Super User Daily: 2026-10-05
Saturday was the day OpenClaw got its eulogies and its best field reports at the same time. The feed filled with people saying Dots, Muse and Instinct made the Mac mini setup obsolete, while the people who stayed posted the most concrete work of the day: an agent that phoned a Tokyo restaurant and booked a table, an interior designer who has a second agent maintain the first one, and a full forensic write-up of a failed upgrade and how it was rolled back file by file. On the Claude Code side the money stories were unusually specific. A solo founder had an agent read 83,982 tax-preparer websites to decide what each cold email should say, a Japanese creator stopped paying 800,000 to 1,000,000 yen a month for video editing, and a team proved with a 12-task test that making the agent tick every box on a checklist changes nothing. Video made entirely in code kept spreading, from a 127,445-tile mosaic to a wedding-style montage and a $240 music video, and the complaints were about staying logged in, staying unbanned and seeing what the agent actually did.
@Crackx17_ [Claude Code]
https://x.com/Crackx17_/status/2106308419770527951
Crackx17_ is marketing a new SaaS called dokutrak, which chases accounting and law-firm clients for documents, and built the whole outbound pipeline with Claude Code. Because the IRS publishes every US tax preparer, an agent read 83,982 of their websites to see how each firm collects documents today (upload page, TaxDome, ShareFile, Google Forms) and the cold email is chosen per tool, with no AI-written lines. A second agent scores everyone who liked or commented on LinkedIn posts about the same problem; a regex filter had thrown out 2,460 of 14,778 decision makers because of job titles, while the AI scores context (a small-firm paralegal got 94, a maintenance tech got 1). Week one: 4,507 leads scored, 3 signups, 0 conversions, and the founder is now fixing onboarding rather than the list.
@masahirochaen [Claude Code]
https://x.com/masahirochaen/status/2106375997373468856
masahirochaen says a video-editing skill built on Claude Code and Codex cut outside editing costs by 800,000 to 1,000,000 yen a month, and all outsourcing has stopped. The method is specific: instead of asking for a nice edit, the skill encodes editing rules, step order and the code to run; one reference video is supplied with the instruction to match it; Claude Code acts as the hub that calls video, audio and editing tools through APIs stage by stage. Skills are split by format (long-form, shorts, news), and the agent takes screenshots at the end to check its own output for drift. Fine adjustments can still be exported to Premiere Pro.
@argofowl [Claude Code]
https://x.com/argofowl/status/2106389732322091475
argofowl found that Codex's computer-use engine is just a local MCP server inside the ChatGPT Mac app, so Claude Code can call it, even headless with claude -p. On the same 8-task test, Opus 5.5 driving Codex's engine scored 6 of 8, Codex itself 6 of 8, and the CUA driver 3 to 4 at four times the cost. The difference is mechanical: CUA's background mode on Mac goes through accessibility APIs, so canvas clicks and drags never land, while Codex's engine sends real clicks without touching the user's cursor. The post includes the exact registration prompt, with the warning that it is unofficial and has no hover support.
@mielke_jack [Claude Code]
https://x.com/mielke_jack/status/2106460142699151477
mielke_jack has listened to more than 20 hours of a personal AI podcast that Claude Code produces twice a day, at 6am and 5pm, delivered over Telegram, roughly 35 minutes in total. It reads meeting transcripts from Granola and Wispr Flow, messages from iMessage, Telegram, Slack and Gmail, and every GitHub commit labelled by who really wrote it, plus the user's own Claude Code transcripts. Sections include a Funny shit roundup of the day's absurd quotes and a Sharpening segment on technical habits, and an overnight agent builds one new project per night, recently a 3D Venus colonization game. The whole thing is now packaged as a skill others can ask their agent to build.
@enrivasgarcia [Claude Code]
https://x.com/enrivasgarcia/status/2106405495317901579
enrivasgarcia tested a Claude Code mod that redacts sensitive values before they reach the model: a 37-line filter blocked 26 values. The first version leaked an email because the domain contained the Danish letter ø and the regex missed it. The subtler failure was downstream: Claude received placeholders, believed the file really looked like that, and told the user there were no CPR numbers in it. The fixed version adds a warning line saying how many values were substituted and leaves the original file untouched, which the author sums up as: a filter can cut data, it cannot cut the model's errors of judgment.
@sebuzdugan [Claude Code]
https://x.com/sebuzdugan/status/2106427372388130952
sebuzdugan ran a small controlled test on day 17 of a 100-day series: make Claude Code write a TASKS.md and tick every box before finishing. With the checklist it passed 10 of 12 tasks, without it 9 of 12, which is no real difference. The individual runs were more telling: one ticked all 16 boxes, never ran the test suite and failed, while another left 7 boxes open, wrote All tasks complete and passed. A checklist the agent fills in itself measures its self-report, not the work.
@Rocky_Chuck [OpenClaw]
https://x.com/Rocky_Chuck/status/2106382966436123113
Rocky_Chuck posted a full incident report for an OpenClaw upgrade from 2026.7.35 to 8.35 on Ubuntu that crash-looped the gateway and then refused to roll back. The new version demanded a workspace migration via doctor --fix, which could not run while the service was up; after 13 unclean boots the restart breaker suppressed channels. Restoring the 7.35 package was not enough because 8.35 had already rewritten openclaw.json and moved the SQLite state to schema 15, and a full tar restore filled the disk. What worked was extracting three individual files from the pre-update backup and moving the new ones aside, a level of detail that shows how much state a long-running agent now carries.
@vishnuunhsiv9 [OpenClaw]
https://x.com/vishnuunhsiv9/status/2106445451524211108
vishnuunhsiv9 had an OpenClaw agent call a Japanese restaurant and make a reservation for an upcoming trip, using Twilio and ChatGPT's live voice model. In the replies the setup turned out to be almost conversational: a travel subagent had already built the itinerary, so it picked the date, time and party size itself, and then walked the owner through setting up Twilio and enabling calls to Japan. The transcript shows the staff repeating the date and party size back, and the bot's first line is an apology for calling, which commenters found very Japanese. Whether the table actually holds is still an open question until the visit.
@XMihura [OpenClaw]
https://x.com/XMihura/status/2106341399947821476
XMihura is an interior designer, not a developer, and is building a personal assistant called Gertru with a two-agent setup. An old 2015 Surface Pro with 8GB of RAM runs Linux and the ChatGPT app with sudo access, so a coding agent has full control of that machine and can be reached from an Android phone. Gertru itself is a Pi-based agent in a Linux VM, with terminal and browser control and a Telegram bot as its interface. When Gertru gets stuck, the owner tells the coding agent to read Gertru's message history, tool calls and errors and fix the cause, which is how the assistant improves through use rather than through the owner programming it.
@mathieuhq [Claude Code]
https://x.com/mathieuhq/status/2106410676239347712
mathieuhq described a pipeline where nobody writes code. The day starts in Hermes, which has access to the author's and clients' repos, inspects the code and turns an idea into a complete GitHub issue using mattpocock's grill-with-docs, to-spec and to-tickets skills. Once an issue is clear enough to run without questions, Hermes labels it agent-ready; a self-hosted GitHub Actions runner on a VPS picks it up, and Claude Code creates an isolated worktree, implements, runs tests, pushes and opens a PR. Claude Code has no merge rights on main, so the human role is reduced to specifying and reviewing pull requests.
@BugBountyDEFCON [Claude Code]
https://x.com/BugBountyDEFCON/status/2106187665611636759
BugBountyDEFCON shared a talk by Jason Haddix on Armada, Arcanum's in-house hackbot running on Claude Code. A single six-hour run produced 4 confirmed critical and 17 confirmed high-severity findings. The talk is notable for walking through every failure it took to get there rather than only the result. It is one of the clearest numbers so far on what an offensive-security harness built on a coding agent yields per run.
@codyschneider [Claude Code]
https://x.com/codyschneider/status/2106429078555247020
codyschneider runs three small media properties for a business with AI agents and shared the numbers. An industry directory site, planned and built by Claude Code in about 30 minutes after an agent researched every target company, reached about 100 clicks a day after 90 days and now carries ads for the product. An industry news podcast and newsletter, researched by an agent and voiced by ElevenLabs, grew to 20,000 subscribers in six months. A TikTok slideshow farm of 10 cloud accounts does 300,000 impressions at about a $3 CPM.
@AaronxShepherd [Claude Code]
https://x.com/AaronxShepherd/status/2106172563676144060
AaronxShepherd's agency sends more than 2 million cold emails a month and replaced manual Apollo list building with Claude Code. Claude first interviews the user about the ideal customer one question at a time and saves an icp.md file, then turns that file into an Apollo search, excludes named categories like vendors and insurance brokers, and scores every lead 1 to 10 against the same file. The last run returned 50 contacts across 25 mid-size US hospitals in 22 states. The chain is wrapped in a skill, so the next list is one command and calls go only to the 10s.
@33339999g [Claude Code]
https://x.com/33339999g/status/2106368681886543964
33339999g runs a business with monthly fixed costs above 1 million yen, of which AI is only 70,000 to 80,000. Claude Code Max, Codex and Genspark together cost about $300 a month and handle admin, prospect lists and proposal drafts, a workload the owner estimates would take five or six people. The suggestion is practical: before arguing about whether AI is expensive, calculate its share of fixed costs, because that number makes it easier to decide what to hand over next.
@David_Ouyang [Claude Code]
https://x.com/David_Ouyang/status/2106244814073364979
David_Ouyang is building an interactive echocardiography tutorial with Claude Code. Chapter 7 shows a 3D heart, the standard imaging views, the probe position for each, and simulated ultrasound video. Medical education material of this kind usually needs a dedicated developer and 3D artist, and here it is being produced chapter by chapter by a domain expert.
@redp314 [Claude Code]
https://x.com/redp314/status/2106324559805010004
redp314 made a mosaic portrait that does not exist: 127,445 tiles, each its own 3D object, placed one by one in code by Claude Code with Opus 5.5. The camera starts inside the subject's eye and ends on a pearl. A separate post about a Roman mosaic of 7,878 stones, built the same way in about four hours of back-and-forth, shows this has become a genre: no image or video model, no Blender, just code that renders in a browser.
@maxescu [Claude Code]
https://x.com/maxescu/status/2106395987996536913
maxescu released a music video in which the song, footage, edit, motion graphics and grading are all AI, with only the idea, curation and art direction done by a person. Claude Opus 5.5 orchestrated it through Claude Code and subagents, calling Higgsfield, ElevenMusic, Seedance, Nano Banana Pro and GPT Image, plus a long list of local audio and vision tools: Demucs for stem separation, Whisper and a forced aligner for lyrics, beat tracking, face landmarks and background removal. Total cost was about $240, with a few skipped frames and lip-sync the author rates as 99% there.
@Yumzlef [Claude Code]
https://x.com/Yumzlef/status/2106383842382602644
Yumzlef is publishing small science visualizations written entirely by Claude Code. One is the Gray-Scott reaction-diffusion model: the same six seed squares grow into spots, coral, a maze or holes depending only on two rates, in 128 lines of plain JavaScript, 18,480 steps rendered at 1080p in 94 seconds. Another trains a neural network from scratch on the three-arm spiral, with forward pass, backprop and Adam in 169 lines and no ML library, going from 24.5% to 99.2% on unseen points. Both are the kind of explainer that used to require a dedicated science animator.
@RoundtableSpace [Claude Code]
https://x.com/RoundtableSpace/status/2106244100303270276
RoundtableSpace highlighted Floodplain, a free Dutch low-country city builder with water tables, dike upkeep and walker-delivered goods, built in JavaScript and three.js over a few months of evenings with Claude Code. Claude handled the simulation, procedural 3D models for every house, mill, church and boat, and headless testing with bot players that reported where they got stuck. Where it fell short is equally specific: drawing a campaign map failed four times until it was rendered with the game's own engine, chapter one asked for 300 people and a cathedral before a human noticed it was too hard, and the call to make warnings bigger had to come from the developer.
@shinshin86 [Claude Code]
https://x.com/shinshin86/status/2106369796895846842
shinshin86 is turning a quoting mechanism into a tool that lets anyone build a Live2D-style mesh avatar from a single illustration. After the Claude Code version, a second illustration was run through Codex alone and worked well from the start. Eye and mouth open-close states are better drawn as separate images, so the pipeline asks Codex's image generation for those diffs. The author's current best practice: let an agent build the Live2D-style mesh base, then patch details with image generation.
@tokkyo [Claude Code]
https://x.com/tokkyo/status/2106240136778575897
tokkyo built a visualization of Claude Code's work history on a whim, and now sees at a glance when each piece of work happened. The post drew more than 370,000 views, which says as much about demand as about the tool: once people run several sessions a day, a timeline of what the agent did becomes something they want to look at, not dig for.
@is_llll [Claude Code]
https://x.com/is_llll/status/2106253961170919529
is_llll turned off Claude Code's auto memory, reviewed CLAUDE.md line by line, cut it from 1,070 lines to 344, and banned the agent from editing it again. The result was that Opus 5.5 suddenly seemed much smarter. The author's hypothesis is that Opus 5.5 is more sensitive to context pollution than Fable, which matches a recurring theme this week: the file the agent maintains for itself grows faster than its usefulness.
@yupi996 [Claude Code]
https://x.com/yupi996/status/2106232523751673974
yupi996 followed Matt Pocock's advice that keeping an agent from wandering around a repo is an underrated token saver, and checked the session logs of one project. In 9 of the last 10 sessions the agent was reading and searching the same ten-thousand-word notes file. Pocock's /retro command reads recent sessions and reports what was hard to find and which docs are stale; its most common advice is to stop adding rules to CLAUDE.md, which should only hold pointers to where things live. Practical notes: install it yourself because the marketplace copy is outdated, name the project in the prompt, and discard any suggestion it cannot tie to a specific session.
@Damir_Akaza [Claude Code]
https://x.com/Damir_Akaza/status/2106333764548702392
Damir_Akaza no longer lets Claude close a task until a model from another company has reviewed it. A Stop hook fires when Claude says done and runs codex review --uncommitted, which only reads the changes and returns a list of problems; Claude then fixes them or explains why it disagrees. The second attempt to finish passes without review, so there is no infinite loop, and the hook stays quiet when nothing changed. The reasoning: a model checking its own code shares the blind spots that produced it, while a different vendor's model has different ones.
@jackfriks [Claude Code]
https://x.com/jackfriks/status/2106401479519318402
jackfriks keeps 10 remote-control Claude Code sessions running at all times on a docked machine, with caffeinate keeping it awake. The sessions are managed through iTerm2, and a shortcut that Claude itself built resumes all of them on demand. It is a small detail that shows where heavy users are: the unit of work is no longer a session but a standing fleet that needs to survive sleep and restarts.
@altryne [Claude Code]
https://x.com/altryne/status/2106173279660855596
altryne has to SSH in and log Claude Code back in every two to three days, which makes remote control close to useless for an always-on setup. The complaint is aimed at Anthropic directly: the push toward cloud sessions is understood, but long-lived local sessions are what people are building on today. It is the most common infrastructure complaint of the day, from someone who clearly runs agents unattended.
@leaf_sanren [Claude Code]
https://x.com/leaf_sanren/status/2106379623865221274
leaf_sanren's Claude account was frozen, and GPT picked up the work without a gap, because Claude Code stores full conversations as plaintext in the local home directory where any other agent can read them, while Claude chat history is encrypted and disappears with the account. The advice is to use Claude Code rather than chat whenever possible and to write a habit of saving run logs and summaries into the global agent contract file. A follow-up post describes the cost of switching: a $20 Codex plan burned its five-hour quota on just importing the Claude Code config, skills and scheduled tasks and reading one project.
@heliotsx [Claude Code]
https://x.com/heliotsx/status/2106371589759541606
heliotsx bought a used M4 Mac mini for R$4,600 to self-host a Grok Bot-style agent instead of paying $300 a month for Grok Bot plus macOS actions for iOS testing. The setup also lets the team use native Claude Code on the $200 Anthropic plan they already pay for. The current workaround is telling: the Grok-based bot installs Claude Code on its own machine and delegates every serious task, debugging and development, to it, because Grok is still well behind Opus 5.5 on quality. The author expects the Mac mini to pay for itself within six months.
@timothyachumba [Claude Code]
https://x.com/timothyachumba/status/2106489967254098009
timothyachumba traced a year of trying to orchestrate agents across work and life. It started with five or six terminal tabs, then Claude Code subagents with one orchestrator per project, then Paperclip with a custom iMessage agent, a pages system so agents could write reports and morning briefs, and custom orchestration on T3 Code. Visibility was the recurring sticking point: what the agents were doing, what had finished, where they were blocked. Dot is the first setup that seems to hold the broad picture while connecting to the computer and agents, though it is still being wired up.
@L_go_mrk [Claude Code]
https://x.com/L_go_mrk/status/2106353572849787250
L_go_mrk describes a video production company making a 55-second brand film entirely through instructions and review in Claude Code. The key prep was 11 company documents in one folder (overview, workflow, numbers, brand, logo, colors, typeface) that Claude Code reads before starting. Higgsfield handles footage, ElevenLabs voice, music and effects, After Effects motion and compositing, with Claude Code driving the whole sequence. The instructions specify size, length, full narration, a structure matched to the music's beats, design and build order, so the expert's tacit decisions become reusable material for the next job.
@uniyume [Claude Code]
https://x.com/uniyume/status/2106295668830556340
uniyume used to shoot and edit same-day wedding end-roll videos alone in three hours, and now turns 15 seconds of AI footage into a 30-second piece with motion added purely through conversation in Claude Code with Opus 5.5 and HyperFrames. Even the side-by-side before-and-after video was one request: put two phones next to each other and keep the visual world consistent. Timing tweaks, like moving a shutter sound a few seconds later, are also just words now. The stack: ChatGPT for storyboards, MiniMax H3 via fal for video, Suno for music.
@AI_swwww [Claude Code]
https://x.com/AI_swwww/status/2106489746344358222
AI_swwww handed an entire music video to Claude Code: lyrics, composition, vocals, dance and editing, with the agent coordinating across tools. The human contribution was handing over one image and saying which parts were not good enough. The post maps the flow as a tree, which is becoming the standard way creators document these multi-tool productions.
@seiiiiiiiiiiru [Claude Code]
https://x.com/seiiiiiiiiiiru/status/2106310907899105704
seiiiiiiiiiiru made a fictional shampoo commercial with every step routed through Claude Opus 5.5 and automated end to end in Claude Code. The chain: Nano Banana Pro and Kling for live-action shots, one frame converted to anime style, Python to turn frames into manga panels that keep color only on hair, glasses and product, Higgsfield animation converted panel by panel, After Effects reading a spread with a 3D camera, Kling to make characters leap out of the manga, and Gemini TTS for narration. The author's takeaway is that with this much range, choosing the direction is now the scarce skill.
@FFBuncho [Claude Code]
https://x.com/FFBuncho/status/2106386541547536509
FFBuncho had Claude Code add subtitles to a Grok-generated video. Claude cannot hear audio, so the only extra install was Whisper, and after that the subtitles were done a few minutes after upload. The motivation is ordinary: X's auto-captions mishear, and adding captions manually means editing on the phone at upload time. It worked well enough that the author admits to being slightly unnerved.
@eternityspring [Claude Code]
https://x.com/eternityspring/status/2106391529082495202
eternityspring saw the alvdansen/h3-keyframe-animation repo on the timeline, dropped the link into Claude Code with Opus 5.5 and asked for an evaluation. Half an hour later an evaluation video was ready, which the author compared to making a wish. The same author also used reelbench-skills to break down a reference video of the viral comedic grandma format and had Claude rewrite the script and produce a new video, noting that without the skill the fight scenes were weak.
@nathannam [Claude Code]
https://x.com/nathannam/status/2106404547652681921
nathannam saw customers of a previous observability company ask why they should train engineers on the product when everyone uses coding agents, and demand a CLI, skills and an MCP server instead. At Physical AI the author built that interface for the Newton simulator: every demo app and end-to-end example became a reusable skill, and the contribution graph went from 378 to 2,161 contributions in a year. In hands-on workshops, one attendee said that within 15 minutes of using Newton through Claude Code they felt they could build anything.
@mary_nekyia [Claude Code]
https://x.com/mary_nekyia/status/2106340464203047378
mary_nekyia asked Fable in a cloud Claude Code session to review and upgrade a Discord MCP server. A maintainer reviewed and merged the change the next morning. The MCP now uses updated Discord API endpoints, has better tools, and can act as an admin on servers. It is a clean example of an agent doing maintenance work on open-source infrastructure that a human then signs off.
@fayequepeerzade [Claude Code]
https://x.com/fayequepeerzade/status/2106404138577957230
fayequepeerzade gave the same prompt to Opus 5.5 on medium reasoning, with no skills, in Google Antigravity and in Claude Code. Antigravity took 8 minutes and consumed 21% of the weekly limit and 52% of the five-hour limit. Claude Code finished in 2 minutes. With Antigravity having just added Claude models for Google AI Pro subscribers, this kind of same-model, different-harness comparison is exactly what people are trying to work out.
@agtmengineer [Claude Code]
https://x.com/agtmengineer/status/2106341952300204283
agtmengineer's first day with Instinct ended with a network-tracking system: every person met is saved into Notion from a WhatsApp text or voice note, and a live Pokédex-style app of contacts is built on Zite. Twice a day a Claude Code routine reads the new entries, enriches what is missing, scores contacts against the author's GTM notes and drafts follow-ups, and nothing is sent without approval. A Sunday digest lists who was met and pending actions. The broader point: personal assistants already hold the context, so the opportunity is building services on top of them.
@vkryukov [Claude Code]
https://x.com/vkryukov/status/2106260902798582233
vkryukov describes an always-on development setup: a Claude Code CLI on a VPS with the right skills and GitHub as the source of truth. The author talks to a chief-of-staff agent while running; it tells an Idea Capturer to write a detailed, cross-linked GitHub issue and a Builder to execute it, handles coordination and quality control, and pings for escalations. A tester runs a full end-to-end QA routine periodically, and another specialist captures incoming feedback and passes it back to the Idea Capturer. Grok 4.x was dropped for coding because it needs to get faster and better.
@stretchcloud [Claude Code]
https://x.com/stretchcloud/status/2106190245402476695
stretchcloud built Campfire after noticing that power users juggle many subscriptions by hand, with one well-known creator running six Claude and three Codex subscriptions. It runs Claude Code, Codex, Aider and Goose side by side in one browser tab, isolated by git worktrees. Its distinctive feature is permission voting across backends: a majority approves an action, any single deny blocks it, and a 30-second timer keeps things moving. It can also race the same task across three backends and shows all subscriptions as one cost number.
@bokuwalily [Claude Code]
https://x.com/bokuwalily/status/2106295124653391978
bokuwalily, a former CMO who builds apps solo with Claude Code and almost no hand-written code, writes the App Store description before writing any code. There are always three sticking points: the first line, which must say who opens the app and when; the third screenshot, which tests whether there is a reason to come back; and the how-we-differ field, where writing only AI-powered means competitors say the same. If any of them stalls, the spec memo gets rewritten before Claude Code ever sees it, because stopping at the description stage is far cheaper than discovering it mid-build.
@davekiss [Claude Code]
https://x.com/davekiss/status/2106448502133842387
davekiss compiled every Claude Code mod worth installing from the first two days and added one of their own: env, an environment-variable manager where Claude never sees the values. The list shows what people built first: pasted-image thumbnails, a live tree of the agent, subagents and every tool call, a view of child chats and unmerged worktrees, a context-window weather forecast, a prompt-cache timer, an in-app installer, and multiplayer Doom or code typing races while waiting. Almost all of them are about seeing the agent, not changing what it does.
@nateherk [Claude Code]
https://x.com/nateherk/status/2106412212978098295
nateherk built a Claude Code mod that gives a visual of /goal prompts, showing tasks, progress and ETA. It is most useful with several goals running at once, because all of them can be viewed in one place. It is a good illustration of how /goal and mods combine: one sets the finish line, the other makes progress toward it visible.
@dani_avila7 [Claude Code]
https://x.com/dani_avila7/status/2106455605925822967
dani_avila7 recommends the Cache Control mod as one of the biggest token savers. It adds a bar showing Claude's five-minute prompt-cache window and sends notifications to help keep the cache warm, so sessions last longer before limits hit. The post drew close to 190,000 views, a sign that cache economics has become everyday knowledge for heavy users.
@AniC_dev [Claude Code]
https://x.com/AniC_dev/status/2106406423122878870
AniC_dev gathered feature, pricing and funding data for more than 366 sandbox providers by asking Claude Code to run 100 cloud VMs in parallel, each spawning a browser and a subagent that scraped a different subset with human-like mouse movement, then merge everything into a CSV and a hosted report. The VMs come from a sandbox provider that combines anti-bot browsers with persistent files and hosting, and residential proxies can be plugged in through nftables. It is a concrete template for mass browser work that a single local agent could not do.
@theo [Claude Code]
https://x.com/theo/status/2106473854113894520
theo improved the usage view in T3 Code so it shows where spend goes and how different models behave on the user's own data. The detail that matters is scope: it measures all Claude Code and Codex usage on the machine, not just usage inside T3 Code. Replies called this the spend view that actually matters, since most heavy users now split work across two vendors.
@ledevultime [Claude Code]
https://x.com/ledevultime/status/2106471672689922141
ledevultime made a short film using only Claude Desktop with Claude Code and Opus 5.5, the author's own voice, and a human-made base illustration. Claude generated the variations of that illustration by driving Grok through its CLI. No HyperFrames, no Remotion. The point is that the orchestration layer can now call other vendors' generators directly from the command line.
@paonx_eth [Claude Code]
https://x.com/paonx_eth/status/2106376383085633960
paonx_eth sent a crawler through GitHub's AI repos from the last 30 days: one crawler fanned out to 128, read 2,000 READMEs, and asked a decision model one question per repo, whether it was worth the reader's time. 99.3% were discarded as wrappers, copies, empty READMEs or lists of lists, and 9 survived, including a browser skill that lets Claude Code work in a logged-in browser and an internal code reviewer said to use about one ninth of Claude Code's tokens on the same model. Four of the nine came from Alibaba, Cloudflare and Tencent.
@EthereumThaila1 [OpenClaw]
https://x.com/EthereumThaila1/status/2106398778567827933
EthereumThaila1 runs OpenClaw on a local machine with Sonnet to take on tasks and shared the unit cost: a little over 1 baht per instruction. It finishes quickly and neatly, but at 100 tasks a day that is 100 baht, which the author calls expensive. In a separate post the same user explained staying on OpenClaw rather than switching to a hosted assistant, because it gives more freedom to back up files, and routing it to a Chinese model on Tencent Cloud for the cheapest per-call price.
@rozhkov_ [OpenClaw]
https://x.com/rozhkov_/status/2106333394422640659
rozhkov_ created a product-owner agent for each project plus a portfolio lead agent that controls them, all in OpenClaw. Setup is still rough, but the author sees no other choice for scaling personal attention across projects. The stated need is simple: an entity that keeps track of things and pushes them forward.
@lmedinas [OpenClaw]
https://x.com/lmedinas/status/2106302063801987363
lmedinas already runs a home setup with OpenClaw: the agent controls a mini-display at home showing the family's calendars, sets music on the Sonos to match, and controls Home Assistant on demand. It was posted in reply to a call for ambient agent hardware, and it shows the kind of household integration long-time OpenClaw users built before personal-agent products arrived.
@juancahurtadoai [OpenClaw]
https://x.com/juancahurtadoai/status/2106504880089420005
juancahurtadoai says their whole ad system has been automated on OpenClaw for a while, especially Google and Bing Ads campaigns. The APIs are connected, reports arrive in Telegram every morning, and the agent raises or lowers campaigns according to a target ROAS. Combined with brand-kit creatives, campaigns can also be launched and pruned automatically.
@zengbozb [OpenClaw]
https://x.com/zengbozb/status/2106325602336907650
zengbozb never ran OpenClaw, but read the 2.0 source last month and lifted two ideas into a personal setup: the task ledger and the capability manifests. Both are still in daily use. As OpenClaw's user mindshare fades, its design ideas are being harvested by people building their own agents.
@moriyorihayash1 [OpenClaw]
https://x.com/moriyorihayash1/status/2106342879324979533
moriyorihayash1 describes devoting most of 2026 to babysitting an OpenClaw agent on a Mac mini, and credits it with an X account that went viral and launched a business. Dot now feels like that same experience released to everyone: a VM instead of a physical box, with connectors, scheduled tasks and message collection that need no effort. The author still uses OpenClaw but advises newcomers to master Dot.
@DeFi_Hanzo [OpenClaw]
https://x.com/DeFi_Hanzo/status/2106481452191789448
DeFi_Hanzo has used OpenClaw since January and lists the downsides plainly: high costs, frequent fixes and limited abilities. GrokBot, tried this month, has low limits, struggles with simple things and makes many mistakes, though it is good at agentic web tasks. The distinction drawn is useful: OpenClaw is something you build from scratch with room to shape it, and the real barrier to trying Dots is that moving everything from one agent system to another is a nightmare.
@kgsi [OpenClaw]
https://x.com/kgsi/status/2106456113084518895
kgsi says Dot and Cursor's bot have made OpenClaw completely obsolete, and the bot installed on a Mac mini no longer even starts. The author grants that OpenClaw shaped every autonomous bot that followed, but the pace of churn feels like it is burning out the brain. It captures the day's mood among early adopters: gratitude and exhaustion in the same post.
@federicolentini [OpenClaw]
https://x.com/federicolentini/status/2106395834682167386
federicolentini updated OpenClaw from 9.7 and it broke the setup completely, with Codex trying to revive it for half an hour. Others in the same thread joked about peaceful hours while a different AI figures out why the update broke, and one team announced a rollback to the July version. Updates breaking long-running agents was the most consistent OpenClaw complaint of the day, and the release's own maintainers posted a same-day follow-up because the headline model support had not fully shipped.
🗣 User Voice
User Voice
Staying logged in and staying unbanned is now a workflow concern. @altryne has to re-login over SSH every two to three days, which breaks unattended remote control, and @leaf_sanren lost a Claude account and only kept working because Claude Code logs are plaintext on disk. In Chinese, @ai_funss is comparing residential IPs and App Store gift cards as a ban-avoidance stack. Users want session longevity and a clear ban policy more than new features.
People want to see what the agent did and verify it independently. Mods are overwhelmingly visibility tools, and @trevin and @aiwithsally both liked You Should Know for catching what they would have missed. @sebuzdugan showed that a self-ticked checklist proves nothing, and @Damir_Akaza now routes every finish through another vendor's reviewer. @Rencrypta and @Joshua_WD flagged the catch: mods run with the user's permissions and no sandbox.
Remote and mobile control remains incomplete. @rishabhnasa wants to reach the same Claude Code instance from a phone when the laptop cannot come along, @KirkMarple likes the iOS app but wants to see shell commands and their output, and @jan_volad wants background processes that wake the agent when they finish rather than a polling heartbeat.
Context hygiene and limits keep surfacing together. @is_llll got a visibly smarter Opus 5.5 by cutting CLAUDE.md to a third, @btc_MasterPlan burned 51% of a weekly limit in two days, and @ScarletKc asks whether a Fable advisor is still worth its extra quota now that Opus 5.5 handles most tasks.
OpenClaw users want upgrades that do not break a running agent. @federicolentini, @jlowin and @ValmereTheory all describe updates that took a working setup down, @Rocky_Chuck documented a rollback blocked by schema migrations, and @absol_89 could not get the Windows installer to run a local gateway. Meanwhile @kgsi and @DeFi_Hanzo say hosted personal agents feel easier, though migrating is the real cost.
Staying logged in and staying unbanned is now a workflow concern. @altryne has to re-login over SSH every two to three days, which breaks unattended remote control, and @leaf_sanren lost a Claude account and only kept working because Claude Code logs are plaintext on disk. In Chinese, @ai_funss is comparing residential IPs and App Store gift cards as a ban-avoidance stack. Users want session longevity and a clear ban policy more than new features.
People want to see what the agent did and verify it independently. Mods are overwhelmingly visibility tools, and @trevin and @aiwithsally both liked You Should Know for catching what they would have missed. @sebuzdugan showed that a self-ticked checklist proves nothing, and @Damir_Akaza now routes every finish through another vendor's reviewer. @Rencrypta and @Joshua_WD flagged the catch: mods run with the user's permissions and no sandbox.
Remote and mobile control remains incomplete. @rishabhnasa wants to reach the same Claude Code instance from a phone when the laptop cannot come along, @KirkMarple likes the iOS app but wants to see shell commands and their output, and @jan_volad wants background processes that wake the agent when they finish rather than a polling heartbeat.
Context hygiene and limits keep surfacing together. @is_llll got a visibly smarter Opus 5.5 by cutting CLAUDE.md to a third, @btc_MasterPlan burned 51% of a weekly limit in two days, and @ScarletKc asks whether a Fable advisor is still worth its extra quota now that Opus 5.5 handles most tasks.
OpenClaw users want upgrades that do not break a running agent. @federicolentini, @jlowin and @ValmereTheory all describe updates that took a working setup down, @Rocky_Chuck documented a rollback blocked by schema migrations, and @absol_89 could not get the Windows installer to run a local gateway. Meanwhile @kgsi and @DeFi_Hanzo say hosted personal agents feel easier, though migrating is the real cost.
📡 Eco Products Radar
Eco Products Radar
Codex: named in more than 100 of the top posts, now mostly as the second agent next to Claude Code, from reviewer to computer-use engine.
Dots / Dot, Muse, Grok Bot, Instinct: the hosted personal agents framed as OpenClaw's successors, compared on setup friction and limits.
Hermes Agent: the open harness people pick when they want model choice and their own server.
Pi / Pi Durable: the minimal harness under several self-built assistants and research setups.
Opus 5.5 and Fable 5.1: the default model plus the on-call advisor, with users now questioning whether the advisor still pays for itself.
Claude Code Mods and You Should Know: day three of mods, still dominated by visibility, cost and redaction tools.
Jev: the decision model used for routing and filtering inside pipelines.
Twilio: the phone layer behind agents that make real calls.
Higgsfield, Seedance, Nano Banana Pro, HyperFrames, ffmpeg: the video stack behind code-driven film and ads.
Telegram: the default interface for always-on personal agents.
Mac mini: the hardware of the OpenClaw era, now either repurposed or gathering dust.
Codex: named in more than 100 of the top posts, now mostly as the second agent next to Claude Code, from reviewer to computer-use engine.
Dots / Dot, Muse, Grok Bot, Instinct: the hosted personal agents framed as OpenClaw's successors, compared on setup friction and limits.
Hermes Agent: the open harness people pick when they want model choice and their own server.
Pi / Pi Durable: the minimal harness under several self-built assistants and research setups.
Opus 5.5 and Fable 5.1: the default model plus the on-call advisor, with users now questioning whether the advisor still pays for itself.
Claude Code Mods and You Should Know: day three of mods, still dominated by visibility, cost and redaction tools.
Jev: the decision model used for routing and filtering inside pipelines.
Twilio: the phone layer behind agents that make real calls.
Higgsfield, Seedance, Nano Banana Pro, HyperFrames, ffmpeg: the video stack behind code-driven film and ads.
Telegram: the default interface for always-on personal agents.
Mac mini: the hardware of the OpenClaw era, now either repurposed or gathering dust.
Comments