Super User Daily: 2026-11-09
The most interesting Claude Code work today came from people who don't write software for a living. Someone who isn't a furniture designer drew a table part by part and got a bill of materials that undercuts a premium brand by 85%. Someone else cleared 400 GB off a Mac with one prompt, and a third person read the revenue of every startup on TrustMRR in eight minutes. The second thread is a quieter one: less scaffolding keeps winning. The ISUCON 2026 winner threw away three months of hand-built multi-agent harness because the bare agent scored higher, and a designer shipped a logo animation in under an hour after deleting six versions that took all day. The loudest worry in the replies is privacy, after a report that Claude Code was putting the account email into the User-Agent of its own web requests.
@levelsio [Claude Code]
https://x.com/levelsio/status/2119132202952798369
Claude Code designed a dining table part by part and then produced a bill of materials so each piece can be ordered from suppliers and assembled at home. The table was drawn by @theannalux, who is neither a furniture designer nor a coder and simply had a style in mind and worked in the terminal. The post argues that with Opus 5.5 driving free tools like Blender and FreeCAD, buying raw materials directly saves about 85% against brands like Poliform, so a $30,000 sofa becomes $4,500 and an $8,000 table becomes $1,200. It is the clearest case today of an agent cutting out a middleman in a physical market, not a software one.
@srt_taka [Claude Code]
https://x.com/srt_taka/status/2118892700179276029
A recap of a podcast episode on how kawasy won ISUCON 2026: on the day, the human only logged into the portal, and a single Claude Code session ran on its own for about eight hours and took first place. The three months of prep are the interesting part. A four-phase multi-agent harness stuffed with human best practices was built first, then in late September a nearly bare agent was rerun and scored far higher, so the whole harness was thrown out and only thin fixes were added back where problems showed. Evaluation used seven problems, five past ones, one practice and one self-written, with the self-written one kept as the uncontaminated check for generalization, and a Stop hook that tells the agent how many minutes remain and what to do next fixed its habit of declaring victory after two hours.
@shinmomo0512 [Claude Code]
https://x.com/shinmomo0512/status/2119215043753717791
A 45-second vertical fan PV for Gesshoku Kitan in which all 34 characters burst out of gold light at the song's peak and parade past the camera, made without ever opening the Blender interface. Claude Code wrote, rendered and checked Python scripts that drive Blender, and the first instruction was a single line; it asked back three questions, how many characters, how long and which song, and those plus whether to render and whether to fix were the only human decisions. The preview took about 13 minutes and the final render two and a half hours on a Mac Studio. The best detail is the bug: a waving hand was poking into the neighbour's face, a geometry check found 32 overlapping pairs in just two frames because the hand swung out 0.93 m against 0.9 m spacing, and now overlap gets checked by machine before every render.
@alphabatcher [Claude Code]
https://x.com/alphabatcher/status/2119110041596747968
One job for Claude Code: read TrustMRR, a site that pulls revenue straight from Stripe, and it got through all 123 pages and 12,244 startups in eight minutes. Of those, 522 sell an AI agent, 284 of them made $0 in the last 30 days, and the 15 that cleared $10,000 took 84% of everything the 522 earned. Nine of those 15 sell the same thing, more customers, through sales, ads, SEO, social posts or video. Four are one-person companies started this year, including an ad-running agent with 189 followers that did $70,229 last month.
@karpathy [Claude Code]
https://x.com/karpathy/status/2119114537442189652
Tristan published a list of favourite YouTube channels, and it got turned into TristanTube, a browsable site with 47K very nerdy videos and a TV mode for lean-back watching. The whole thing was one-shotted with Claude Code in three minutes. It is a tiny build, but it shows the new default for personal tools: someone shares a list, and a few minutes later it is a product. It was also the most viewed Claude Code post of the day by a wide margin.
@andrew_jennings [Claude Code]
https://x.com/andrew_jennings/status/2119152409595224279
A multi-brand business owner lists the roles replaced this year, and the finance one is the Claude Code piece: reporting was rebuilt so bank feeds and store data go in and daily profit per brand, cash and forecasts come out, and the monthly P&L someone used to email is gone. Around it sit a Claude chat that reads email and calendar and sends a morning brief with sales, cash, ad spend and stock for each brand, a command center where every pending decision lands as a card to approve or skip, and an ops wiki in plain text with a 60-line rules file per agent saying what it can touch and who it asks. Human creators still shoot the videos; only the editing got automated. The structure is the useful part: every agent has written limits, and every decision still goes through one person.
@_aashish_singh_ [Claude Code]
https://x.com/_aashish_singh_/status/2119098894902755819
The Mac's disk kept filling up, so Claude Code was started from the home folder, which the author admits is a risky allowance, with one prompt: list what eats the most storage in a table, sorted by size, with what each thing is for and whether deleting it is safe. The answer was about 100 GB of Docker cache untouched for months, about 100 GB of unused Android emulators and hundreds of GB of Xcode and Android build caches. Claude deleted them safely and freed roughly 400 GB. The author calls it a naive use case that turned out to be a game changer, with one rule: push critical projects to GitHub before you try it.
@ayodeji0x0 [Claude Code]
https://x.com/ayodeji0x0/status/2119107970164236473
An experiment originally tried with Grok Bot was rerun with Claude Code: build and operate a full uptime monitor across eight providers, Fly, Neon, Cloudflare, Vercel, Modal, Runpod, Resend and Supabase, all through their control planes. Claude Code found the tools on its own, deployed, hit errors, fixed them and checked its own work. The human only stepped in for OAuth scopes and token permissions. The demo is live and lets visitors simulate a provider failure, which makes this an operations job done by the agent, not just code written by it.
@Mochi_AGI [Claude Code]
https://x.com/Mochi_AGI/status/2118944760580968706
A dancing lyric video built from GIF animations with Claude Code on Opus 5.5 plus ChatGPT Images 2.6, at no extra cost because the API credits that come with ChatGPT Pro covered it. Claude Code wrote the song through a Suno MCP and the author picked a take by ear, then six nine-frame GIFs were generated at roughly $0.12 each. The craft is in the fixes: GIF frames redraw and shake the background, so each cut gets one fixed empty background; characters are generated on green screen, keyed out and given shadows from their own silhouettes; every lyric character bounces with motion chosen per word. Upscaling happens only after the motion is locked, so nothing expensive gets thrown away.
@camperaigg [Claude Code]
https://x.com/camperaigg/status/2119052253982306393
One prompt asked Claude to ride a lightning bolt from cloud to ground, and Fable 5.5 at xhigh inside Claude Code ran for 182 minutes and $17.57 in API costs without the author touching code. It simulated the storm on a 30.5M-cell grid with 63 logged flashes, grew the bolt with a dielectric breakdown simulation and rode the camera along all 60 steps of the leader. Two self-corrections stand out: it noticed the bolt stalling 752 m above the ground and resumed the sim instead of restarting, and it scanned every frame, found a full-screen strobe at 4 to 6 Hz and fixed it before the author ever saw it. The result is a 28-second unbroken 1080p60 shot in three.js and Taichi.
@marouane53 [Claude Code]
https://x.com/marouane53/status/2119130135186252213
A recommender where 432 shows the author loved vote on what to watch next, packaged as a Claude Code and Codex skill. Under it sits a database of 78,802 films and series from IMDb, TMDB, TVmaze and Wikipedia, plots embedded with a Qwen3 model for meaning-based search, and 32M MovieLens ratings, with every tweak benchmarked against real viewer data and kept only if it helped; Wikipedia premises alone gave up to 15% better matches. The lessons read like election design: superhero shows voted as a bloc so seats became proportional, classics like Mad Men had no recent neighbours so every title gets a full ballot, and franchises get one seat each. Ask for Breaking Bad but lighter and it explains its picks and where to stream them.
@tanakitto [Claude Code]
https://x.com/tanakitto/status/2118933303139168263
After two years of sending out CVs during COVID, the author built a personal job hunting system called Ikigai that scores openings on what you love, what you are good at, what the world needs and what pays. The prototype is Google Apps Script, Google Sheets, n8n and the Gemini API, with Claude Code connected to n8n over MCP so it builds and runs the flows while the author just talks to it. It pulls jobs from LinkedIn and JobsDB alerts, APIs and search, scores four categories out of five with reasons and keeps the rejects, and on Apply it tailors a master CV to the job description using only real experience in an ATS-friendly format. Rejections and ghostings are logged with reasons, and the author is weighing privacy before opening it to anyone else.
@dan__rosenthal [Claude Code]
https://x.com/dan__rosenthal/status/2119082513100407239
A 22-person team where everyone uses Claude Code every day, and the lead's point is that buying seats was the easy part. The layers under it, context, integrations and execution, had to be built once by the org and kept alive by a small dedicated team. The recipe: dump every SOP and spreadsheet in one place, have Claude set up folders like company-wiki and raw-context, connect every tool through MCP or CLI (HubSpot, GitHub, Pinecone, Notion, Instantly, HeyReach), let Claude turn it all into agent-ready markdown, and promote any end-to-end SOP into a skill such as /tam-mapping or /campaign-builder. Data syncs keep the context fresh, skills fire on schedules or through managed agents, and improvements come back as pull requests.
@Cayden970207 [Claude Code]
https://x.com/Cayden970207/status/2118979605499633998
The author uses Hermes, Claude Code and Codex every day and got tired of being the copy-paste layer between them, moving progress across, re-explaining the same error and hoping nobody overwrote the other's files. So the author built Thoth, a Hermes plugin that turns them into one team while each keeps its own harness: decisions flow straight into the next agent's brief, they review each other's work before main, and each gets its own workspace. The numbers from its own logs: 92% of prompt tokens came from cache and 99% of the agents' transcripts stayed out of the coordinator's context. Building the same CRM twice, Thoth used 26% fewer tokens and was 35% faster than handing work between agents by hand.
@Pluvio9yte [Claude Code]
https://x.com/Pluvio9yte/status/2119080693233754506
A hands-on switch to Vercel Labs' open-source agent-desktop, which lets Claude Code or Codex operate desktop apps by reading the accessibility tree instead of guessing click coordinates from screenshots. It doesn't move your mouse or pull windows to the front, so the author kept chatting in WeChat the whole time without losing focus once. Timings on a Mac: reading the whole System Settings window in about 0.3 s, clicking Bluetooth in 0.17 s, typing mixed Chinese and English into a background text editor in 0.15 s with an exact readback, and Electron apps were readable too. The previous tool depended on a runtime inside the ChatGPT desktop app and broke on its updates, so it was uninstalled the same day.
@initcore [Claude Code]
https://x.com/initcore/status/2119198891086385568
Claude Code built a one-million-zombie scene in a single session: each zombie has its own GPU-simulated steering, crowd avoidance and ragdoll flight when hit. Bullet time comes from a simulation clock written into the shader rather than from slowing the video afterwards. The 12-second 1080p clip rendered in 21 minutes on one RTX 5090. It was a first spin of a tool called vork, and it shows how far a single session can now push real-time graphics work.
@Lummox_eth [Claude Code]
https://x.com/Lummox_eth/status/2119027995239194880
The same prompt, a cinematic lighthouse in a storm, went through four AI workflows with the bill and the clock recorded. Blender MCP took $67 and 20 minutes and delivered the most cinematic scene, GPT-6 Astra $40 and 11 minutes, Kimi K3.1 $12 and 3 minutes, and Claude Code $102 and 28 minutes, which makes Kimi 8.5 times cheaper and 9.3 times faster. The author ran a second prompt, a desert campsite, and Claude Code again came last on price at $104 with the most cartoonish look. Two runs is not a benchmark, but it is a useful reminder that for 3D scenes the tool driving Blender matters more than the model's general coding strength.
@doooyle [Claude Code]
https://x.com/doooyle/status/2118862321170743641
A founder building a marketing factory added a PR skill: find articles in the space, check where the founder's own data overlaps, and pitch the journalist on including it. On a creator economy story, Claude Code pulled numbers on who is increasing ad spend and pitched the charts. Eleven pitches produced one positive reply, and Claude then built out the supporting page and drafted the response. The author is candid about the downside: sharing tactics like this speeds up how fast every tactic turns to slop.
@Mrhage28 [Claude Code]
https://x.com/Mrhage28/status/2118905027167453578
A blogging instructor with more than ten years of experience is growing a history site from scratch with Claude Code, publicly, with the goal of passing a million monthly page views. October numbers: 1,445 search clicks and 4,208 page views, no revenue yet. The milestone the author cares about is breaking 100 views a day and nearly clearing 200, which in their experience means 1,000 a day is close, helped by demand around the year's historical TV drama. The promise is to post wins and accidents alike, with numbers.
@wy_ones [Claude Code]
https://x.com/wy_ones/status/2118859311673983388
After Thursday's US market drop, the author went through every explanation that could be found and turned the process into a prompt to hand Claude Code or Codex next time there is a selloff. Rule one: browse and open the original reports instead of trusting search snippets, and write not found rather than answering from memory. Rule two: ask about sectors and the index separately; that day the Philadelphia semiconductor index fell 12.25% against 4.36% for the S&P 500, and the author estimates about 60% of the drop came from tech. Rule three: recheck every intraday conclusion after the close, because two of the author's own midday conclusions flipped.
@jilaiqian [Claude Code]
https://x.com/jilaiqian/status/2119118276089245782
A week of switching agent workspaces, herdr to Orca to herdr-app and finally T3 Code, with an honest note on each. herdr is tmux for agents, minimal and fast, but has no phone access and stays limited to what the Claude Code or Codex harness offers. Orca is the hyped Swiss army knife, artifacts, subscription switching, mobile, GitHub and Linear, and the author says it works very well but felt too heavy. herdr-app added customizable layouts and plugins, and the author built one: a git panel with a commit graph and staged and unstaged changes.
@KameAIHacks [Claude Code]
https://x.com/KameAIHacks/status/2118864829389619596
A walking turtle robot video that is not generative AI video: Claude Code wrote all 1,227 parts as code, counted down to individual screws and washers. The legs are computed every frame so it actually walks, and the sound is synthesized to match the motion. The only AI image was a single reference drawing, the base was a free skill called teardown, and no paid API was used. The author wants to use the same approach for product explainers and kids' teaching material.
@Phavorism [Claude Code]
https://x.com/Phavorism/status/2119092001434223008
A full day with Claude Code trying to get dust to spiral into the Elysium logo for @SmartXTerminal. The dust kept circling and then snapping into the logo instead of forming it, and even at version six it still looked crowded. So everything was thrown out, and the version that shipped took under an hour. It is the same lesson as the ISUCON win in miniature: past a point, iterating on a bad base costs more than starting clean.
@EngincanVeske [Claude Code]
https://x.com/EngincanVeske/status/2118964973040701800
A working pattern for big tasks: instead of letting one session do everything and fill its context, the first thing the author asks is for Claude Code to spin up the sub-agents the job needs, research, coding, image generation, bug checking, each in its own context window that only sends back a result. The main session stays the coordinator, turns a finished research pass into a plan and spawns the next agents from it, and the author steers the whole flow with /btw while they run. With more than one subscription, each agent can get a different model, Opus for coding, Sonnet for bug checks, Gemini 4 Argon for research, and Gemini or Codex for images.
@tifoden [Claude Code]
https://x.com/tifoden/status/2118978653979562115
Grok Bot's usage limits make it a poor place for heavy engineering, so complex coding and PRs stay in Claude Code or Codex, and the gap was context: the bot could only see what landed in a PR or in git, not the tests tried or the work in progress. The fix was to run both personal coding and delegated coding inside Devin cloud sessions, and give the bot the Devin API so it can look up sessions and send instructions. Now the bot can pick up work the author started, the author can take back anything delegated, and attempts made before a PR are captured in the session history. Because the sessions live in the cloud, the Mac can be off and there is no approval card on every handoff.
@OkaTakuma1 [Claude Code]
https://x.com/OkaTakuma1/status/2118900828795011190
A cat trivia short, why cats slow-blink, made end to end by Claude Code: script, footage gathering, AI narration, subtitles and the edit are all automated. The first versions looked cheap, and it took several rounds of improvement to get to the posted one. The author admits the voice still sounds a bit like AI. The interesting part is that a single hobbyist now runs a whole short-video production line, with the only remaining weak spot being the voice.
@_nogu66 [Claude Code]
https://x.com/_nogu66/status/2118864503513436193
The author audited their own prompt collection against current models and got 141 flags, using the prompt-audit check inside Claude Code. The most common problems were outdated facts and skills that conflict with each other. The takeaway is the uncomfortable one: insurance instructions written to protect against older models' weaknesses are now actively holding the new model back.
@illyism [Claude Code]
https://x.com/illyism/status/2119097342347624571
Most MCP servers only wrap GET requests, dump 20,000 tokens of JSON into the chat and still leave you to open ten browser tabs to do the work. So the author upgraded five products with full read and write remote MCP servers, from 7 to 14 tools each. The design choice worth copying is a two-step spend gate: the backlink tool previews the exact quote first and only charges the saved card when the agent passes an explicit confirmation flag. Inside Claude Code that means checking SEO numbers, approving publisher sites and placing an order without ever leaving the terminal, but never paying by accident.
@turingou [Claude Code]
https://x.com/turingou/status/2119096744051130515
Several vertical personal agents the author is building all need image and video generation, so it had to be solved at the platform layer. Hosted APIs like fal are bound by commercial and cost rules, and every image and video service seems to draw a different baseline for what it will freely generate. Claude Code's approach is to handle it at the director level with scripts and pre-edits, so results stay highly consistent whether the model is a ComfyUI workflow or a hosted API. The catch is that this benchmark needs a local machine where a director model reviews output over and over, which leads to the conclusion that for large-scale, consistent long-form video, renting cloud GPUs is the only real option.
@RealYDT [Claude Code]
https://x.com/RealYDT/status/2118895733512572956
After reading Anthropic's engineering post on agent memory and dreaming, the author had Claude go through their own MEMORY.md: merge duplicates and move the change history piled up in the index back into the topic files. The file went from 12.5 KB to 10.7 KB. It matters because Claude Code's auto memory only loads the first 200 lines or 25 KB as an index, so anything below that line is invisible. The author's analogy: a new hire saves every email, a veteran keeps one sticky note on the desk and looks the rest up.
@liorbuilds [Claude Code]
https://x.com/liorbuilds/status/2119170764826087718
One habit that made Claude Code better inside an Obsidian vault: a single note of learned rules, capped at 30. Every correction goes in, but if a matching rule already exists it gets sharpened instead of a new one being added. Claude reads the note before bigger tasks. The cap is the clever part, because it forces the list to stay short enough that the model actually follows it, and the author reports fewer repeated mistakes.
@naw103 [Claude Code]
https://x.com/naw103/status/2119060304336204019
A small, real annoyance: delete a Claude Code routine's old session files from disk and the desktop app keeps listing them as Session not found on disk, because it holds the Runs list in memory. So the author built a Routine Cleanup skill that plans the cleanup in a script and then deletes through the app rather than behind its back. It is open source.
@thorstenball [Claude Code]
https://x.com/thorstenball/status/2118950540021203169
A pushback on the claim that you can't run Postgres or MySQL with Claude Code and that remote coding is still painful. The author's team runs its full stack on remote boxes called orbs: several dev servers and services, Postgres, Redis and Docker. The argument is that once you are no longer typing code yourself, you no longer need LSPs or an editor on the remote machine, and configuring services or getting headless browsers working is now easy with an agent. The advice: put your code on a remote box, headless, and ask an agent to run through it end to end.
@cfenghub [Claude Code]
https://x.com/cfenghub/status/2119212049569890357
A pixel-style black hole and nebula animation made entirely in the terminal with Claude Code, with no video editor or timeline tool involved. Opus 5.5 wrote a single Python script that draws the accretion disk, galaxy and starfield frame by frame, and the 8-bit chiptune soundtrack is also generated by code instead of coming from a sound library. Claude rendered previews, spotted visual glitches, debugged the script until the frames passed, and then muxed the final video with sound. The author's estimate is that a game intro or channel opener like this now costs little more than electricity and a few dollars of API.
@MariusDevo [Claude Code]
https://x.com/MariusDevo/status/2119176507247690126
TreeBoss puts one coding agent in charge of the others: Claude Code opens Cursor and Codex sessions in the middle of its own turn, each in its own git worktree, then reviews their diffs and merges them. It runs on the subscriptions you already pay for, and it is free and source-available. It is one more entry in a crowded category this week, tools that make Claude Code the manager rather than one of the workers.
@sortofadev [Claude Code]
https://x.com/sortofadev/status/2118885396918735297
A few months ago the author did not know what a terminal was, and last month a rental car platform built with Claude Code, GPT and T3 Code did just under $4k in MRR. It was built mostly from 9 to 11 pm after the family went to sleep, alongside a day job as COO of a marketing agency. The author counts about 1,100 merged PRs so far and credits pulling from open-source projects on GitHub. It is the clearest non-developer-to-revenue story in today's data, with the time cost stated plainly.
@AiAircle34052 [Claude Code]
https://x.com/AiAircle34052/status/2118878610195120234
A recap of Iron Curtain Call, a fast single-player FPS in a 1930s cartoon style made by one developer with a team of AI agents on Claude Code and the free Godot engine. Development started on September 29 and a free demo was on Steam by October 23, with 4,558 commits, almost none of the code written by hand according to the developer. ElevenLabs supplies the voices. The store page draws a clear line on AI use: voices, some music, 2D art and cutscenes were made with AI help, and nothing is generated by AI while you play.
@dextune [Claude Code]
https://x.com/dextune/status/2119173045797708268
New motions for a Live2D model without the Cubism Editor, by treating the motion file as what it is: JSON keyframe curves for each model parameter. As long as you stay inside the parameters a model already has, new motions can be written as text with no rigging or mesh work. The repo adds seven motions to the official sample model Hiyori, including wink, nod, thinking and shy, and none were written by hand; a generator script builds each one, an independent validator checks it, and a headless browser confirms it plays. An AGENTS.md lets Claude Code or Codex add a new motion on request.
@satoru_obento [Claude Code]
https://x.com/satoru_obento/status/2118880777454911724
An SEO specialist with 25 years in web marketing, much of it producing content for big newspaper and electronics brands, has put that know-how into Claude Code and is mass-producing content for owned sites. With an SSH connection, site updates run almost on autopilot, and the setup publishes three pieces a day of 7,000 to 10,000 characters each. The bet is that owned media is about to matter again. Volume like that is exactly the kind of output that needs the human expertise feeding it, which is the point the author is making.
🗣 User Voice
User Voice
Privacy is the top worry. @CryptoNews_op lays out how Claude Code loaded the account email into context and the model reused it as a contact address in User-Agent headers, with reports going back months, while @theaisignals points out the stopgap filter misses scripts the model writes itself. A separate report shared by @maddie72keth says passwords and keys approved with Don't Ask Again ended up in files committed to public repos.
People want to know what they are paying for. @BhurkeSiddhesh cites Claude Code on Opus 5.5 using 14.5M tokens and 146 turns per task against 3.2M and 39 for Codex, @NikitaD26506 found 98% of an agent's bill was it re-reading its own context, and @nyxienyxie reads the docs to show ultracode moves the plan out of context rather than saving tokens.
Trust the diff, not the summary. @ShaneBOly's house rule is that when the recap says everything went smoothly and git log disagrees, the code wins, and @HochstatMichael warns that a Claude-built wiki can merge pages badly and then look authoritative, so keep raw sources untouched.
Handoffs between agents are still manual. @StarlingMage coordinates Claude Code, Codex and Antigravity through shared Obsidian notes and says plainly that the human is the relay, while @evvvvince wishes Claude Code had the kind of mid-run recovery other harnesses are shipping.
Parallel agents hit a human ceiling. @stretchcloud notes most people top out at three to six agents on one repo because someone still has to read every diff before it merges.
Privacy is the top worry. @CryptoNews_op lays out how Claude Code loaded the account email into context and the model reused it as a contact address in User-Agent headers, with reports going back months, while @theaisignals points out the stopgap filter misses scripts the model writes itself. A separate report shared by @maddie72keth says passwords and keys approved with Don't Ask Again ended up in files committed to public repos.
People want to know what they are paying for. @BhurkeSiddhesh cites Claude Code on Opus 5.5 using 14.5M tokens and 146 turns per task against 3.2M and 39 for Codex, @NikitaD26506 found 98% of an agent's bill was it re-reading its own context, and @nyxienyxie reads the docs to show ultracode moves the plan out of context rather than saving tokens.
Trust the diff, not the summary. @ShaneBOly's house rule is that when the recap says everything went smoothly and git log disagrees, the code wins, and @HochstatMichael warns that a Claude-built wiki can merge pages badly and then look authoritative, so keep raw sources untouched.
Handoffs between agents are still manual. @StarlingMage coordinates Claude Code, Codex and Antigravity through shared Obsidian notes and says plainly that the human is the relay, while @evvvvince wishes Claude Code had the kind of mid-run recovery other harnesses are shipping.
Parallel agents hit a human ceiling. @stretchcloud notes most people top out at three to six agents on one repo because someone still has to read every diff before it merges.
📡 Eco Products Radar
Eco Products Radar
Codex: the default second agent and comparison point in almost every setup, 167 mentions in the posts read
Cursor: still listed next to both, 68 mentions
Gemini: the research and image model people route to beside Claude, 35 mentions
Grok Bot: the personal agent people delegate to, then move heavy coding off because of limits, 25 mentions
OpenCode: the open-source harness, now with a browser that uses your real logins, 22 mentions
DeepSeek: the cheap model people plug in behind Claude Code, 20 mentions
Hermes: the self-hosted agent that coordinates Claude Code and Codex in several builds, 19 mentions
Blender: driven headless by Claude Code scripts for promos, scenes and furniture, 19 mentions
Muse: the free personal agent many compare against, 17 mentions
ffmpeg: the last step in nearly every code-made video, 15 mentions
Claude Duet: now steers Claude Code sessions from the phone, 12 mentions
ElevenLabs and Seedance: the voice and video layers in agent-made media, 12 and 11 mentions
Three.js, Remotion and T3 Code: code-drawn animation, programmatic video and the multi-agent console, 9, 8 and 8 mentions
Orca: runs a team of coding agents on your own subscriptions, 7 mentions
Codex: the default second agent and comparison point in almost every setup, 167 mentions in the posts read
Cursor: still listed next to both, 68 mentions
Gemini: the research and image model people route to beside Claude, 35 mentions
Grok Bot: the personal agent people delegate to, then move heavy coding off because of limits, 25 mentions
OpenCode: the open-source harness, now with a browser that uses your real logins, 22 mentions
DeepSeek: the cheap model people plug in behind Claude Code, 20 mentions
Hermes: the self-hosted agent that coordinates Claude Code and Codex in several builds, 19 mentions
Blender: driven headless by Claude Code scripts for promos, scenes and furniture, 19 mentions
Muse: the free personal agent many compare against, 17 mentions
ffmpeg: the last step in nearly every code-made video, 15 mentions
Claude Duet: now steers Claude Code sessions from the phone, 12 mentions
ElevenLabs and Seedance: the voice and video layers in agent-made media, 12 and 11 mentions
Three.js, Remotion and T3 Code: code-drawn animation, programmatic video and the multi-agent console, 9, 8 and 8 mentions
Orca: runs a team of coding agents on your own subscriptions, 7 mentions
Comments