Super User Daily: 2026-06-13
The clearest pattern today: the frontier crowd has stopped asking Fable 5 to do tasks and started asking it to fix the systems that do tasks. One user pointed Fable at a bloated PowerPoint skill and it rebuilt the tool itself; Anthropic's own team edited the entire Fable launch video with zero traditional editing software; and a solo developer migrated a million lines of Zig to Rust in six days with Dynamic Workflows. Off the coding path, agents are running incident forensics, physics research, real-money trading bots, lead-gen businesses and 3D printers. The other side of the coin is everywhere too: token bills, rate limits and the June 22 Fable deadline are now the loudest complaints on the timeline.
@SantiTorAI [Claude Code]
https://x.com/SantiTorAI/status/2065107488501063822
A breakdown of how Jarred Sumner (creator of Bun) migrated 1,000,000 lines of code from Zig to Rust in six days, file by file, with the original test suite passing at the end. The key was Dynamic Workflows: an orchestrator splits the work into independent tasks, each subagent runs in its own isolated branch using git worktrees, tests validate each piece before integration, and context stays clean across hours of autonomous work. A project that used to be quarters of work for a whole team became one developer plus one orchestration pattern.
@nityeshaga [Claude Code]
https://x.com/nityeshaga/status/2065162130584944886
An Every consultant gave Fable 5 a 6-slide deck task and watched it burn 30 million tokens in 20 minutes doing XML archaeology. Instead of asking for a better deck, he asked Fable to step back and improve the skill it was running inside. Fable diagnosed the real problem (XML flooding the context window), then designed and built its own CLI version of PowerPoint, now open-sourced as Hands-on Deck. Once token costs collapsed, the model spent the savings reviewing its own slides repeatedly without being asked. His takeaway: give Fable the task behind the task.
@dotey [Claude Code]
https://x.com/dotey/status/2064904545298194855
A detailed teardown of how a Claude Code core developer edited an entire launch video using Claude Code plus Fable 5, with no Premiere or Final Cut. The pipeline: local Whisper transcription with word-level timestamps on 25GB of footage, subagents picking the cleanest takes by analyzing the transcript JSON, FFmpeg executing the edit decision list, hand-written LUT code for color grading with an HTML slider page for human tweaks, Remotion React components beat-synced to spoken words, and a Figma MCP round-trip for design polish. The whole edit is text: diffable, greppable, promptable.
@MichLieben [Claude Code]
https://x.com/MichLieben/status/2065073124966895810
The founder of ColdIQ ($7M ARR) rebuilt every business function inside Claude Code and listed the results per team. GTM ships lemlist and Instantly campaigns in under 30 minutes with lists, pain points and copy done. A content workflow trained on past performance pulled 685K LinkedIn impressions in 2 months for one AE. Recruitment cut CV evaluation time 3x. Design ships 5x more pieces via the design system plus Figma MCP. Ads runs $300k/month across Meta, Google and LinkedIn with about 80% of the work handled by Claude Code.
@rui314 [Claude Code]
https://x.com/rui314/status/2065045236125171875
The author of the mold linker (one of the most respected systems programmers alive) threw a one-line request at Fable in Claude Code: add linker script support to mold. The model did it, to the point where mold can now link and boot the Linux kernel. His comment: I knew this day would come, but it is still astonishing to watch it actually happen on a real, hard systems-programming problem.
@bradmillscan [OpenClaw]
https://x.com/bradmillscan/status/2065063193391546555
His OpenClaw install broke after an upgrade: gateway crash loops, dead plugins, no Telegram responses. He gave Fable 5 in the Claude Code desktop app a single /goal with pointers to his Opik logs, gateway logs, 3 months of chat logs and his playbook wiki. It worked for one hour, used 2% of his weekly limit, reconnected plugins, rewrote cron jobs, restored runtime config, migrated him off a broken bridge, asked only 3 questions, and finished with a list of 10 things he should stop doing based on reading 2 months of his own logs.
@robtlee [Claude Code]
https://x.com/robtlee/status/2065044571365532122
A SANS instructor built an autonomous incident response agent on the SIFT Workstation, then typed the prompt: find evil. Claude Code ran a complete C-drive forensic analysis across 200+ tools via MCP, identifying the threat actor, the attack chain, malware deployment method, persistence mechanisms, code injection, C2 infrastructure, the full malicious process tree and a chronological timeline. Two days after he shared the findings, Anthropic published a report on threat actors doing the same thing offensively. His Find Evil hackathon now has 4,178 defenders building IR agents.
@dlouapre [Claude Code]
https://x.com/dlouapre/status/2065059386716811317
A physics team rebuilt their autonomous physics-intern system (which lifted Gemini from 17.7% to 31.4% on the CritPT benchmark) into interactive subagents that plug into Claude Code, because real scientists want a sidekick they can argue with, not an oracle. The author tested it on a problem he abandoned after a 2023 paper: computing resonant modes of 3D meshless shapes with the walk-on-spheres Monte Carlo method. After some back and forth, the agent convinced him his question had no answer as posed, then found the version of the question that did.
@VaibhavSisinty [OpenClaw]
https://x.com/VaibhavSisinty/status/2065081837773627747
The founder of OpenClaw open-sourced the loop he uses to maintain all his repos automatically. It wakes every 5 minutes, scans repos, finds open issues and PRs, writes fixes with tests, has a separate AI review the code, runs CI until green, and merges on its own. When it cannot decide (product choices, security, missing access) it prepares the PR anyway and asks one specific question: land this, delete this, or give me this one thing. He wakes up to mostly shipped work, with the rest needing one decision each.
@MPxbt [Claude Code]
https://x.com/MPxbt/status/2065127934210568300
A non-developer used Claude Code to build a trading bot that trades 5 instruments at once: Nasdaq, Bitcoin, gold, oil and the S&P 500, with automated risk management, execution and daily phone summaries. Mean reversion on indices, breakouts on Bitcoin, trend-following on commodities, ATR position sizing, fixed 1% stop-loss and a correlation filter. 122 trades in 4 weeks, net positive, fully hands-free: she gets a 7am pre-market brief and a 9pm P&L recap, and never touches charts in between.
@waveking1314 [Claude Code]
https://x.com/waveking1314/status/2064939545716834694
A breakdown of a designer who earned $420,000 with a hand-gesture-controlled 3D visual system called touchdesign, built on an $80 webcam, free MediaPipe and TouchDesigner, with Claude Code writing the core Python logic from scratch. MediaPipe tracks 21 keypoints per hand; TouchDesigner turns them into 3D control signals: move hands to move objects, pinch for glitch effects, spread to scale. She now performs live visuals at fashion shows and product launches for $5,000-15,000 per event and sells preset packs at $99. The whole rig was assembled in one weekend.
@0x_fokki [Claude Code]
https://x.com/0x_fokki/status/2065024424554676507
A 22-year-old spent $60 last month and made $8,400 selling the same kind of market intelligence a B2B firm charges $50,000 a year for. Claude Code scraped every competitor in his niche in 45 minutes, mapped pricing across 12 marketplaces in 30 minutes, clustered 50,000 data points into demand gaps in 40 minutes and packaged the gaps as alerts in 20 minutes, with Make handling delivery to subscribers. Total build: 2 hours 15 minutes. The data was free; the structure is what sells.
@VengeonsP [Claude Code]
https://x.com/VengeonsP/status/2064972025169371239
A former soldier in France makes 10-15k euros per month selling leads to artisans, and Claude Code automated the entire technical stack. He builds local SEO sites that generate renovation quote requests; when a homeowner calls, his assistant logs it in a CRM coded with Claude Code, the partner artisan gets the lead pushed to a custom app plus an SMS, and he takes 15% commission on signed quotes. Last month: 79k euros in quotes generated, 30k signed, 4k commission from that channel alone. What stays human: recruiting artisans and answering the phone.
@HlidacStatu [Claude Code]
https://x.com/HlidacStatu/status/2065142070612762806
The Czech transparency watchdog Hlidac Statu pointed Fable 5 at a code review of their 300,000-line codebase. It burned through 4 hours worth of plan tokens in 40 minutes, and found 413 bugs and improvement opportunities. A concrete data point for what a frontier-model whole-repo review actually costs and returns on a real production system.
@ChemistDeFi [Claude Code]
https://x.com/ChemistDeFi/status/2065017388798529865
A team finally let Fable 5 loose on legacy repos nobody dared to open for years. The agent navigated code they were scared of and shipped refactors postponed for quarters. Then the bill arrived: every re-read file, re-sent context window and retry loop was billed, and tokens went much faster than expected. He pairs his experience with the macro data: enterprise AI spend up from $1.2M to $7M average, token prices down 67% while total spend explodes. His conclusion: the real skill now is knowing when the agent is worth keeping the meter running.
@fankaishuoai [Claude Code]
https://x.com/fankaishuoai/status/2064876304412111332
An AI quant trading team of six ships 4-5 releases per day with everyone on Claude Code plus Codex. They started as seven: one leader, one product manager, five engineers, and shipped one release every few days. The founder traced the bottleneck to the leader and the PM, removed both roles, and gave each of six engineers a module with one daily standup. His argument: the PM's core value used to be translating business needs into technical language, and AI is now the best translator, so a PM who cannot execute end to end is left with meetings and message-passing.
@SimonHoiberg [OpenClaw]
https://x.com/SimonHoiberg/status/2065040241971376192
His standard content pipeline now: go for a long walk in the Swiss mountains, talk into Telegram for 30-40 minutes straight, and let his OpenClaw agent structure the rambling into content. By the time he is back, the agent has prepared a full week of carousels, graphics and threads in FeedHive. The ideas stay his; the agent does all the shaping, formatting and scheduling.
@ScottyBeamIO [OpenClaw]
https://x.com/ScottyBeamIO/status/2065103808368091467
An OpenClaw agent on a Mac Mini went from a spoken description to a physical object with zero manual steps: it opened Blender automatically, designed a 3D character, placed it in a scene, animated blinking eyes and stomping feet, built and published a landing page, sliced the model and sent the file to a 3D printer, which started printing. The poster's framing: custom printed figures sell for $30-200 on Etsy, and the bottleneck was never artistry but production speed, so description-to-print autonomy is a business model, not a demo.
@cjzafir [Claude Code]
https://x.com/cjzafir/status/2065104422762684745
A simple routing recipe that cut his weekly Claude Code limit burn by 50%: install the OpenAI Codex plugin inside Claude Code, then use Fable 5 on high for planning, Codex 5.5 xhigh for execution (billed against the Codex plan, no API), and Fable 5 max for review. Two subscriptions, three roles, half the limit consumption on the Anthropic side.
@daniel_mac8 [Claude Code]
https://x.com/daniel_mac8/status/2065066247448821841
The most-shared pattern for using Fable 5 without instantly hitting limits: set the model to Fable 5, reasoning to max, then instruct Claude to run a dynamic workflow where Fable acts only as the orchestrator and delegates the reasoning-heavy phases to Opus (or even Sonnet). Fable is overpowered enough that spending its intelligence on every step is waste; let it manage the cheaper models instead.
@miroburn [Claude Code]
https://x.com/miroburn/status/2065127742891241974
After days of continuous work on his Lab Club app, he reports Fable 5 holds context in one window since Tuesday with no deliberate compaction, picks tasks off the roadmap and feedback on its own, and most strikingly: it batches changes intelligently because it has internalized that his CI/CD pipeline takes 30-40 minutes, sometimes shipping a small fix alone, sometimes bundling a pile of work to optimize throughput from task to production. His verdict: like working with a real CTO who wants to hit KPIs without bothering the CEO.
@Mnilax [Claude Code]
https://x.com/Mnilax/status/2065182511739600950
He rebuilt the same 10-screen dashboard twice as an experiment: about 6 hours of restyling by hand versus about 40 minutes when the fix lived in a machine-readable design system handed to Claude Code, which fixed every screen at once instead of one at a time. He pairs it with Anthropic design lead Jenny Wen's claim that the discover-mock-iterate design process is dying, replaced by design systems that agents execute.
@sewon__min [Claude Code]
https://x.com/sewon__min/status/2065117695025226162
An AI researcher tried manually tracing all of the Olmo model's dependencies, gave up after a few hours, and her students built ModSleuth instead. It turns out Olmo and Nemotron have hundreds of deep, recursive, barely visible dependencies. Her honest field note: she thought it would be a one-week Claude Code project and it was not, because information extraction (where Claude Code shines) was not the hard part; the recursive dependency resolution was. The tool even traces models that declare they used Claude Code, which itself has dependencies.
@monokern [Claude Code]
https://x.com/monokern/status/2065155772368007363
A trader loaded his Obsidian vault with hundreds of chart patterns he had studied, each linked to the outcome, the market context and what he was thinking at the time. Now he types one command and Claude Code finds the relevant sources, runs analysis through NotebookLM and saves everything back structured. The vault stopped being storage and started connecting setups to outcomes, while most traders screenshot charts into Discord and forget them within 48 hours.
@qkl2058 [Claude Code]
https://x.com/qkl2058/status/2065002903899922548
A faceless YouTube Shorts channel about nature facts hit 100K subscribers in 30 days and earned $60,000 over the following 4 months. Claude Code generates the scripts, organizes the nature facts and designs the hooks, compressing what used to be 3 hours per video into minutes. No camera, no face, no team; the channel grew on daily volume and the algorithm's preference for broad, uncontroversial, high-retention content: over 50 million views in the first month.
@sidequestforevr [Claude Code]
https://x.com/sidequestforevr/status/2064935720154743237
A French SEO operator shared the loop he uses to produce ranking articles in 10 minutes: first make Claude Code research and update its own SEO knowledge with sources from Reddit, forums and real experts; then a quality loop where it reviews its own output, scores itself out of 10, and rewrites until it beats 9.5; then encode the whole thing into a CLAUDE.md plus a 5-6 step recipe driven by live keyword data from the Haloscan API. He claims new sites reach thousands of monthly visitors on this method.
@KacperTrzepiec1 [Claude Code]
https://x.com/KacperTrzepiec1/status/2065048769196949894
A Polish founder runs his whole working memory on two tools: Obsidian as the human-readable base of notes, tasks and journal, and Claude Code working on the same files. Every morning the assistant builds his daily summary, books calendar slots for tasks, lists emails awaiting replies, checks the CRM and tells him what to chase: 30 minutes a day gone from the ritual. His 4-person team each runs their own assistant on a shared business context, and the assistants coordinate with each other.
@elie2222 [Claude Code]
https://x.com/elie2222/status/2065214747596005761
He wrote a joke growth hack: install the Kickbacks extension (which pays you for ads shown on the Claude Code spinner), open 50 Claude Code tabs on Haiku, and have each tab ralph-loop counting to a million while ads roll. Then he actually tried it: $1.86 at time of writing, a friend at $13.59, and the tweet itself was scheduled to publish while he slept so he would wake up to either earnings or a ban. Loop engineering at its most absurd and most honest.
@buraksu42 [Claude Code]
https://x.com/buraksu42/status/2065122710867382684
A meticulous 4-server setup for agent-driven development: one machine running claude-code and opencode, one test server, one prod server, and a VPN box, all on a closed Tailscale network. He codes via VS Code Remote SSH plus tmux so sessions survive any local device, never tests on the dev machine (auto-deploy from GitHub instead), manages secrets per-project with a global claude.md convention, runs Dokploy for test/prod, Sentry for bugs, hourly backups to Hetzner plus B2. A complete blueprint for working from anywhere while agents keep running.
@Tz_2022 [Claude Code]
https://x.com/Tz_2022/status/2065125078073802907
His best practice for joint Codex plus Claude Code development: run the Codex app on a MacBook connected to a 24/7 dev host (Linux or Mac Mini) that runs both Codex CLI and Claude Code, with a four-pane layout: project task list on the left, Codex conversation in the middle, and a right pane that switches between file browser, a Claude Code terminal, a browser and diff review, with an SSH terminal at the bottom. The laptop is just a window; the agents never stop.
@IHayato [Claude Code]
https://x.com/IHayato/status/2064949121707032715
He pointed Claude Code's Computer Use at Roblox game development and let it drive the actual editor: it autonomously imported avatars and implemented the needed features, working visually through the GUI. Much slower than a human operating the editor (the demo video is sped up 15x), but fully autonomous, and a preview of what happens to GUI-locked workflows once Computer Use gets fast.
@wangfeng_0128 [Claude Code]
https://x.com/wangfeng_0128/status/2064875051992977569
A veteran Chinese entrepreneur (ex-gaming, now building a robotics company) credits Claude Code with collapsing his hardware iteration cycle: motion trajectory logic discussed with the team one day is running on the workbench the next morning, essentially WYSIWYG for robot control code. He also had two engineers and a PM who had just learned Codex write the front end directly, completing the company's full skill dataset for models and agents in three days.
@Dilmerv [Claude Code]
https://x.com/Dilmerv/status/2065145702653776103
He took a VR game from idea to the official Meta Horizon Store in exactly 4 weeks. The stack: Meta VR CLI, an agentic tools repo full of skills, Claude Code in VS Code, Unity 6 with Unity MCP, Meta XR Core and interaction SDKs, OpenAI Image for concept art and Photoshop for store assets. His takeaway is about what modern XR development looks like when a solo developer pairs an engine with agentic workflows.
@tangero [Claude Code]
https://x.com/tangero/status/2064961524016312620
A Czech publisher kept postponing building a web interface for managing his ad campaigns, until he realized he did not need one: he wrote a conversational skill for Claude Code and now manages campaigns directly from the terminal. His one-line verdict: substantially more efficient than the UI he never built.
@r_monaka_design [Claude Code]
https://x.com/r_monaka_design/status/2065014226931118163
A Japanese designer shipped an interactive site themed on ma, the Japanese concept of negative space, born from the observation that AI is still bad at whitespace. Stories emerge from emptiness; visitors touch and shake the page to play. The production split: images by GPT, implementation by Claude Code, video by Seedance. A small, finished, personal piece of work, the kind of creative shipping that agentic coding quietly enables.
@Andrew_Blumson [Claude Code]
https://x.com/Andrew_Blumson/status/2065142309717697004
He gave Fable one prompt: build a fully playable table tennis game in three.js with WebGPU, no WebGL fallback, PlayStation-4-grade visuals, solid physics, one player versus CPU. One shot, 2.5 hours, 30 agents spawned, playable result. A clean data point on what a single ambitious prompt buys now.
@phoronix [Claude Code]
https://x.com/phoronix/status/2065069031434989751
Phoronix reports YSERVER: a modern X11 server written in Rust with the help of Claude Code. Reimplementing an X11 server is the kind of grinding, spec-heavy systems project that almost nobody volunteers for; an agent doing the heavy lifting changes the economics of even the most thankless infrastructure work.
@CMhOeNnExY [Claude Code]
https://x.com/CMhOeNnExY/status/2065076034849149343
A solo Chinese operator reports his AI business crossed 100,000 RMB in monthly revenue this month, with total inputs being his own time plus Claude Code plus Codex. He works over 10 hours a day and is targeting 1M RMB monthly within two years, building the team out gradually. Unremarkable on the surface, but a concrete picture of the new minimum viable company: one person, two subscriptions.
@measssess [Claude Code]
https://x.com/measssess/status/2065118607043170306
A Turkish researcher highlights academic-research-skills, a Claude Code pipeline that writes academic papers from a single command while attacking AI's nastiest academic failure: hallucinated citations. During literature review it audits whether each found source actually supports the claimed point and rejects the output if not, with chained modules for research, writing, harsh peer-review-style critique and revision.
@MyWestLord [Claude Code]
https://x.com/MyWestLord/status/2064971130746609985
Everyone builds free AI brains in Obsidian; companies pay $20,000-80,000 for the same thing rebuilt in Claude Code, and the difference is one word: permissions. The demoed enterprise brain knows every employee (role, manager, credentials), knows every project's context and strategy, and routes itself across subagents, all on a stack of markdown files, an org chart and access rules. Sales cannot read HR files; juniors cannot see board strategy. Obsidian makes you smarter; Claude Code makes you the vendor.
🗣 User Voice
User Voice
Cost anxiety is the dominant emotion. Microsoft teams reportedly cut Claude Code after blowing annual budgets (@svpino), Uber spent its 2026 AI budget by April, an enterprise engineer with a $3,000/month corporate cap says a $200 Max subscription simply lasts longer (@0xLaughing), and a founder admits accidentally spending $4,000 in 3 days. Users want metering, routing and budget tools, not just bigger models (@ChemistDeFi).
The June 22 Fable deadline is shaping behavior: users describe a 10-day development hackathon to extract maximum value before Fable leaves subscription plans (@imatrix), while others warn the real trap is your brain getting used to Fable-quality output before the two-choice moment arrives: pay double or downgrade (@kinpatsu_cpa).
Safety classifier false positives are the sharpest new pain. One user runs /model every 30 minutes to check whether he has been silently downgraded, scoring 4 fallbacks out of 5 checks (@agupta); another had a mechinterp chat killed by the classifiers (@voooooogel). Anthropic apologized and made fallbacks visible, which users welcomed while bracing for more visible false positives (@oikon48).
Demand for model orchestration is universal: the Fable-as-orchestrator, Opus-as-worker pattern was today's most repeated recipe (@daniel_mac8), and users explicitly ask why Claude Code does not route models per prompt automatically (@trevin).
Memory portability across agents is a rising ask: users juggling Codex, Claude Code, Cursor and OpenClaw complain that every agent is a new employee who forgets the project, and want a shared memory layer rather than client-locked context (@zstmfhy, @servasyy_ai).
Cost anxiety is the dominant emotion. Microsoft teams reportedly cut Claude Code after blowing annual budgets (@svpino), Uber spent its 2026 AI budget by April, an enterprise engineer with a $3,000/month corporate cap says a $200 Max subscription simply lasts longer (@0xLaughing), and a founder admits accidentally spending $4,000 in 3 days. Users want metering, routing and budget tools, not just bigger models (@ChemistDeFi).
The June 22 Fable deadline is shaping behavior: users describe a 10-day development hackathon to extract maximum value before Fable leaves subscription plans (@imatrix), while others warn the real trap is your brain getting used to Fable-quality output before the two-choice moment arrives: pay double or downgrade (@kinpatsu_cpa).
Safety classifier false positives are the sharpest new pain. One user runs /model every 30 minutes to check whether he has been silently downgraded, scoring 4 fallbacks out of 5 checks (@agupta); another had a mechinterp chat killed by the classifiers (@voooooogel). Anthropic apologized and made fallbacks visible, which users welcomed while bracing for more visible false positives (@oikon48).
Demand for model orchestration is universal: the Fable-as-orchestrator, Opus-as-worker pattern was today's most repeated recipe (@daniel_mac8), and users explicitly ask why Claude Code does not route models per prompt automatically (@trevin).
Memory portability across agents is a rising ask: users juggling Codex, Claude Code, Cursor and OpenClaw complain that every agent is a new employee who forgets the project, and want a shared memory layer rather than client-locked context (@zstmfhy, @servasyy_ai).
📡 Eco Products Radar
Eco Products Radar
Products and tools mentioned 3+ times across today's posts:
- Codex (OpenAI) - the default pairing partner: plan in Fable, execute in Codex was today's most repeated cost recipe
- OpenClaw - personal agent platform; founder's automated maintainer loop open-sourced
- Hermes Agent (Nous Research) - self-improving open agent, 190K GitHub stars, often run beside Claude Code
- Cursor - increasingly described as the tool teams churn from
- Obsidian - the favorite human-readable memory layer under Claude Code (second brains, trading vaults, team ops)
- Figma / Figma MCP - design round-tripping in the Fable video-editing and design-system workflows
- Remotion - React-based video graphics rendered by agents instead of After Effects
- FFmpeg - the cutting engine in agent-driven video pipelines
- Mac Mini - the de facto home server for 24/7 agents (one garage farm runs 50 of them)
- Ollama - local model runtime behind no-API-bill setups
- MiMo Code (Xiaomi) - newly open-sourced Claude Code-compatible coding agent with 1M context
- Nessie - context/memory importer praised by Garry Tan for moving ChatGPT/Gemini history into OpenClaw and Hermes
- Linear - issue tracker now dispatching Claude Code and Codex to auto-fix bugs (30% of its own bugs already)
- Kickbacks - the new ad marketplace on the Claude Code spinner; users already loop-farming it
Products and tools mentioned 3+ times across today's posts:
- Codex (OpenAI) - the default pairing partner: plan in Fable, execute in Codex was today's most repeated cost recipe
- OpenClaw - personal agent platform; founder's automated maintainer loop open-sourced
- Hermes Agent (Nous Research) - self-improving open agent, 190K GitHub stars, often run beside Claude Code
- Cursor - increasingly described as the tool teams churn from
- Obsidian - the favorite human-readable memory layer under Claude Code (second brains, trading vaults, team ops)
- Figma / Figma MCP - design round-tripping in the Fable video-editing and design-system workflows
- Remotion - React-based video graphics rendered by agents instead of After Effects
- FFmpeg - the cutting engine in agent-driven video pipelines
- Mac Mini - the de facto home server for 24/7 agents (one garage farm runs 50 of them)
- Ollama - local model runtime behind no-API-bill setups
- MiMo Code (Xiaomi) - newly open-sourced Claude Code-compatible coding agent with 1M context
- Nessie - context/memory importer praised by Garry Tan for moving ChatGPT/Gemini history into OpenClaw and Hermes
- Linear - issue tracker now dispatching Claude Code and Codex to auto-fix bugs (30% of its own bugs already)
- Kickbacks - the new ad marketplace on the Claude Code spinner; users already loop-farming it
Comments