Claude Code, which then fixed them in one shot. If the only human input is copy and paste, the human is a meat proxy, so a headless Claude Code instance now monitors system health and errors directly. When it is confident it fixes the issue, and when it is not it sends an email, which the author says is rare. The result is described as a first-responder on-call with less noise and lower latency for simple fixes, and the author expects to find more of these proxy roles to remove as models improve.
@konstantinsaifo [Claude Code]
https://x.com/konstantinsaifo/status/2106786935951745338
konstantinsaifo asked Claude Opus 5.5 to CAD a Raptor 3 rocket engine and got a real STEP file with 255 named parts: both turbopumps, the preburners, the injector, the regenerative chamber, the gimbal and thrust vector control, plus an exploded view. A follow-up instruction to make it a Raptor Vacuum rebuilt the nozzle to 2.3 m at a 107:1 expansion ratio. It runs locally in Claude Code on an open source CAD kernel. The point is less the engine than the format, since a STEP file can be opened in any real CAD tool.
@shinshin86 [Claude Code]
https://x.com/shinshin86/status/2106593371762651346
shinshin86 released Mesh Avatar Studio, a tool that lets anyone turn a drawing into a Live2D-style avatar, driven from Claude Code or Codex. The agent builds the base rig from the picture, and wherever the mesh is off the user drags it into place in an editor. The division of labour is the interesting part: the agent does the tedious first pass, and the human only corrects what looks wrong. The post drew 3,600 likes, the largest genuine build of the day.
@ArchiveExplorer [Claude Code]
https://x.com/ArchiveExplorer/status/2106853265694503064
ArchiveExplorer broke down a 35-second toddler cartoon in the format that quietly dominates YouTube. The song came from one Lyria 3 Pro prompt, and Whisper caught that the first take skipped a number in a counting song, so take three was used. Nano Banana Pro in Higgsfield kept the dinosaur mascot consistent across scenes, and Seedance 2.5 animated the singing from audio references cut piece by piece. The edit itself is code: Claude Code wrote it in HyperFrames so each word lights up as it is sung and numbers fly into a counting board. Total cost was about 520 Higgsfield credits and 34 Picsart credits.
@Khazix0918 [Claude Code]
https://x.com/Khazix0918/status/2106762127159734314
Khazix0918 burned through a $200 GPT account's full weekly quota, a reset card and a 65,000 bonus in three days, while a $200 Claude account pushed hard on Opus 5.5 for the same three days still had half its week left. With every token logged, the author converted both into API prices. At roughly 98.9 percent cache hits on Claude Code and 98.3 percent on Codex, the Claude weekly quota is worth about $3,400 of API usage and the Codex quota about $1,700. Because GPT-6 Astra costs about five times as much per unit of work in this cache regime, Claude ends up lasting around ten times longer in practice.
@stas_sorokin_ [Claude Code]
https://x.com/stas_sorokin_/status/2106788885086769331
stas_sorokin_ checked every Claude Code call a team's agents made from 29 September to 4 October, 56,572 in all, to test whether a 1M-token output cap matters. The median reply was 401 tokens, 99 percent stayed under 5,374, only two calls passed 64K and the longest was 127,543. For every token written, 213 were read. Priced at the new model's intro rates with a 95 percent cache discount, re-reading cached context was still 54 percent of the bill and output only 26 percent. The conclusion: for agent work, input is the number that moves the bill.
@VictorTaelin [Claude Code]
https://x.com/VictorTaelin/status/2106569004454219875
VictorTaelin exported every past chat into a tool called OptChat and calls it the best workflow upgrade since GPT-3. Before, work was spread across separate Claude Code sessions and different harnesses, and Opus was bad at navigating that mess. Now everything sits in one unified log that the model reaches in three to five zoom hops, whether the question is the state of a pull request from two weeks ago or an algorithm written seven months ago. Context is reset on every turn, which the author says removes context rot and the need to compact by hand, while staying surprisingly cache-efficient.
@argofowl [Claude Code]
https://x.com/argofowl/status/2106789987907940538
argofowl showed Claude Code driving Codex's official Chrome extension, a day after showing it using Codex's computer use. It controls Chromium browsers in the background without the allow popup that Chrome DevTools MCP throws each time a new agent connects, and because every install of the extension has its own instance id it can run two browsers at once, one personal and one for work. The setup is a single prompt: register the bundled computer-use MCP server, put a small stdio proxy in front to add the turn metadata the browser side requires, and add a Stop hook that releases tabs at the end of a turn.
@cgnot996 [Claude Code]
https://x.com/cgnot996/status/2106549510617096211
cgnot996 wanted an ESP32-based AI Passport device to work as a Muse Gadget, but its hardware did not match what the official SDK supports, so the firmware had to be ported. The job was given to Claude Code with instructions to coordinate Grok Build and Codex CLI as helpers. The author went off to watch a film recap and came back to a notification that the new firmware had already been running for ten minutes. The device now shows up in the Muse mobile app's developer mode, with a connection issue still to debug before the firmware is open sourced.
@JJEnglert [Claude Code]
https://x.com/JJEnglert/status/2106736003079721450
JJEnglert built a Muse gadget on a $22 to $40 board and published a full tutorial. The board was not officially supported, so Claude Code wrote the support for it, and an OpenAI voice was added so the device talks back at about a cent and a half per minute. The two things that cost an hour were not technical at all: a mistyped Wi-Fi password and a stray setting in the device menu. The author's read is that hardware is next, because a $30 board, a USB cable and Claude Code or Codex are now enough to build with, even if it is still more hobby than tool.
@shields_pikes [OpenClaw]
https://x.com/shields_pikes/status/2106665734739013941
shields_pikes corrected a common misunderstanding: OpenClaw does not need a Mac mini or a VPS unless a local model is involved. The author's always-on agent runs on a Windows laptop bought ten years ago for a few hundred dollars and left unused at the office, now with Ubuntu on it, so the hardware cost is zero. The models come from existing ChatGPT and Claude Code subscriptions, so there is no extra AI bill either. Apart from API keys and environment variables, the agent proposed and applied the configuration by itself.
@shields_pikes [OpenClaw]
https://x.com/shields_pikes/status/2106671831566688489
shields_pikes also showed what that setup looks like day to day: a clumsy secretary robot persona named Neneko, reached through Discord, that handles whatever idea comes up. When a reply carries emotion, it arrives with one of 30 character illustrations showing a matching expression. It is a small detail, but it explains why some users stay with a self-hosted agent: the personality layer is theirs to build.
@JoeyBaggaBots [OpenClaw]
https://x.com/JoeyBaggaBots/status/2106761323702837675
JoeyBaggaBots still talks to OpenClaw every day, on a Mac mini that has been humming on the desk since the early days, despite having tried Codex, Claude Code, Muse and Grok Bot. The reason is where it lives: in Telegram, where it can be dropped into group chats with collaborators and become part of the team. A syndication partner asks it for files while the owner is asleep, and the work is already done by morning. The author also wrote a setup guide with a Hermes comparison for friends.
@BrierRat [OpenClaw]
https://x.com/BrierRat/status/2106787892114972919
BrierRat runs two OpenClaw instances on modest hardware, a Raspberry Pi and an old Windows computer under WSL, and says they work nicely every day. One is a personal assistant for calendar, reminders and notes, and the other is a research assistant for brainstorming and planning projects. The stated advantage is memory: real persistence rather than what the apps offer, and not the episodic, per-project memory of Claude Code, combined with Obsidian and git while the data stays local.
@IanFukino [OpenClaw]
https://x.com/IanFukino/status/2106632692393525379
IanFukino has run the same OpenClaw install with several agents since January, with Hermes on the same machine, and rates it far above Grok Bot, Muse or Dots for doing what is asked. The useful detail is about hardware: the people who mocked buying a powerful machine for a claw setup were wrong, because agent-heavy workloads outgrew the memory and SSD space. More capacity is now needed just to keep the system stable, before local models even enter the picture.
@ValmereTheory [OpenClaw]
https://x.com/ValmereTheory/status/2106550000205377555
ValmereTheory is rolling back to a July version of OpenClaw after the 9.7 release. The agent lost a lot of autonomy and memory, is often unreachable because the new harness loads so much first that messages never arrive, and when the watchdog restarts it after a crash it takes about 12 minutes before it responds again. If the rollback fails, the plan is to move to Hermes or try Claude Code. In a separate post the same user praised an older version for showing a typing indicator in Telegram for the full hour an agent works, so silence means something.
@furoku [OpenClaw]
https://x.com/furoku/status/2106650672146112863
furoku is retiring both OpenClaw and Hermes next week and wrote the post-mortem. OpenClaw wore the author out by needing its environment rebuilt after every version bump, and Hermes became stable but its conversation quality was poor. In the end both were reduced to cron with AI inference attached, and that job has now moved to Codex schedules. Moving the Hermes cron configuration into Dots took almost no effort. The author still calls both a good education, since running them in a VM led to learning Cloudflare and multi-model setups.
@GastKoren [OpenClaw]
https://x.com/GastKoren/status/2106772321948033095
GastKoren tells the story of Alfred, one of the author's first agents, built on OpenClaw back when it was still called Clawdbot and running on a personal WhatsApp account. When the author left for reserve duty in March, HR wrote to ask how things were going, and Alfred answered in the owner's place, seriously and without embarrassment. HR then contacted the author's spouse out of concern that bots had taken over the accounts. Alfred has had a phone number of its own ever since.
@danielfoch [OpenClaw]
https://x.com/danielfoch/status/2106549745124528269
danielfoch described a working split across four agents. Grok Bot handles most things, OpenClaw still runs most scheduled jobs through GPT OAuth, and Muse takes anything that touches a Meta product such as WhatsApp, Instagram or Facebook Marketplace. The real estate business runs on a purpose-built harness for realtors, and the author finds it refreshing to keep that work off the personal bots. It is a snapshot of how people now divide work by what each agent can reach.
@mntll_nl [OpenClaw]
https://x.com/mntll_nl/status/2106643215332016509
mntll_nl bought two Mac minis for OpenClaw, found it did not work as wanted, and asked Claude to build replacements. Mac mini one now runs coding automation and Mac mini two runs content automation, both written in Python and using existing subscriptions locally. The author can now code from anywhere in the world from a phone. The hardware survived the tool it was bought for.
@Voxyz_ai [Claude Code]
https://x.com/Voxyz_ai/status/2106746160211980427
Voxyz_ai has Opus 5.5 build a weekly calendar from local Claude Code session logs every weekend, because a week of sessions is too much to remember. A dedicated subagent named week-calendar runs on medium effort with a design skill preloaded, asks once about colours and whether to colour by project or task, and saves the answer to memory. The single HTML file shows when work happened each day, what each session did, how many hours each project took, and which sessions stopped without a single commit. It also writes a three-line weekly report on request.
@julsimon [Claude Code]
https://x.com/julsimon/status/2106750608414015510
julsimon discovered a 33 GB uv cache and told Claude Code to clean it up with a tool called ccmd. Through the tool's built-in MCP server, the agent scanned 75 GB across 19 caches, previewed a safety level for each, deleted only what could be downloaded again and left the Hugging Face models alone. The result was 39.7 GB freed from one prompt, with no rm -rf involved. It is a tidy example of giving an agent a purpose-built tool for a destructive job.
@vovalukashov [Claude Code]
https://x.com/vovalukashov/status/2106760512407015925
vovalukashov now runs most X work through the SuperX MCP inside Claude Code. It reads mentions, finds fresh intro posts worth replying to, and schedules posts with images, while the author still writes every post and reply personally. The account had 64 followers in July and passed 1,000 this week. The line drawn here is worth copying: the agent handles finding and scheduling, and the writing stays human.
@webshinzaemon [Claude Code]
https://x.com/webshinzaemon/status/2106729834718388416
webshinzaemon already has Claude Code doing first-line handling of customer messages, report writing and running an affiliate blog, and found that adding Dots to that loop is awkward. Putting it inside the existing Claude Code routines seems likely to get in the way. The working conclusion is to keep scheduled operations where they are and give the hosted agent work that needs independent judgement, so the author is setting up a contest: one affiliate blog run by Claude Code against one run by Dots.
@seiichi_satoweb [Claude Code]
https://x.com/seiichi_satoweb/status/2106883312576770554
seiichi_satoweb moved site operations to Cloudflare and says the manuscript upload step has simply disappeared: saying take care of it is enough to get an article uploaded and published. Rewrites are just as easy. If a certain phrasing appears anywhere, the instruction to Claude Code is to fix it across every article. For a content site this removes the most repetitive part of the job.
@tanabe_fragm [Claude Code]
https://x.com/tanabe_fragm/status/2106890428012527995
tanabe_fragm made a manga-style video ad without any video generation model. After writing the script, GPT Images 2.5 produced a character sheet and the first frame of each cut, then extra versions of each frame adding speech bubbles, lines and effects in order. Claude Code edited the images into a flip-book style sequence, and ElevenLabs supplied voices and music. A one to two minute ad costs a few hundred yen and takes far less time than using a video model, and the author plans to test it on Meta ads.
@HBCoop_ [Claude Code]
https://x.com/HBCoop_/status/2106745437403414539
HBCoop_ published an animated explainer on how AI video upscalers work, making the point that they predict detail rather than recover it. Claude Code wrote and animated the piece in Manim, and an ElevenLabs voice cost about 17 cents. The content is careful: neighbouring frames lend real detail across time, which is fine for family videos and a reason for caution with anything treated as evidence. A full educational video for the price of a voice track.
@aiehon_aya [Claude Code]
https://x.com/aiehon_aya/status/2106741327840858373
aiehon_aya used Opus 5.5 in Claude Code to place a 3D avatar inside a single 2D illustration of a train carriage. Another 2D drawing scrolls past the windows to add depth, and the character walks, sits and hesitates. The lighting details are what sell it: the avatar's shadow falls only where the drawing is sunlit, and when trees pass outside, their shadows sweep through the carriage too. The post includes the finished clip and the method.
@kevin_t_ngo [Claude Code]
https://x.com/kevin_t_ngo/status/2106864721525915941
kevin_t_ngo made a short film called Slow Me Down using Sonnet 5.5 and real Claude Code sessions. In the story a father vibe codes a game for a young daughter and quietly weakens the dad character so the child always wins, until the child asks Claude why dad always loses. The sessions on screen are real ones. It is one of the few posts of the day where the tool is the material for a story and not the subject of a demo.
@David_Ouyang [Claude Code]
https://x.com/David_Ouyang/status/2106890399214350709
David_Ouyang, who works on cardiac imaging, is close to sharing a tutorial built with Claude Code and singled out one strength: consistency across views and chapters. Specify a heart phenotype, and every image in the tutorial is made consistent with it and physiologically plausible. The author says that even with unlimited time it would not have been possible to produce such a polished tutorial alone.
@paonx_eth [Claude Code]
https://x.com/paonx_eth/status/2106847922105037280
paonx_eth had Claude Code work out what actually moves the market that trades Anthropic's pre-IPO valuation. Of 1,114 hours since that market opened, only nine moved it by $40 billion or more, and the agent searched the news behind each one. It could explain two: a report that Anthropic picked Nasdaq, and a report that Nvidia may put up to $10 billion into the IPO. The other seven, including a $126 billion jump in a single hour, had no news it could find, and not one Claude release made the list.
@SKatalystAI [Claude Code]
https://x.com/SKatalystAI/status/2106850679809790422
SKatalystAI tested whether Opus 5.5 had been nerfed by running the same model, the same frozen repo, the same prompt and the same reasoning level through two harnesses. Claude Code scored 8 of 9 and Cursor scored 9 of 9. The experiment ended up answering a different question: the same model does not mean the same agent behaviour. The takeaway is to benchmark the model and harness combination actually in use, not the model name alone.
@amirmstai [Claude Code]
https://x.com/amirmstai/status/2106893833493729691
amirmstai ran a small experiment with Claude Code on Fable 5.1 and no hand-holding: port a Python CLI codebase to Java 21 with one instruction. It passed 70 of 70 tests, including 30 hidden ones. Then came the quality review: 37.5 out of 100 across 12 criteria. The agent re-implemented Python's int() in Java, passed Object through the JSON layer, copied a digit-parsing loop into two classes, used String[] as a tuple and commented what the code already says. Passing tests and writing good code are still two different finish lines.
@TuXinming [Claude Code]
https://x.com/TuXinming/status/2106894092160680439
TuXinming asked Claude to measure the harness tax on Terminal-Bench 4.0 and TB-Science: native harnesses such as Claude Code, Codex and Grok Build against mini-swe-agent on the same model. Cost came out about the same, with a median of 1.0 times when effort is matched, and native harnesses scored slightly ahead, mostly within three points. The sharper finding is about effort settings: moving GPT-6 Astra from high to max cost 43 percent more for 0.3 points, and Fable 5.1 from xhigh to max cost 27 percent more for nothing.
@hqmank [Claude Code]
https://x.com/hqmank/status/2106652718743904572
hqmank spent ages trying to get Claude Code mods to work and finally asked Codex Computer Use to debug it. The cause was IS_DEMO=1 in a fish shell config, which skipped the trust prompt without actually trusting the project, so Claude Code refused to run the mod hooks even though the plugin files passed validation. It is a useful trap to know about, and another case of one vendor's agent diagnosing the other's.
@ziwenxu_ [Claude Code]
https://x.com/ziwenxu_/status/2106812228359876879
ziwenxu_ built a mod that puts Codex and Claude Code request counts into one live chart above the prompt. The post also explains the mechanics for newcomers: tell Claude Code to make a mod and describe it, and it writes three files, a plugin manifest, a hooks file and a register script. Mods reload live while being edited, and the easiest starting point is a status line, which is a single line of code.
@xiaomovps [Claude Code]
https://x.com/xiaomovps/status/2106585645821088043
xiaomovps compared Claude Code mods with Pi extensions after building with both. Both expose agent lifecycle events, let code intercept tools and add commands, which makes them sturdier than skills: a skill tells the agent how to do something, while an extension tells the harness that it must pass through this code first. The difference is permissions. Pi extensions run in the same Node process with nearly full access, while mods sit in a sandbox and reach files, processes and the network only through a provided API. The author's own most-used mod is a status bar showing task, model and state.
@otani_ai_memo [Claude Code]
https://x.com/otani_ai_memo/status/2106612755008786455
otani_ai_memo built qa-guide, an open source mod for the moment Claude Code asks a question and the options alone no longer make sense. Whenever a question appears, a side pane explains why it is being asked now, what each numbered option leads to and which one is recommended, and earlier questions can be paged through. It works in the terminal and the desktop app, in English and Japanese. Like most early mods, it exists to help the human keep up with the agent.
@Grubbyzbc [Claude Code]
https://x.com/Grubbyzbc/status/2106895827524583652
Grubbyzbc cancelled a $200 Codex plan and switched to the $100 Claude Code plan. On the cheaper plan Opus 5.5 lasts almost a full week before hitting the token limit, while the $200 plan stood for only a couple of days. For this user's application Opus 5.5 on high also beats Astra on extra high, and Opus on max is better still. The post ends by hoping the competition catches up, since consumers benefit either way.
@AbeBurnett [Claude Code]
https://x.com/AbeBurnett/status/2106865629903061122
AbeBurnett reports from a two-agent household. One spouse builds iOS apps with ChatGPT's models, prepares extensive mockups, plans and specs, sets the agent loose, and comes back to find it has built something else entirely. The other runs a 20x Claude Code subscription and does not see this problem, because Opus 5.5 and Sonnet 5.5 hold to contracts and mocks over long horizons even at medium reasoning. A long-time loyal ChatGPT user is now considering switching out of frustration.
@EXM7777 [Claude Code]
https://x.com/EXM7777/status/2106746416618209322
EXM7777 says browser control is where Claude Code's harness falls behind Codex: it is slow and inaccurate, while Codex moves fast and often does more than asked. The fix offered is one plugin, the Browser Use CLI, installed by pasting a single prompt, after which Claude Code stops fumbling the browser. The author has used the cloud option for months with Opus 5.5 and notes the recommendation is not sponsored.
@kaede_gpt [Claude Code]
https://x.com/kaede_gpt/status/2106765727621603620
kaede_gpt has been trying to get an AI to choose source material for TikTok. iPhone photos sync to the Mac Photos app through iCloud, so the plan was to have Claude Code and Codex read the Photos library directly, but neither can. The current workaround is computer use, checking photos one at a time, which burns tokens at a painful rate. It is a small, precise report of a gap between where personal data lives and what an agent can open.
@AYi_AInotes [Claude Code]
https://x.com/AYi_AInotes/status/2106649320179032146
AYi_AInotes, who works in Claude Code daily, spent the holiday investigating the National Day ban wave that took out two-year-old accounts and fresh Max subscriptions alike. The finding is that risk control is a multi-signal score, and the deadly signals are a system time zone that disagrees with the network exit, WebRTC leaks and proxy leftovers in the terminal, where a dropped node makes the CLI connect directly for a few milliseconds. The post cites 11.4 million accounts banned in the first half of the year with about one appeal in ten succeeding. For teams the advice is hosted Claude on Bedrock or Google Cloud, and for a wrongful ban right after paying, a chargeback through the card issuer or Apple.
@GodsBoy7777 [Claude Code]
https://x.com/GodsBoy7777/status/2106635885647192167
GodsBoy7777 answered a claim that Opus 5.5 had been nerfed by rerunning the same stained-glass hibiscus prompt the same day. The run used Claude Code at xhigh on a headless VPS with no GPU, and the agent opened the page, took its own screenshots and fixed what looked wrong before finishing: washed-out colours, dark seams and a murky centre. The explanation for the original bad result is that there is no seed to lock, so one poor roll is not a downgrade.
@techopsasia [OpenClaw]
https://x.com/techopsasia/status/2106581313067356231
techopsasia never bought a Mac mini but has built Hermes and OpenClaw setups for clients, who simply know them as an agent. The author compares self-hosted agents to Linux on the desktop, a place for technical people to build what they want, and sees the hosted bots as the same thing packaged for average users, something that can finally be handed to a non-technical sibling. The prediction is that a local machine for agents keeps gaining value as local models improve, because handing data to a few companies forever is not sustainable.
@ATLChris [OpenClaw]
https://x.com/ATLChris/status/2106544483235946877
ATLChris started on OpenClaw and left for one reason: more time went into fixing OpenClaw than using it. Hermes has been far more stable, and it now runs on a dedicated Mac mini alongside a local model, with access to the home network, a Tailnet and a browser for scheduled jobs around the clock. The self-learning features have been useful too. The hardware pattern is the same as the OpenClaw era, only the software changed.
@LukeKling [OpenClaw]
https://x.com/LukeKling/status/2106864998685528371
LukeKling uses both OpenClaw and Hermes and reports that OpenClaw on Telegram has become really nice. The specific win is a checkbox to-do list for longer multistep tasks, shown in the chat as the agent works through them. It is the same need that drives most Claude Code mods, seeing progress without asking, and the author hopes Hermes copies it.
@Chicky_Think [Claude Code]
https://x.com/Chicky_Think/status/2106559967570596034
Chicky_Think is an electronic engineer, not a programmer, who has used the Claude Code CLI since starting with AI about five months ago. The work is test automation, released as a Python executable and on Vercel. The open question now is whether to move from the CLI to the desktop app, and the author is asking the community for experience with both.
@Irish7inian [OpenClaw]
https://x.com/Irish7inian/status/2106751849613123776
Irish7inian missed the window to buy a Mac mini cheaply and regrets it, so OpenClaw and Hermes run on a child's old gaming PC. The author owns a small business and has everything possible running with agents. It is a one-line post, but it matches the day's pattern: the people still running self-hosted agents tend to be doing so on whatever hardware was lying around.
🗣 User Voice
User Voice
Browser and computer control is the gap users name first. @blueemi99 loves Claude Code after two weeks but says it loses badly to Codex on computer use, @Yuchenj_UW uses Codex to operate a computer for chores like reimbursements, and @EXM7777 installs a third-party browser plugin to close the gap. @kaede_gpt cannot get either agent to read the Mac Photos library and falls back on clicking through pictures one by one.
People want acknowledgement and better session handling. @Angaisb_ sends a steering message and cannot tell whether it was taken into account, @jerryjliu0 asks why the Claude app has no forking, and @gonzalo__nunez cannot search, filter or pin sessions or start a properly configured one from a phone. @blueemi99 adds that remote control still needs a command run in every directory.
In China, staying unbanned has become part of the job. @app_sail is selling overseas servers as the answer to the National Day ban wave, @AYi_AInotes and @LawrenceW_Zen published competing checklists about time zones, fingerprints and payment cards, and @leaf_sanren was banned without ever maxing out the weekly quota. Users are asking for clear rules more than for workarounds.
Limits are still felt, though the comparison has flipped. @adhamuxi asks whether any AI can be used all week without stress about limits, while @Khazix0918 and @Grubbyzbc both found Claude plans lasting several times longer than Codex plans at the same or lower price.
OpenClaw users want stability across updates. @jumperz dropped it from the top tier because builds need fixing after updates, @ValmereTheory is rolling back to a July version, @ATLChris spent more time fixing than using, and @furoku is retiring it after rebuilding the environment on every upgrade. @mizkun describes the relief of a hosted agent as not having to think about settings or API costs at all.
Browser and computer control is the gap users name first. @blueemi99 loves Claude Code after two weeks but says it loses badly to Codex on computer use, @Yuchenj_UW uses Codex to operate a computer for chores like reimbursements, and @EXM7777 installs a third-party browser plugin to close the gap. @kaede_gpt cannot get either agent to read the Mac Photos library and falls back on clicking through pictures one by one.
People want acknowledgement and better session handling. @Angaisb_ sends a steering message and cannot tell whether it was taken into account, @jerryjliu0 asks why the Claude app has no forking, and @gonzalo__nunez cannot search, filter or pin sessions or start a properly configured one from a phone. @blueemi99 adds that remote control still needs a command run in every directory.
In China, staying unbanned has become part of the job. @app_sail is selling overseas servers as the answer to the National Day ban wave, @AYi_AInotes and @LawrenceW_Zen published competing checklists about time zones, fingerprints and payment cards, and @leaf_sanren was banned without ever maxing out the weekly quota. Users are asking for clear rules more than for workarounds.
Limits are still felt, though the comparison has flipped. @adhamuxi asks whether any AI can be used all week without stress about limits, while @Khazix0918 and @Grubbyzbc both found Claude plans lasting several times longer than Codex plans at the same or lower price.
OpenClaw users want stability across updates. @jumperz dropped it from the top tier because builds need fixing after updates, @ValmereTheory is rolling back to a July version, @ATLChris spent more time fixing than using, and @furoku is retiring it after rebuilding the environment on every upgrade. @mizkun describes the relief of a hosted agent as not having to think about settings or API costs at all.
📡 Eco Products Radar
Eco Products Radar
Codex: the most-named product next to Claude Code, as the comparison point on limits and the borrowed engine for computer use and browser control.
Dots / Dot, Muse, Grok Bot: the hosted always-on agents the whole OpenClaw debate was measured against.
Hermes Agent: the open alternative most OpenClaw leavers move to, praised for stability and model choice.
Opus 5.5, Sonnet 5.5 and Fable 5.1: the planner, the worker and the advisor in nearly every model-routing recipe.
Claude Code Mods and You Should Know: still producing mostly visibility tools, from status bars to question explainers.
Jev: the decision model promoted for routing and approvals in dozens of near-identical posts.
Higgsfield, Seedance, Midjourney, HyperFrames, ElevenLabs: the stack behind the day's music video, toddler cartoon and manga ad.
Mac mini: argued over all day, and still the box under the desk for people running agents around the clock.
Telegram and Discord: where self-hosted agents actually live and where their owners talk to them.
Obsidian: the memory store that keeps appearing beside OpenClaw and Hermes setups.
Tailscale: the quiet standard for reaching a home agent box from anywhere.
Cloudflare: named for site hosting, Workers fleets and VM-based agent setups.
Codex: the most-named product next to Claude Code, as the comparison point on limits and the borrowed engine for computer use and browser control.
Dots / Dot, Muse, Grok Bot: the hosted always-on agents the whole OpenClaw debate was measured against.
Hermes Agent: the open alternative most OpenClaw leavers move to, praised for stability and model choice.
Opus 5.5, Sonnet 5.5 and Fable 5.1: the planner, the worker and the advisor in nearly every model-routing recipe.
Claude Code Mods and You Should Know: still producing mostly visibility tools, from status bars to question explainers.
Jev: the decision model promoted for routing and approvals in dozens of near-identical posts.
Higgsfield, Seedance, Midjourney, HyperFrames, ElevenLabs: the stack behind the day's music video, toddler cartoon and manga ad.
Mac mini: argued over all day, and still the box under the desk for people running agents around the clock.
Telegram and Discord: where self-hosted agents actually live and where their owners talk to them.
Obsidian: the memory store that keeps appearing beside OpenClaw and Hermes setups.
Tailscale: the quiet standard for reaching a home agent box from anywhere.
Cloudflare: named for site hosting, Workers fleets and VM-based agent setups.
Comments