Super User Daily: 2026-10-07
Monday's feed was dominated by two stories that are not use cases, so they get one sentence each: The Information reported that Meta halved its Claude Code seats and Microsoft cut internal Claude spend by a third, and Anthropic moved every new Cowork task for Pro and Max users to the cloud, which pushed a wave of people back to Claude Code as the last local option. Underneath that noise the actual work was unusually candid about failure. One user installed every tool their social feed recommended, 800 skills and 31 MCP servers and 68 hooks, and watched Claude Code auto-compact every half second until it could not run the prompt asking it to fix itself. Another caught the bash tool hanging for 12 minutes on a cargo test and had to stop it by hand. On the productive side, a team of three engineers has run on Claude Code alone since May, a high school timetable got solved as a constraint problem, a security professional described onboarding OpenClaw the way you would an executive assistant, and a Figma file became a client site in one prompt. The OpenClaw thread was a migration ledger: a $32 droplet shut down for a $16 Muse plan, a $700 Mac mini untouched for six months, and a handful of people who went the other way and said they want their old setups back.
@nonstopmatt [Claude Code]
https://x.com/nonstopmatt/status/2106964291584725057
nonstopmatt spent one night forcing an Instagram feed to serve only videos about repos that would upgrade Claude Code, saved every link, and installed all of it the next day: 800 skill files, 31 MCP servers, 68 hooks and 21 plugins, all loading every session. Context went from 100 percent to zero on launch and Claude Code auto-compacted every half second. The 285 agent descriptions alone totalled 16k tokens, over the 15k limit Claude Code had been warning about the whole time, and a memory plugin injected 24.8k tokens per session whose last line said to continue a CRM deployment, so every new session redid finished work. A self-improvement skill accounted for 26 percent of usage. The lesson in the post is that these tools do not cooperate, they stack, and every one adds text to every prompt whether it is used or not.
@EXM7777 [Claude Code]
https://x.com/EXM7777/status/2107108825761558799
EXM7777 logged 70-plus hours with Opus 5.5 inside Claude Code and distilled it to a setup list: keep effort on medium because higher never produced better output, keep the context window under 400k tokens, run a single global CLAUDE.md and let the code and knowledge base do the rest, use git worktrees with one main chat as the router that spawns each project's thread in its own worktree, lean heavily on subagents to keep the main context clean, skip plan mode in favour of the grill-me or Compound engineering plugins, and start projects with heavy research plus a pile of references before trusting Opus overnight. The two most contrarian lines are to disable 90 percent of installed skills and to have Opus maintain an artifact that maps the full scope of each project and updates it at every milestone.
@yifanxu_ephai [Claude Code]
https://x.com/yifanxu_ephai/status/2107118307837145595
yifanxu_ephai has been saying for months that the Claude Code bash tool is a poor fit for long-running tests, and posted a concrete hang to prove it. A cargo test for a newly added concurrent append-and-tail test ran for over 12 minutes with its output piped through grep, so nothing printed, and the processes had already exited by the time the command was stopped by hand. Claude's own post-mortem admitted it was a hang rather than slowness, could not identify the cause, marked the test as ignored, added a drop guard and a 60 second timeout, and noted the fix compiled but had never actually been run. The author's preference is the Codex exec plus write-stdin polling model.
@chrisholland [Claude Code]
https://x.com/chrisholland/status/2106936972979696068
chrisholland reports that their whole team has run entirely on Claude Code since May 2026, with no harness beyond a lightweight plugin they built themselves, on a Claude Max subscription at $200 per engineer per month. The notable data point is that they very rarely hit daily or weekly limits, which the post attributes to keeping the workflow lean rather than stacking tools on top.
@nb4ld [Claude Code]
https://x.com/nb4ld/status/2107207404488155219
nb4ld answered the question of whether an LLM can optimise a high school timetable with all its constraints, and the answer was a method rather than a prompt. Use a tool that writes and runs code, such as Claude Code, Codex, Hermes or Mistral Vibe. First prompt: here are all my constraints in heterogeneous formats, define a common language for them, identify each format, build a converter for each, then run a deterministic script to push every source through. Second: now write a constraint solver that proposes the most optimal timetables, with optimal defined as x, y, z. Third: add that one teacher is on maternity leave half the year and building A has works every morning. Fourth: run the solver for a first draft and show me the cases that remain problematic. It is a clean example of turning a school administration problem into a formalise-then-solve loop.
@robtlee [OpenClaw]
https://x.com/robtlee/status/2106905072248442968
robtlee, a security professional, uses OpenClaw and does not buy the fearmongering, but also does not connect it to everything at once. The post describes granting access slowly, the way you would onboard an executive assistant, so it takes over the things the author would rather not spend time on. The rest of the post is a field guide to the five personal agents that shipped recently, Grok Bot on its own cloud machine, Muse with email and payments, Dots inside ChatGPT for Slack and Teams, Hermes Bot Mode as named bots passing work to each other, and Cue giving each agent an email address, phone number and wallet.
@JussiKorhonen [Claude Code]
https://x.com/JussiKorhonen/status/2107077848506737134
JussiKorhonen's client shipped a site using nothing but Claude Code and the designer's Figma file, and the first prompt got more than 90 percent of the work right: key CSS classes matched the Figma frame names, animations came out as specced, SEO and responsiveness were present. The author's estimate is that a few years ago this was easily 10,000 euros of developer build work. The thing that made it work was structure behind the visuals, variables for everything and important frames named like CSS classes, which is why the post frames it as senior designers suddenly being able to deliver far more to a business client.
@rvaniaaaa [Claude Code]
https://x.com/rvaniaaaa/status/2107221534821498977
rvaniaaaa runs Claude Code as a team rather than one agent and posted the full configuration: a Critic on Fable 5.1 that checks in at three points, where modules meet, after a test fails twice, and before a PR goes out; an Architect on Opus 5.5 at high effort that owns the plan and the final merge; Developers on Sonnet 5.5 at medium effort for UI and backend, each owning its own files; and Jev picking which file to open and which tool to run in under half a second. Teammates share one task list and claim work with built-in file locking. The setup prompt enables the experimental agent-teams flag in settings, sets high effort inside the Opus model settings since a top-level effort level does not apply to Opus 5.5, writes the subagent definitions, and adds the critic rule to CLAUDE.md, showing every config diff before applying.
@FavioVaz [Claude Code]
https://x.com/FavioVaz/status/2107178413241106799
FavioVaz took the memory Claude Code had kept for a project, 19 notes over 10 days, turned it into a git repo, and pointed a fresh Devin session at it. Devin recalled what Claude Code had learned and then dreamed on it: 21 commits, the 19 notes consolidated to 14, duplicates merged and stale notes removed. It is a small but clean demonstration of one agent curating another agent's memory when the memory lives in plain files under version control.
@guilleflorvs [Claude Code]
https://x.com/guilleflorvs/status/2107116888341803159
guilleflorvs relayed a Ben Horowitz podcast segment on why software companies trade at low multiples, and the host James Norman was the evidence. About a year ago Norman started using Claude Code, cancelled HubSpot the next day to build a CRM, had it a week later, then cancelled Affinity and built a fundraising platform. The rule now is that a tool has to be very special for Norman not to build it in-house, which is the customer-side version of the thesis that even a customer can rewrite Salesforce.
@borjafat [Claude Code]
https://x.com/borjafat/status/2107138649150124258
borjafat made Opus 5.5 in Claude Code the chief of SEO backlinks and organised it as six skills by the work they do: sources finds the pages AI answers cite by asking buyer questions and keeps the publishers worth pitching; reclaim fixes pages that receive links but return 404s and finds unlinked brand mentions; gift sends publishers verified corrections, fresh screenshots and diagrams; listicles finds roundups featuring competitors and checks for products that shut down; drip reviews one exchange placement a day; and linkreview runs every Monday to confirm links are live and compare Search Console clicks. Claude researches each opportunity and drafts the pitch, while the human approves placements and sends every email.
@naomu_ai [Claude Code]
https://x.com/naomu_ai/status/2107102661053223252
naomu_ai makes the point that everyone now builds AI employees with Claude and Codex but few manage them, and describes using Dots as the online secretary that checks on each employee's session progress. The concrete example is a booking task handed to the clerical employee through Dots, which went from reserving a rental space to completing payment on a single instruction. The post also notes that routing Claude to Codex and Codex to Claude Code works without touching the CLI from the desktop apps.
@Sarut0biSasuke [Claude Code]
https://x.com/Sarut0biSasuke/status/2107076101293969587
Sarut0biSasuke works across several brands that do not share a mailbox, Drive or Slack, so the single-company Grok Bot setups did not fit. Over a weekend the author built an agent fleet: each brand's Grok Bot scans its own tools and writes one note a day into its own folder of an Obsidian vault that is also a git repo, discovered tasks are delegated through an agent inbox inside that repo, and Claude Code, Codex and Grok each have a task inbox for heavy lifting such as a Figma banner, a detailed X scan or a website build. The coding agent does the work and replies to the inbox. This gets around the one-account-per-connector limit on Gmail, Fireflies and Slack, and the next step is Discord monitors and a scoped Telegram MCP per brand.
@AaronErickson [Claude Code]
https://x.com/AaronErickson/status/2107207001759793457
AaronErickson lists what non-engineers around them do with Claude Code from the CLI: deal with medical claims, renew DMV paperwork, find subscriptions they no longer use, auto-filter email scams and draft expense reports. The first time the author handed an executive a Claude Code setup earlier this year, the response was I am in love, which after 46 years of writing code was the first time anyone had said that about a program. The post's complaint is that the barrier is imagination, not capability.
@egr_investor [OpenClaw]
https://x.com/egr_investor/status/2107185571852583377
egr_investor runs chief-of-staff style workflows at home and at work built on OpenClaw-like automation: they scan email, calendars and Slack, surface things that might be missed, and draft or send emails. The same setup produces weekday morning briefings, weekend briefings and a weekly executive briefing covering priorities and loose ends. The reported result is dropping the ball on fewer things, which the author ties to having only so many threads to hold between work and a toddler at home. Twitter DMs are the next input to wire in.
@nikhilaravi [OpenClaw]
https://x.com/nikhilaravi/status/2107146119147135289
nikhilaravi shut down an OpenClaw setup to use Muse only, and the post is a one-line cost comparison: about $32 a month for a DigitalOcean droplet that hosts OpenClaw and needs updates and manual maintenance, against $16 a month for Muse. The author calls it a no-brainer, and it is the cleanest single data point in a day full of OpenClaw migration posts.
@treeonchain [OpenClaw]
https://x.com/treeonchain/status/2107141417550193015
treeonchain admits to spending $700 on a Mac mini eight months ago for an OpenClaw agent, using it for two months and not touching it since. The post has no setup details, but it is the hardware-side version of the migration story and drew a lot of agreement in the replies.
@m0xt_ [OpenClaw]
https://x.com/m0xt_/status/2107142523185721742
m0xt_ moved all operations onto Grok Bot after hopping between OpenClaw, Hermes and Claude for most of the year. The reasoning is specific: on Claude it is hard to spin out a team of agents that each do one job and pass work along, and the OpenClaw and Hermes agents felt less capable and buggy, while on Grok Bot a request goes to a chief of staff and simply happens. The author cites the Latent Space comparison that Grok Bot feels like unboxing a MacBook and OpenClaw feels like Linux, and notes that since the setup is mostly plain files the lock-in cost is low.
@sola_vrc [OpenClaw]
https://x.com/sola_vrc/status/2106910392266133877
sola_vrc, who contributed to OpenClaw more than once, gave up on it recently because it broke too easily and the maintenance cost was too high. The replacement is Dots, and the author is explicit about the trade: accept the context lock-in risk in exchange for a fully managed service, which on its own was worth it.
@JulianRaxworthy [OpenClaw]
https://x.com/JulianRaxworthy/status/2107134137706709109
JulianRaxworthy is using Grok Bot and Dots to do what their OpenClaw bots did six months ago, mainly because the new ones are less brittle, and the Meta and Microsoft news made the author want to go back and rejuvenate the OpenClaw setup, since the project has shipped a lot of features in the meantime. It is the counter-current in the migration thread: people who left for stability and are now weighing whether to return.
@suyuan1711 [Claude Code]
https://x.com/suyuan1711/status/2107083405296795724
suyuan1711 describes a setup someone used to give one Claude subscription to a team of dozens without getting banned: a Mac mini in the office logged into Claude Max x20, Claude Code bridged to a Feishu bot through a skill, and every colleague given permission to talk to the bot in their own isolated session. From Anthropic's side it looks like one device and one user, with no reverse proxy. The post names two mature bridges, lark-coding-agent-bridge, which gives every private chat and group its own session with an invite command, and cc-connect at 15k stars, which covers Feishu, DingTalk and Slack for both Claude Code and Codex. The recommended first step is to hand the link to Claude Code and ask it to wire itself up.
@rewind02 [Claude Code]
https://x.com/rewind02/status/2107086047985258639
rewind02 built a small playable game in Unreal Engine with Claude Opus 5.5 working inside the running editor, with no blueprints by hand. The setup is Unreal 5.8 and a Third Person project under source control, Unreal MCP with all toolsets enabled, and Epic's free Claude Code plugin started in the project folder. The build order is greybox first, a path, a walled village, a visible tower, then the rules, three orbs, a locked gate, a win message and restart key, then one enemy, health and attack one feature per prompt, and only then real assets and lighting. Every prompt follows four rules: say what done means, say what must not change, ask for a list of everything touched, and no silent workarounds.
@yonsan434343 [Claude Code]
https://x.com/yonsan434343/status/2107023395372056689
yonsan434343 released a Shiren-style roguelike set in a steam city, where the mechanic is to stop runaway clockwork beasts and repair them later so they move into town and fight for you. Both the game and its trailer were made with generative tools: game development plus trailer capture and editing in Claude Code, the thumbnail and clock art from Codex image generation, gear animation from MiniMax H3, and music from Lyria 3 Pro. The trailer's gameplay footage was produced by running the actual game automatically and recording it frame by frame.
@nakazakifam [Claude Code]
https://x.com/nakazakifam/status/2107208767305601135
nakazakifam reproduced the fast-cut typography style of the YouTube channel NextCore as a Japanese homage, and the pipeline is the interesting part. The brief was deliberately rough, mimic the way text appears and do whatever with the footage, and Opus 5.5 picked a psychology topic on its own, the illusion of explanatory depth. Claude Code wrote the structure, script and code; visuals were rendered with Node.js and a Skia canvas; narration came from Gemini TTS, stills from GPT-Image through Codex CLI, one cafe scene from MiniMax; subtitle timing from Soniox, a read-back check with local Whisper; and the background music was synthesised note by note in NumPy and SciPy, then mixed and loudness-normalised with FFmpeg.
@sammyo_official [Claude Code]
https://x.com/sammyo_official/status/2106983853948842264
sammyo_official posted a full credits list for a music video where Claude Code wrote the lyrics, Suno composed, Nano Banana Pro through Higgsfield generated the source stills, Kling 3.0 and Seedance 2.5 through the Higgsfield API generated the footage, and the edit was done in After Effects operated by Claude Code. The After Effects control is the unusual piece and the author points to a guide on how to drive it from Claude Code.
@masahirochaen [Claude Code]
https://x.com/masahirochaen/status/2107080438158442535
masahirochaen makes two or three decks a day and iterated a home-made slide tool in Claude Code until it beat anything else they had used, with attention to design, ease of editing and collaboration. The demo video was also made with Claude Code. The post is short on internals but it is a clear case of a non-engineer building their own daily-driver tool because the workload justified it.
@TheObiLeonard [Claude Code]
https://x.com/TheObiLeonard/status/2106999623931891920
TheObiLeonard shipped a website, an admin dashboard, a gamified Mission Control interface and a user dashboard in four days without opening Figma once, with everything built in Claude Code and Codex. There is no breakdown of the split between the two tools, but four surfaces in four days with no design step is a concrete throughput number.
@tikhbana [Claude Code]
https://x.com/tikhbana/status/2106926193006059702
tikhbana sat next to someone on a flight who spent most of the journey in Claude Code and turned out to be a motion designer, not a software engineer. The conversation covered how the tools of the trade, Figma and After Effects, are now driven through agents, and the most striking detail was that the designer's current workflow had been created the previous week. The author's takeaway is that constantly rebuilding how you work is now the job.
@oliviscusAI [Claude Code]
https://x.com/oliviscusAI/status/2107042468852117742
oliviscusAI shared an AGENTS.md that targets three habits of Codex and Claude Code after months of daily use: writing backward-compatible code for a project nobody uses yet, keeping old code and docs it no longer needs, and adding a pile of tests to change a button colour. The core rule is that if the project has no real users yet, skip compatibility and migrations and build the simplest thing that works, then follow docs first, tests, smallest implementation, clean up.
@matiasdev_ [Claude Code]
https://x.com/matiasdev_/status/2107196994901151989
matiasdev_ picked one idea out of the harness engineering paper as the useful one: Hermes's verify-on-stop, which refuses to let a turn close if it touched code and there is no fresh verification evidence. The author has something similar wired up with Stop hooks in Claude Code and says it noticeably changed the quality of what gets delivered.
@BenENewton [Claude Code]
https://x.com/BenENewton/status/2106946540174492158
BenENewton noticed that Claude Code subagents inherit the parent model, so a subagent sent to count Swift files runs on Opus. The fix was a mod that asks Jev to pick the agent and model for every subagent call. On lookup tasks it cut cost by 51 percent and every fact still checked out, and it took three rounds to get there.
@gxcsoccer [Claude Code]
https://x.com/gxcsoccer/status/2107253592809984102
gxcsoccer's take on Claude Code mods is that code runs inside the session, sees every streamed token, draws its own UI, adds commands and hot-reloads live, and you do not even write it, you say make a mod that and Claude does. The author's own mod is a taxi meter that runs during the session.
@Voxyz_ai [Claude Code]
https://x.com/Voxyz_ai/status/2107184044228452491
Voxyz_ai recommends adding a cache countdown to the Claude Code status line and explains why: every message resends the whole conversation, the cache lets the model skip reprocessing what came before, and the first message after expiry reprocesses everything from scratch. The status line shows minutes left and, once expired, how many tokens the next message will re-cache. The post also surfaces documented details many users miss: on a subscription the main conversation cache lasts an hour while subagents get five minutes, on an API key or usage credits the main conversation drops to five minutes too, switching models mid-session recomputes everything and with opusplan every move into or out of plan mode counts as a switch, editing CLAUDE.md mid-session does not break the cache but also does not apply until clear, compact or restart, and rewind can pick up an existing cache while compact builds a new one.
@yifanxu_ephai [Claude Code]
https://x.com/yifanxu_ephai/status/2106960305490387390
yifanxu_ephai observed that apply_patch and the edit tool are close to dead in practice: of seven files written in Codex, only one went through apply_patch and the rest were exec_command plus a Python script, and in Claude Code the day before, Opus 5.5 was likewise mostly calling bash plus Python to rewrite files. It is a small window into how the models actually use the tools they are given versus how the harness designers intended.
@morganlinton [Claude Code]
https://x.com/morganlinton/status/2107227644454711314
morganlinton has been enjoying cloud computers in Claude Code as a streamlined way to move from the desktop to the phone, slower than running locally but with plenty of use cases. The other note is that Opus 5.5 on low effort has become the go-to for planning because it is detail-oriented and pleasant to talk to. The closing thought is that computers are becoming thin clients.
@shi3z [Claude Code]
https://x.com/shi3z/status/2107107346443100472
shi3z was drinking tea in Golden Gai, thought up a new game for an AI arcade, and fed the prompt to Claude Code from the bar through nagi, expecting the first prototype to be finished by the time they got home. It is a short post but a clean picture of the idea-to-prototype gap collapsing to the length of a train ride.
@kawapo_jp [Claude Code]
https://x.com/kawapo_jp/status/2107126533001494612
kawapo_jp asked Claude Code with Opus 5.5 on medium for a realistic, high quality Japanese garden in Three.js and got one in roughly 20 minutes, impressive enough that the author did not expect it to be built unattended. The practical note is that the MacBook Air ran hot, and the scene became heavy once water came in.
@SterlingCooley [Claude Code]
https://x.com/SterlingCooley/status/2107248010904375742
SterlingCooley had Opus 5.5 produce an explainer video on 39 Hz ultrasound straight out of the Claude Code app, using a combination of Python and ElevenLabs, and describes the session as very hands-off. It is one of several posts in the window where video came out of the terminal rather than an editor.
@TDataScience [OpenClaw]
https://x.com/TDataScience/status/2107119124413350187
TDataScience shared a Samir Saci piece in which an OpenClaw agent runs a dozen supply chain personas that track air-freight orders from a Milan warehouse and explain late deliveries. It is a rare operations-side OpenClaw case on a day when most OpenClaw posts were about leaving it.
@dango333 [OpenClaw]
https://x.com/dango333/status/2107227251490295958
dango333 listed the automations already running: AI news research delivered to OpenClaw at 5:30, senior AI reskilling research delivered to Opus at 6:00, Threads post drafts delivered by an AI employee, and cooking videos produced and auto-posted to three social networks. The post asks what other automations people recommend.
@Youssofal_ [OpenClaw]
https://x.com/Youssofal_/status/2107094198918799414
Youssofal_ switched an OpenClaw agent's 24-hour morning brief from Opus to Astra, with subagents to save money, and posted the first day's output side by side with the Opus version. The comparison itself is in the image, but the post is a concrete example of the model-per-task swapping that OpenClaw users do to control cost.
@0xTone [Claude Code]
https://x.com/0xTone/status/2107183017164947752
0xTone uses Hermes as the harness and has it orchestrate the Claude Code agents it spawns, both powered by a single Claude Max subscription, running on a Hetzner VPS and operated mostly from mobile and Telegram. The post calls the setup cozy, and it is a tidy example of one harness driving another.
@itsdidijoy [Claude Code]
https://x.com/itsdidijoy/status/2107146422696984762
itsdidijoy gave the same job in the same words to Grok Bot and Claude Code: alert when Bitcoin closes a four-hour candle above the high of the last 20 candles, paper only. Grok Bot found that its email alerts were down, asked for Telegram or a webhook, then switched to Kraken price data and saved a check every four hours. Claude Code wrote a 90-line Python script, tested it on fake candles, asked before running it, then pulled Binance data. Both landed on the same answer, Bitcoin 3.09 percent below a breakout. The author's rule from the test: Grok Bot keeps watching on its own, Claude Code hands you the tool and you decide where it runs.
@Yerocode0 [Claude Code]
https://x.com/Yerocode0/status/2106992425587618096
Yerocode0 tells the story of someone who left a bot running overnight expecting to wake up at zero and instead went from 300 to 14,000, having written no code: they took the 5min-btc-Polymarket repo, spent a couple of hours cleaning it up with Claude Code and hit run. The strategy enters Polymarket's five-minute BTC markets with about two minutes left, and if the price has already moved 70 to 100 dollars it buys the side of that move between 0.80 and 0.99 and waits for settlement at 1, hedging only if the book goes to 95/5. The author's own warning is the most useful part: this is not a system, it is a market that currently pays dearly for slowness, the edge is in not touching it when fear kicks in, and the day the market stops paying it becomes a hole.
@Mikadzyki_NFT [Claude Code]
https://x.com/Mikadzyki_NFT/status/2107194797546987925
Mikadzyki_NFT summarised the Lex Fridman episode in which DHH, who criticised AI agents and only installed Claude Code half a year after release, said that for the last two months of Omarchy 4.0 a machine produced 100 percent of the code and DHH never looked at it. The method runs against common advice: keep the task vague, delete CLAUDE.md and AGENTS.md or cut them by 80 percent, and hand all coding to agents. What DHH keeps is using the software and judging the result, through what the post calls differential evaluation, asking for three versions and picking the best. DHH also admits the approach gets noticeably harder on Basecamp and HEY with years of legacy code.
@yhslgg [Claude Code]
https://x.com/yhslgg/status/2106961380456996971
yhslgg summarised a widely shared Chinese guide to the National Day ban wave and stripped it to its thesis: risk control is a score, not an IP check, so a residential line or an aged account alone does not protect you, 11.4 million accounts were banned in the first half of the year and roughly a tenth of appeals succeeded. The common failure modes are many accounts on one device, a new account opening straight on Max 20x, virtual cards on the web, subscriptions wired to third parties and long sessions on shared nodes. The 11-step floor includes a fixed timezone, a single dedicated residential IP, TUN plus global routing, leak checks, no leftover proxies, one account per machine, in-app purchase over cards, climbing tiers slowly, and for Claude Code specifically the official client only, telemetry off per the docs, no binary edits, and no base URL pointed at a relay.
@gtmom [OpenClaw]
https://x.com/gtmom/status/2107162373287010460
gtmom, self-described as basically nontechnical but willing to play heavily, describes the mess that followed Cowork's launch, projects and documents created everywhere, some local and some in Obsidian vaults syncing across machines. Last week the fix was to take the Mac mini formerly used for OpenClaw and Hermes, log it into Claude and Perplexity, run the automations there, and control Claude Code from the phone. The author doubts this will make sense for most people.
@olanetsoft [Claude Code]
https://x.com/olanetsoft/status/2107032322583216211
olanetsoft lost track of agent sessions for months, had something local that solved it, and after seeing someone else ask for the same thing cleaned it up and open-sourced it as Agent Lookout: one local page for all Claude Code and Codex sessions that shows which Claude Code sessions are waiting on you. Support for more agents is next.
@hkdom [OpenClaw]
https://x.com/hkdom/status/2107117009763029261
hkdom has been running multi-agent setups since the OpenClaw boom in early 2026, when different machines' OpenClaw instances talked on a message board or Discord, then Hermes profiles with its built-in cross-agent kanban, peaking at more than ten agents before settling at four or five. The current tool is Herdr, which lets each vendor's own harness or an open agent run its own model and coordinates them without ACP or A2A plumbing. The day's experiment compared DeepSeek's new desktop harness, Manus 2.0 and Herdr with seven models on the same hour-long task from a PDF and a few sites, and the author was most surprised by the depth of the DeepSeek harness on Thinking Max. The conclusion is that more agents is not necessarily better, and a good model with a good harness does not have to be Opus 5.5.
@i_love_profit [Claude Code]
https://x.com/i_love_profit/status/2106927321131471148
i_love_profit's assessment of Claude Code is that its business judgment is still poor, but once the human makes the one decision that matters it has more than enough capability to execute it mostly the right way. The author puts it at the level of a business manager and says $200 a month feels like hiring one person at that level, and is about to sign up for a third account.
@dotey [Claude Code]
https://x.com/dotey/status/2107144243726729645
dotey studied amontlabs/lcu, which extracts Codex's Computer Use so Claude Code, Codex CLI and Pi can call it. It is not a reimplementation: it finds the installed ChatGPT desktop app, launches OpenAI's own computer-use runtime inside it, and exposes it to your agent over MCP, so the official desktop app must be installed but no ChatGPT or Codex login is required. The model gets only two tools, js and js_reset: it writes JavaScript against a preloaded cua object, calling getState for the desktop state and instructions, getApp to pick a window, then click, typeText, pressKey and getScreenshot in one call, with state preserved between calls and accessibility tree access so it is not purely coordinate-driven. Repo notes say the init instructions match official Codex byte for byte on macOS; Apple Silicon macOS and X11 Linux are supported, Chrome control goes through OpenAI's extension with per-site approval.
@shields_pikes [OpenClaw]
https://x.com/shields_pikes/status/2106939410675900427
shields_pikes recalls the period when OpenClaw could no longer use the Claude Code subscription allowance, switched development tasks to the API, and ran up roughly 30,000 yen in a single day. The post frames it as a fond memory now that subscription access has been reinstated and Codex subscription quota is also usable, but it is a concrete number for what always-on agents cost at list price.
🗣 User Voice
User Voice
Local is the last line, and people are drawing it around Claude Code. The Cowork cloud-only change produced the same sentence from several directions: @CrazyAITech said choosing a local folder now uploads files and caps them at 100, so a loyal Cowork user is forced into Claude Code; @hammer_mt pointed out that Cowork existed precisely to give normal people Claude Code with the scary parts removed, and @paradite_ asked why anyone needs Grok Bot, Instinct or Dots when Claude Code already runs everything expressible as logic.
The harness changes the model more than the model does. @intellectronica ran the same Opus and Sonnet 5.5 through Pi, OpenCode and Copilot and found Claude Code barely usable by comparison, @kgsi got clearly different document quality asking Gemini through Claude Code versus the Antigravity CLI with identical skills, and @vincit_amore described the CLI Opus as a wizard and the app Opus as something to regret engaging with.
Permission modes are either too loose or in the way. @kr0der says auto mode refuses too many normal tasks and stops to ask approval for work that was just approved, @jaivinwylde mocks anyone still in ask-permission mode, and @AriZerner is weighing handing Claude Code a DigitalOcean API key on the box it runs on.
Nobody can tell which agent owns a task. @gocronk wants to know, after a handoff between agents, which one owns the job and which one will report back; @awakecoding cannot prompt the popup that splits discovered work into its own worktree session; @PracticalAIbook wants voice mode or read-aloud in Cowork and Claude Code sessions on mobile.
OpenClaw's maintenance tax is now the stated reason for leaving. @senamakel timed a simple search opening a full browser for over ten minutes, @JacksonW29657 says updates left almost nothing working, @banjtheman has five personal agents asking for Gmail and nothing useful for any of them to do, and @xdotli's ranking calls OpenClaw rough before 2.0 and quiet after, with a mobile app that never got updated.
Local is the last line, and people are drawing it around Claude Code. The Cowork cloud-only change produced the same sentence from several directions: @CrazyAITech said choosing a local folder now uploads files and caps them at 100, so a loyal Cowork user is forced into Claude Code; @hammer_mt pointed out that Cowork existed precisely to give normal people Claude Code with the scary parts removed, and @paradite_ asked why anyone needs Grok Bot, Instinct or Dots when Claude Code already runs everything expressible as logic.
The harness changes the model more than the model does. @intellectronica ran the same Opus and Sonnet 5.5 through Pi, OpenCode and Copilot and found Claude Code barely usable by comparison, @kgsi got clearly different document quality asking Gemini through Claude Code versus the Antigravity CLI with identical skills, and @vincit_amore described the CLI Opus as a wizard and the app Opus as something to regret engaging with.
Permission modes are either too loose or in the way. @kr0der says auto mode refuses too many normal tasks and stops to ask approval for work that was just approved, @jaivinwylde mocks anyone still in ask-permission mode, and @AriZerner is weighing handing Claude Code a DigitalOcean API key on the box it runs on.
Nobody can tell which agent owns a task. @gocronk wants to know, after a handoff between agents, which one owns the job and which one will report back; @awakecoding cannot prompt the popup that splits discovered work into its own worktree session; @PracticalAIbook wants voice mode or read-aloud in Cowork and Claude Code sessions on mobile.
OpenClaw's maintenance tax is now the stated reason for leaving. @senamakel timed a simple search opening a full browser for over ten minutes, @JacksonW29657 says updates left almost nothing working, @banjtheman has five personal agents asking for Gmail and nothing useful for any of them to do, and @xdotli's ranking calls OpenClaw rough before 2.0 and quiet after, with a mobile app that never got updated.
📡 Eco Products Radar
Eco Products Radar
Codex: the default second agent in nearly every workflow post, from Figma-to-site builds to review loops.
Opus 5.5: the model behind most of the day's long-running builds, with effort settings now a tuning topic.
Hermes: harness of choice for people orchestrating Claude Code from Telegram, and the verify-on-stop idea others copy.
Muse: the destination in most OpenClaw migration posts, usually with a price comparison attached.
Dots: OpenAI's always-on agent, used as a commander over other AI employees or as a fully managed OpenClaw replacement.
Grok Bot: the other migration destination, praised for the chief-of-staff experience and brand-per-bot setups.
Jev: the small decision model showing up inside Claude Code mods to route subagents and pick models.
Cowork: the day's lightning rod after the cloud-only change, pushing users back to Claude Code.
Mac mini: the always-on agent box that people are now either repurposing or admitting they stopped using.
Obsidian: the shared memory and task-inbox layer in several multi-agent setups.
Pi: cited as a lighter harness that makes Claude Code look heavy, and as a target for computer-use bridges.
MCP: the plumbing in almost every integration mentioned, from Unreal to computer use to Feishu bridges.
Codex: the default second agent in nearly every workflow post, from Figma-to-site builds to review loops.
Opus 5.5: the model behind most of the day's long-running builds, with effort settings now a tuning topic.
Hermes: harness of choice for people orchestrating Claude Code from Telegram, and the verify-on-stop idea others copy.
Muse: the destination in most OpenClaw migration posts, usually with a price comparison attached.
Dots: OpenAI's always-on agent, used as a commander over other AI employees or as a fully managed OpenClaw replacement.
Grok Bot: the other migration destination, praised for the chief-of-staff experience and brand-per-bot setups.
Jev: the small decision model showing up inside Claude Code mods to route subagents and pick models.
Cowork: the day's lightning rod after the cloud-only change, pushing users back to Claude Code.
Mac mini: the always-on agent box that people are now either repurposing or admitting they stopped using.
Obsidian: the shared memory and task-inbox layer in several multi-agent setups.
Pi: cited as a lighter harness that makes Claude Code look heavy, and as a target for computer-use bridges.
MCP: the plumbing in almost every integration mentioned, from Unreal to computer use to Feishu bridges.
Comments