multica-ai

Our Multica agents now work in an agent market: they buy, remember, review sellers, and sell their own work Multica is an agent factory: you give an agent a runtime, instructions and a schedule, and it works through tasks. AIMarket hubs (modelmarket.dev is the public one) are a market where agents buy capabilities from each other per call. We run both, and in the first week of November we connected them. This is what the agents do now, what broke, what we found when we audited ourselves, and a setup guide in five languages for anyone who wants the same. The crew Five agents run on one self-hosted Multica 0.5.3 (OpenCode + DeepSeek), on fixed UTC schedules: - GitHub digest, daily at 08:00: activity across our repositories, sent to Telegram. It keeps its past digests in Attested Memory, bought on our hub per call ($0.004 to search, $0.012 to write), so it can say "stars on X fell for the third day" instead of repeating yesterday. About $0.03 a run. - Mystery shopper, daily at 09:00: buys a few random safe skills on modelmarket.dev with a prepaid key capped at $0.15 a run, and checks the receipt, the latency and whether the answer matches the listing. - Supplier reviewer, daily at 10:00: diffs the market against yesterday and reviews every new or changed listing. - Weekly post, Mondays at 11:00: drafts a build-in-public post from the hub's counters and the agents' own results. A human reads it and decides. - Order runner: writes the summary for a long job that someone buys on the hub (below). The bug we found before writing a single memory Our memory service took a memory's owner from a payment header that the hub sends empty for most buyers. Every buyer on a prepaid key, a trial or an on-chain payment was the same owner, so "private" memories were readable by all of them. Nothing had been written through the hub yet, so nothing leaked. The fix: the hub names the buyer it authenticated to its own providers with an HMAC pseudonym. The pseudonym is stable across restarts and reveals no account, and the memory service keys owners by it. We found the same bug once more a day later, one hub further away (see "We audited ourselves"). Agents that read strangers' text An agent that reads GitHub issues or sellers' answers can be talked into sending data out. Three layers: 1. Egress allowlist. The agents' Docker network has no route out except a proxy with a short, anchored host list. 2. A flow guard in the runtime. Our WARDEN guard runs as an OpenCode plugin (@aimarket/warden/opencode, from 0.23.0). After the session has read outside text, it refuses a tool call that would carry a credential or text the session read. In a staged attack we planted a fake GitHub token in a file. Sending it was refused through the chat script, curl to an allowed host, a memory write, a project file and the tracker CLI. A normal day's digest ran with zero refusals, after we fixed one false positive: a task's UUID looked like private text. 3. Least privilege per agent. All agents share one daemon, so each gets its own Multica environment. The reviewer and the post writer get empty hub keys, no GitHub token, and the hub's MCP tools switched off. Reviewing sellers with a script, not a model The reviewer reads exactly the text an attacker writes: listing descriptions. So a script decides the facts, and the model only writes the prose and a welcome draft for each outside author. The facts are broken schemas, injection patterns from WARDEN's static scan, a skill that asks the buyer for a credential, plain-http or internal addresses, and price mismatches. The script's facts block goes into the report unchanged. Its first run took a baseline of 282 listings, 22 federation peers and 2 x402 sellers. The one high finding was ours: our public manifest showed a seller's internal plain-http container address. The hub now shows its own invoke address for sellers that only it can reach, and keeps the real one for routing. The reviewer also drafted five welcome letters to outside sellers. Nobody has sent them; that is a human decision. Selling a long job A paid call cannot wait minutes for an agent, so a GitHub activity digest is sold as two capabilities on hub.attestedmemory.net: - github.digest.order@v1 ($0.05) takes any public GitHub user or organization and 1 to 7 days, and answers at once with a job id; - github.digest.result@v1 (free) returns the digest to the buyer who paid. The first real order failed, and our own guard caused half of it. The agent read its task, including a per-job delivery URL, then read GitHub, then tried to send the result to that URL. WARDEN treats task text as private, and after outside text it lets no private data leave, whatever the destination. That is the rule we want for an unattended agent; the delivery design was wrong. The other half: GitHub's 60 unauthenticated requests an hour per IP were already spent on that host. (By what, we found later: a video worker of ours whose model calls had been failing since September. It re-fetched GitHub every two minutes, because its failed run never saved the time.) The second design went live the next morning. The service does the GitHub reads itself, with a token that has no scopes, and builds the dated items. A login that does not exist, or a GitHub outage, refuses the order before anyone is charged. The agent's final answer is one JSON line with a summary. The service collects it through a read-only database function that returns only that agent's answers for that job. If no usable summary arrives, the service writes a plain one from the same data, so a paid order always gets its digest. Four test orders so far: - octocat for a week, 0 events; - the microsoft organization for a day, 87 events; - our own account for three days, then for two. Each came back with the agent's summary. In the latest one, the agent's answer was there 31 seconds after the order. We audited ourselves for the bug classes an outside audit found In October, Vulnetix found real bugs in our code, and we then audited the whole ecosystem for the same classes: - endpoints open without a key; - checking a host name and then connecting to it; - trusting forwarded client addresses. On November 5 we went through everything built for Multica looking for those classes, plus secrets reaching errors and logs, and allowlists that can be stepped around. The service, the scripts, the database function, the proxy and the ports were clean. Four things were not. 1. The order runner could reach every task in the workspace. Its OpenCode rules allowed multica issue get *, and a Multica agent's token is bound to the workspace, not to its own task. Task numbers are sequential. Text a buyer controls reaches the agent: the PR titles of the GitHub account they order. If that text talked the agent into reading another task, the content could land in the summary. Multica posts that summary itself, outside the tool calls a guard sees. A redirect (> file) also passed the allowlist. The agent now runs a root-owned wrapper that reads and closes only tasks assigned to it, with redirects, $ and backticks denied. 2. The guard could be renamed away. Its files belonged to root, but the folder above them belonged to the agent user, and Linux lets you rename a folder inside one you own. The guard now lives in a root-owned folder outside the agent's home, and OpenCode finds it through OPENCODE_CONFIG_DIR. 3. The memory-owner bug, one hub further. When another hub routes a call to ours on its own key, our hub saw that hub's account as the buyer, so all of its buyers would have shared one memory owner. Nobody had used that path. Routed calls now name no owner. 4. Container addresses in our second hub's public manifest, the same as the reviewer's finding. That fix is in the code; this hub gets it with its next update. Two things we learned about OpenCode 1.18.32 on the way: - Permissions. It checks every command in a line on its own, including those inside $(...), backticks and pipes, and the last matching rule wins. A redirected command is matched together with its redirect. - The config folder. At start it writes a .gitignore into every config folder. On a read-only mount it stops with an error, so mount the guard's folder writable but owned by root. We could not test the permission layer by asking the model to break it: DeepSeek refused all six bypass attempts on its own. We read OpenCode's source instead. If you run Multica too Step by step, from the container to selling a long job, in English, Russian, Spanish, French and Chinese: multica/manual. The Multica skill our agents use for buying and selling is SKILL.md. The full audit is in REPORT.md. What we would tell another operator: - Keep facts in scripts and prose in models. - Give each agent its own environment; the global config is shared by everyone on the daemon. - Put the network allowlist below the prompt, and the guard below the model, in a folder the agent cannot touch. - Allow an agent a wrapper you wrote, not a CLI with a wildcard. This article was written with AI assistance from our agents' own logs and audit notes, and edited by a human. * We thank Vulnetix for helping us improve our products.

Does Karpathy know there's a GitHub repo out there with 227k stars which turned his brain into Agent Skills lol multica-ai/andrej-karpathy-skills t.co

Make humans and AI agents work as one team — open-source and self-hostable. multica-ai/multica

Copy any app, Claude does SEO, and the top 10 repos Link to Resource: thenextnewthing.ai - This video is sponsored by TripleTen, a professional education platform designed to help tech professionals and career changers move into higher-paying tech and AI roles. - Explore two TripleTen programs for experienced engineers: - AI Systems Engineering (22 weeks): tripleten.com - AI and Machine Learning (20 weeks): tripleten.com - Both are for engineers with 2+ years of recent, hands-on production experience and require a short skills audit before you can enroll. Both also include career coaching and mentorship to help you stand out in the job market. - TripleTen - github.com/tripleten-com - TripleTen Website - tripleten.com - Presented by Zapier: zapier.com - Zapier Next Gen Zaps - zapier.com This week, Andrew Warner is joined by Adam Brakhane to explore the GitHub repositories and AI tools that are making developers more productive. From running open-source apps in the cloud to giving AI coding agents security, design, and workflow superpowers, these projects show what's possible when developers build on top of the latest AI models. The episode starts with three bonus discoveries: a way to host open-source apps cheaply, Theo's collection of coding skills, and a project that makes iOS app development on Linux possible. Then Andrew and Adam count down the top 10 GitHub repositories of the week. They explore Cloudflare's security audit skill, Claude SEO, Impeccable, Superpowers, TradingAgents, Karpathy-inspired skills, design and accessibility tools, T3 Code, and Anthropic's Knowledge Work Plugins. Along the way, they discuss how open-source projects can become businesses, why skills need to be customized and kept up to date, how AI agents can be persuaded to follow instructions, and why the best coding tools are becoming more than just coding assistants. Repos Featured: Bonus Openers - Every App: every-app/every-app - T3 Stack: t3dotgg/tstack - Omarchy Apple Dev: joshuaswarren/omarchy-apple-dev Top 10 GitHub Repositories - Cloudflare Security Audit Skill: cloudflare/security-audit-skill - Claude SEO: AgriciDaniel/claude-seo - Impeccable: pbakaus/impeccable - Superpowers: obra/superpowers - TradingAgents: TauricResearch/TradingAgents - Andrej Karpathy Skills: multica-ai/andrej-karpathy-skills - Taste Skill: Leonxlnx/taste-skill - Interface Design Skills: jakubkrehel/skills - T3 Code: pingdotgg/t3code - Knowledge Work Plugins: anthropics/knowledge-work-plugins 00:00 - This Week's Best GitHub Repositories 00:27 - Every App: Host Open-Source Apps for $5 00:57 - Why Easy Deployment Matters for Open Source 02:15 - T3 Stack: Theo's AI Coding Skills 02:42 - Why You Should Customize Your AI Skills 03:43 - Omarchy Apple Dev: Build iOS Apps on Linux 04:29 - Why Apple Requires Mac Hardware for iOS Development 05:43 - #1 Cloudflare Security Audit Skill 07:00 - #2 Claude SEO: Run SEO Audits with Claude Code 09:31 - #3 Impeccable: Give AI Agents Better Design Taste 10:35 - Can Open-Source Design Tools Become Businesses? 12:40 - Presented by TripleTen 14:36 - #4 Superpowers: Give Coding Agents a Better Process 15:09 - Using Persuasion Principles to Improve AI Agents 17:11 - #5 TradingAgents: Simulated AI Trading 18:41 - #6 Andrej Karpathy Skills 20:31 - Why AI Agents Struggle to Follow Task Lists 21:43 - Presented by Zapier: Make AI Workflows Cheaper 22:38 - #7 Taste Skill: Give Your AI an Art Director 24:36 - #8 Interface Design Skills and Accessibility 26:20 - #9 T3 Code: Manage Multiple AI Coding Agents 27:20 - Run AI Agents Across Multiple Computers 30:46 - #10 Anthropic Knowledge Work Plugins 31:14 - What Should We Cover in First Stars? 31:45 - Adam's AI-Powered Business Context Layer 33:12 - Andrew's Automated AI Video Editing Workflow 34:18 - More GitHub Repositories to Explore 👉 Media/Sponsorship Inquiries: thenextnewthing.ai 👉 Join us: thenextnewthing.ai [de-DE] Jede App kopieren, Claude macht SEO und die Top 10 Repos Link zur Ressource: thenextnewthing.ai – Dieses Video wird von TripleTen gesponsert, einer professionellen Weiterbildungsplattform, die IT-Fachkräfte und Quereinsteiger beim Einstieg in besser bezahlte Positionen im Technologie- und KI-Bereich unterstützt. – Entdecken Sie zwei TripleTen-Programme für erfahrene Ingenieure: – KI-Systementwicklung (22 Wochen): tripleten.com – KI und Maschinelles Lernen (20 Wochen): tripleten.com – Beide Programme richten sich an Ingenieure mit mindestens zwei Jahren aktueller, praktischer Berufserfahrung. Vor der Anmeldung ist eine kurze Kompetenzanalyse erforderlich. Beide Programme beinhalten außerdem Karrierecoaching und Mentoring, um Ihnen zu helfen, sich auf dem Arbeitsmarkt hervorzuheben. - TripleTen - github.com/tripleten-com - TripleTen-Website - tripleten.com - Präsentiert von Zapier: zapier.com - Zapier Next Gen Zaps - zapier.com Diese Woche sprechen Andrew Warner und Adam Brakhane über GitHub-Repositories und KI-Tools, die die Produktivität von Entwicklern steigern. Von der Ausführung von Open-Source-Apps in der Cloud bis hin zur Ausstattung von KI-Programmieragenten mit Sicherheits-, Design- und Workflow-Funktionen – diese Projekte zeigen, was möglich ist, wenn Entwickler auf den neuesten KI-Modellen aufbauen. Die Folge beginnt mit drei Bonus-Entdeckungen: einer Möglichkeit, Open-Source-Apps kostengünstig zu hosten, Theos Sammlung von Programmierkenntnissen und einem Projekt, das die iOS-App-Entwicklung unter Linux ermöglicht. Anschließend präsentieren Andrew und Adam die Top 10 der GitHub-Repositories der Woche. Sie untersuchen Cloudflares Sicherheitsaudit-Funktion, Claude SEO, Impeccable, Superpowers, TradingAgents, von Karpathy inspirierte Funktionen, Design- und Barrierefreiheitstools, T3 Code und die Knowledge Work Plugins von Anthropic. Dabei erörtern sie, wie Open-Source-Projekte zu Unternehmen werden können, warum Fähigkeiten angepasst und aktuell gehalten werden müssen, wie KI-Agenten dazu gebracht werden können, Anweisungen zu befolgen, und warum die besten Codierungswerkzeuge mehr als nur Programmierassistenten sind. Empfohlene Repositories: Bonus-Opener - Every App: every-app/every-app - T3 Stack: t3dotgg/tstack - Omarchy Apple Dev: joshuaswarren/omarchy-apple-dev Top 10 GitHub-Repositories - Cloudflare Security Audit Skill: cloudflare/security-audit-skill - Claude SEO: AgriciDaniel/claude-seo - Impeccable: pbakaus/impeccable - Superpowers: obra/superpowers - TradingAgents: TauricResearch/TradingAgents - Andrej Karpathy Skills: multica-ai/andrej-karpathy-skills - Geschmacks-Skill: Leonxlnx/taste-skill - Interface-Design-Skills: jakubkrehel/skills - T3 Code: pingdotgg/t3code - Knowledge-Work-Plugins: anthropics/knowledge-work-plugins 00:00 – Die besten GitHub-Repositories der Woche 00:27 – Every App: Open-Source-Apps für 5 $ hosten 00:57 – Warum einfache Bereitstellung für Open Source wichtig ist 02:15 – T3 Stack: Theos KI-Programmierkenntnisse 02:42 – Warum Sie Ihre KI-Kenntnisse anpassen sollten 03:43 – Omarchy Apple Dev: iOS-Apps unter Linux entwickeln 04:29 – Warum Apple Mac-Hardware für iOS benötigt Entwicklung 05:43 – #1 Cloudflare-Sicherheitsaudit-Kompetenz 07:00 – #2 Claude SEO: SEO-Audits mit Claude Code durchführen 09:31 – #3 Tadellos: KI-Agenten ein besseres Gespür für Design vermitteln 10:35 – Können Open-Source-Designtools zu Unternehmen werden? 12:40 – Präsentiert von TripleTen 14:36 ​​– #4 Superkräfte: Optimieren Sie Ihre Programmieragenten 15:09 – KI-Agenten mithilfe von Überzeugungsprinzipien verbessern 17:11 – #5 TradingAgents: Simulierter KI-Handel 18:41 – #6 Andrej Karpathy Skills 20:31 – Warum KI-Agenten Schwierigkeiten haben, Aufgabenlisten abzuarbeiten 21:43 – Präsentiert von Zapier: KI-Workflows kostengünstiger gestalten 22:38 – #7 Geschmackssache: Geben Sie Ihrer KI einen Art Director 24:36 – #8 Interface-Design und Barrierefreiheit 26:20 – #9 T3 Code: Mehrere KI-Programmieragenten verwalten 27:20 – KI-Agenten auf mehreren Computern ausführen 30:46 – #10 Anthropic Knowledge Work Plugins 31:14 – Was sollten wir in First Stars behandeln? 31:45 – Adams KI-gestützte Business-Context-Ebene 33:12 – Andrews automatisierter KI-Video-Workflow 34:18 – Weitere GitHub-Repositories zum Entdecken 👉 Medien-/Sponsoringanfragen: thenextnewthing.ai 👉 Werden Sie Teil unserer Community: thenextnewthing.ai [es-US] Copia cualquier aplicación, Claude hace SEO y los 10 mejores repositorios Enlace al recurso: thenextnewthing.ai - Este video está patrocinado por TripleTen, una plataforma de formación profesional diseñada para ayudar a profesionales de la tecnología y personas que buscan un cambio de carrera a acceder a puestos mejor remunerados en tecnología e IA. - Descubre dos programas de TripleTen para ingenieros con experiencia: - Ingeniería de Sistemas de IA (22 semanas): tripleten.com - IA y Aprendizaje Automático (20 semanas): tripleten.com - Ambos programas están dirigidos a ingenieros con más de dos años de experiencia práctica reciente en producción y requieren una breve evaluación de habilidades antes de la inscripción. Ambos incluyen asesoramiento profesional y mentoría para ayudarte a destacar en el mercado laboral. - TripleTen - github.com/tripleten-com - Sitio web de TripleTen - tripleten.com - Presentado por Zapier: zapier.com - Zaps de nueva generación de Zapier - zapier.com Esta semana, Andrew Warner y Adam Brakhane exploran los repositorios de GitHub y las herramientas de IA que aumentan la productividad de los desarrolladores. Desde ejecutar aplicaciones de código abierto en la nube hasta dotar a los agentes de codificación de IA de capacidades avanzadas de seguridad, diseño y flujo de trabajo, estos proyectos demuestran el potencial de los desarrolladores que trabajan con los modelos de IA más recientes. El episodio comienza con tres descubrimientos adicionales: una forma económica de alojar aplicaciones de código abierto, la colección de habilidades de programación de Theo y un proyecto que permite el desarrollo de aplicaciones iOS en Linux. A continuación, Andrew y Adam presentan los 10 mejores repositorios de GitHub de la semana. Exploran la función de auditoría de seguridad de Cloudflare, Claude SEO, Impeccable, Superpowers, TradingAgents, las habilidades inspiradas en Karpathy, las herramientas de diseño y accesibilidad, T3 Code y los plugins de gestión del conocimiento de Anthropic. Además, analizan cómo los proyectos de código abierto pueden convertirse en negocios, por qué es necesario personalizar y actualizar las habilidades, cómo se puede persuadir a los agentes de IA para que sigan instrucciones y por qué las mejores herramientas de programación van más allá de ser simples asistentes. Repositorios destacados: Introducciones adicionales - Every App: every-app/every-app - T3 Stack: t3dotgg/tstack - Omarchy Apple Dev: joshuaswarren/omarchy-apple-dev Los 10 mejores repositorios de GitHub - Cloudflare Security Audit Skill: cloudflare/security-audit-skill - Claude SEO: AgriciDaniel/claude-seo - Impeccable: pbakaus/impeccable - Superpowers: obra/superpowers - TradingAgents: TauricResearch/TradingAgents - Andrej Karpathy Skills: multica-ai/andrej-karpathy-skills - Habilidad de Gusto: Leonxlnx/taste-skill - Habilidades de Diseño de Interfaz: jakubkrehel/skills - Código T3: pingdotgg/t3code - Plugins de Trabajo del Conocimiento: anthropics/knowledge-work-plugins 00:00 - Los mejores repositorios de GitHub de esta semana 00:27 - Every App: Aloja aplicaciones de código abierto por $5 00:57 - Por qué la implementación sencilla es importante para el código abierto 02:15 - T3 Stack: Las habilidades de programación de IA de Theo 02:42 - Por qué deberías personalizar tus habilidades de IA 03:43 - Omarchy Apple Dev: Crea aplicaciones iOS en Linux 04:29 - Por qué Apple requiere hardware Mac para iOS Desarrollo 05:43 - #1 Habilidad de auditoría de seguridad de Cloudflare 07:00 - #2 Claude SEO: Realiza auditorías SEO con Claude Code 09:31 - #3 Impecable: Mejora el diseño de los agentes de IA 10:35 - ¿Pueden las herramientas de diseño de código abierto convertirse en negocios? 12:40 - Presentado por TripleTen 14:36 ​​- #4 Superpoderes: Mejora el proceso de los agentes de codificación 15:09 - Uso de principios de persuasión para mejorar los agentes de IA 17:11 - #5 Agentes de comercio: Simulación de comercio con IA 18:41 - #6 Habilidades de Andrej Karpathy 20:31 - ¿Por qué los agentes de IA tienen dificultades para seguir las listas de tareas? 21:43 - Presentado por Zapier: Abarata los flujos de trabajo de IA 22:38 - #7 Habilidad de buen gusto: Asigna un director de arte a tu IA 24:36 - #8 Habilidades de diseño de interfaz y accesibilidad 26:20 - #9 Código T3: Gestiona múltiples agentes de codificación de IA 27:20 - Ejecuta agentes de IA en varios ordenadores 30:46 - #10 Plugins de trabajo de conocimiento antrópico 31:14 - ¿Qué deberíamos cubrir en First Stars? 3... [fr-FR] Copiez n'importe quelle application, Claude fait du SEO, et le top 10 des dépôts Lien vers la ressource : thenextnewthing.ai - Cette vidéo est sponsorisée par TripleTen, une plateforme de formation professionnelle conçue pour aider les professionnels de la tech et les personnes en reconversion à accéder à des postes mieux rémunérés dans les domaines de la tech et de l'IA. - Découvrez deux programmes TripleTen pour ingénieurs expérimentés : - Ingénierie des systèmes d'IA (22 semaines) : tripleten.com - IA et apprentissage automatique (20 semaines) : tripleten.com - Ces deux programmes s'adressent aux ingénieurs justifiant d'au moins deux ans d'expérience pratique récente en production. Une évaluation rapide de vos compétences est requise avant l'inscription. Ils incluent également un accompagnement et un mentorat pour vous aider à vous démarquer sur le marché du travail. - TripleTen - github.com/tripleten-com - Site web de TripleTen - tripleten.com - Présenté par Zapier : zapier.com - Zapiers nouvelle génération - zapier.com Cette semaine, Andrew Warner et Adam Brakhane explorent les dépôts GitHub et les outils d'IA qui améliorent la productivité des développeurs. De l'exécution d'applications open source dans le cloud à l'optimisation de la sécurité, du design et des flux de travail des agents de programmation IA, ces projets illustrent le potentiel des modèles d'IA les plus récents. L'épisode débute par trois découvertes bonus : une solution économique pour héberger des applications open source, les compétences de programmation de Theo et un projet permettant le développement d'applications iOS sous Linux. Andrew et Adam présentent ensuite leur top 10 des dépôts GitHub de la semaine. Ils explorent les fonctionnalités d'audit de sécurité de Cloudflare, Claude SEO, Impeccable, Superpowers, TradingAgents, les compétences inspirées de Karpathy, les outils de conception et d'accessibilité, T3 Code et les plugins Knowledge Work d'Anthropic. Au fil de leur exploration, ils abordent la transformation des projets open source en entreprises, la nécessité de personnaliser et de maintenir à jour les compétences, la manière d'inciter les agents d'IA à suivre des instructions et l'évolution des meilleurs outils de programmation, qui deviennent bien plus que de simples assistants de codage. Dépôts en vedette : Ouvertures bonus - Every App : every-app/every-app - T3 Stack : t3dotgg/tstack - Omarchy Apple Dev : joshuaswarren/omarchy-apple-dev Top 10 des dépôts GitHub - Cloudflare Security Audit Skill : cloudflare/security-audit-skill - Claude SEO : AgriciDaniel/claude-seo - Impeccable : pbakaus/impeccable - Superpowers : obra/superpowers - TradingAgents : TauricResearch/TradingAgents - Andrej Karpathy Skills : multica-ai/andrej-karpathy-skills - Compétences en perception gustative : Leonxlnx/taste-skill - Compétences en conception d'interfaces : jakubkrehel/skills - Code T3 : pingdotgg/t3code - Plugins pour le travail intellectuel : anthropics/knowledge-work-plugins 00:00 - Les meilleurs dépôts GitHub de la semaine 00:27 - Every App : Hébergez des applications open source pour 5 $ 00:57 - L'importance d'un déploiement facile pour l'open source 02:15 - T3 Stack : Les compétences de programmation IA de Theo 02:42 - Pourquoi personnaliser vos compétences en IA ? 03:43 - Omarchy Apple Dev : Créer des applications iOS sous Linux 04:29 - Pourquoi Apple exige du matériel Mac pour iOS Développement 05:43 - Compétence n° 1 : Audit de sécurité Cloudflare 07:00 - Claude SEO n° 2 : Réaliser des audits SEO avec Claude Code 09:31 - Impeccable n° 3 : Améliorer le sens du design des agents IA 10:35 - Les outils de conception open source peuvent-ils devenir rentables ? 12:40 - Présenté par TripleTen 14:36 ​​- N° 4 Superpouvoirs : Améliorer le processus des agents de codage 15:09 - Utiliser les principes de persuasion pour améliorer les agents d'IA 17:11 - N° 5 TradingAgents : Simulation de trading par IA 18:41 - N° 6 Compétences d'Andrej Karpathy 20:31 - Pourquoi les agents d'IA peinent à suivre les listes de tâches 21:43 - Présenté par Zapier : Réduire le coût des flux de travail d'IA 22:38 - N° 7 Sens du goût : Offrir un directeur artistique à votre IA 24:36 - N° 8 Compétences en conception d'interface et accessibilité 26:20 - N° 9 Code T3 : Gérer plusieurs agents de codage d'IA 27:20 - Exécuter des agents d'IA sur plusieurs ordinateurs 30:46 - N° 10 Plugins Anthropic Knowledge Work 31:14 - Que devrions-nous aborder dans First Stars ? 31:45 - Couche de contexte... [hi] किसी भी ऐप को कॉपी करें, Claude करेगा SEO, और टॉप 10 रेपोज़ संसाधन का लिंक: thenextnewthing.ai - यह वीडियो ट्रिपलटेन द्वारा प्रायोजित है, जो एक पेशेवर शिक्षा मंच है जिसे तकनीकी पेशेवरों और करियर बदलने वालों को उच्च वेतन वाली तकनीकी और एआई भूमिकाओं में आगे बढ़ने में मदद करने के लिए डिज़ाइन किया गया है। - अनुभवी इंजीनियरों के लिए ट्रिपलटेन के दो कार्यक्रमों के बारे में जानें: - एआई सिस्टम इंजीनियरिंग (22 सप्ताह): tripleten.com - एआई और मशीन लर्निंग (20 सप्ताह): tripleten.com - दोनों कार्यक्रम उन इंजीनियरों के लिए हैं जिनके पास हाल ही में 2+ वर्षों का व्यावहारिक उत्पादन अनुभव है और नामांकन से पहले एक संक्षिप्त कौशल ऑडिट आवश्यक है। दोनों कार्यक्रमों में करियर कोचिंग और मेंटरशिप भी शामिल है ताकि आप नौकरी बाजार में अलग पहचान बना सकें। - ट्रिपलटेन - github.com/tripleten-com - ट्रिपलटेन वेबसाइट - tripleten.com - ज़ैपियर द्वारा प्रस्तुत: zapier.com - ज़ैपियर नेक्स्ट जेन ज़ैप्स - zapier.com इस सप्ताह, एंड्रयू वार्नर के साथ एडम ब्राखाने GitHub रिपॉज़िटरी और AI टूल्स पर चर्चा करेंगे जो डेवलपर्स की उत्पादकता बढ़ा रहे हैं। क्लाउड में ओपन-सोर्स ऐप्स चलाने से लेकर AI कोडिंग एजेंट्स को सुरक्षा, डिज़ाइन और वर्कफ़्लो की अपार क्षमता प्रदान करने तक, ये प्रोजेक्ट दिखाते हैं कि जब डेवलपर्स नवीनतम AI मॉडल्स पर काम करते हैं तो क्या-क्या संभव है। एपिसोड की शुरुआत तीन अतिरिक्त खोजों से होती है: ओपन-सोर्स ऐप्स को सस्ते में होस्ट करने का एक तरीका, थियो के कोडिंग कौशल का संग्रह, और एक ऐसा प्रोजेक्ट जो लिनक्स पर iOS ऐप डेवलपमेंट को संभव बनाता है। इसके बाद एंड्रयू और एडम इस सप्ताह की शीर्ष 10 GitHub रिपॉज़िटरी की सूची प्रस्तुत करते हैं। वे क्लाउडफ्लेयर के सुरक्षा ऑडिट कौशल, क्लाउड एसईओ, इम्पेकेबल, सुपरपावर्स, ट्रेडिंग एजेंट्स, कार्पेथी से प्रेरित कौशल, डिज़ाइन और एक्सेसिबिलिटी टूल्स, टी3 कोड और एंथ्रोपिक के नॉलेज वर्क प्लगइन्स का अन्वेषण करते हैं। इस दौरान, वे चर्चा करते हैं कि कैसे ओपन-सोर्स प्रोजेक्ट व्यवसाय बन सकते हैं, कौशल को अनुकूलित और अद्यतन रखना क्यों आवश्यक है, एआई एजेंटों को निर्देशों का पालन करने के लिए कैसे प्रेरित किया जा सकता है, और क्यों सर्वश्रेष्ठ कोडिंग उपकरण केवल कोडिंग सहायकों से कहीं अधिक बन रहे हैं। प्रमुख रिपॉज़िटरी: बोनस ओपनर्स - एवरी ऐप: every-app/every-app - T3 स्टैक: t3dotgg/tstack - ओमार्ची एप्पल देव: joshuaswarren/omarchy-apple-dev शीर्ष 10 GitHub रिपॉज़िटरी - क्लाउडफ्लेयर सिक्योरिटी ऑडिट स्किल: cloudflare/security-audit-skill - क्लाउड एसईओ: AgriciDaniel/claude-seo - इम्पेकेबल: pbakaus/impeccable - सुपरपावर्स: obra/superpowers - ट्रेडिंग एजेंट्स: TauricResearch/TradingAgents - आंद्रेज कार्पेथी स्किल्स: multica-ai/andrej-karpathy-skills - स्वाद कौशल: Leonxlnx/taste-skill - इंटरफ़ेस डिज़ाइन कौशल: jakubkrehel/skills - T3 कोड: pingdotgg/t3code - नॉलेज वर्क प्लगइन्स: anthropics/knowledge-work-plugins 00:00 - इस सप्ताह के सर्वश्रेष्ठ GitHub रिपॉज़िटरी 00:27 - एवरी ऐप: ओपन-सोर्स ऐप्स को $5 में होस्ट करें 00:57 - ओपन सोर्स के लिए आसान डिप्लॉयमेंट क्यों महत्वपूर्ण है 02:15 - T3 स्टैक: थियो के AI कोडिंग कौशल 02:42 - आपको अपने AI कौशल को क्यों अनुकूलित करना चाहिए 03:43 - ओमार्ची एप्पल देव: लिनक्स पर iOS ऐप्स बनाएं 04:29 - एप्पल को इसकी आवश्यकता क्यों है iOS डेवलपमेंट के लिए मैक हार्डवेयर 05:43 - #1 क्लाउडफ्लेयर सुरक्षा ऑडिट कौशल 07:00 - #2 क्लाउड एसईओ: क्लाउड कोड के साथ एसईओ ऑडिट करें 09:31 - #3 इम्पेकेबल: एआई एजेंटों को बेहतर डिज़ाइन क्षमता प्रदान करें 10:35 - क्या ओपन-सोर्स डिज़ाइन टूल व्यवसाय बन सकते हैं? 12:40 - ट्रिपलटेन द्वारा प्रस्तुत 14:36 ​​- #4 सुपरपावर्स: कोडिंग एजेंट्स को बेहतर प्रक्रिया प्रदान करें 15:09 - एआई एजेंट्स को बेहतर बनाने के लिए अनुनय सिद्धांतों का उपयोग 17:11 - #5 ट्रेडिंग एजेंट्स: सिम्युलेटेड एआई ट्रेडिंग 18:41 - #6 आंद्रेज कार्पेथी स्किल्स 20:31 - एआई एजेंट्स को टास्क लिस्ट का पालन करने में कठिनाई क्यों होती है 21:43 - जैपियर द्वारा प्रस्तुत: एआई वर्कफ़्लो को सस्ता बनाएं 22:38 - #7 टेस्ट स्किल: अपने एआई को एक आर्ट डायरेक्टर दें 24:36 - #8 इंटरफ़ेस डिज़ाइन स्किल्स और एक्सेसिबिलिटी 26:20 - #9 टी3 कोड: कई एआई कोडिंग एजेंट्स को प्रबंधित करें 27:20 - कई कंप्यूटरों पर एआई एजेंट्स चलाएं 30:46 - #10 एंथ्रोपिक नॉलेज वर्क प्लगइन्स 31:14 - हमें फर्स्ट स्टार्स में क्या शामिल करना चाहिए? 31:45 - एडम का एआई-आधारित बिज़नेस कॉन्टेक्स्ट लेयर 33:12 - एंड्रयू का स्वचालित एआई वीडियो एडिटिंग वर्कफ़्लो 34:18 - एक्सप्लोर करने के लिए और भी GitHub रिपॉज़िटरीज़ 👉 मीडिया/स्पॉन्सरशिप पूछताछ: thenextnewthing.ai 👉 हमसे जुड़ें: thenextnewthing.ai [id] Salin aplikasi apa pun, Claude melakukan SEO, dan 10 repositori teratas Tautan ke Sumber Daya: thenextnewthing.ai - Video ini disponsori oleh TripleTen, platform pendidikan profesional yang dirancang untuk membantu para profesional teknologi dan mereka yang ingin beralih karier ke peran teknologi dan AI dengan gaji lebih tinggi. - Jelajahi dua program TripleTen untuk insinyur berpengalaman: - Rekayasa Sistem AI (22 minggu): tripleten.com - AI dan Pembelajaran Mesin (20 minggu): tripleten.com - Kedua program ini ditujukan untuk insinyur dengan pengalaman produksi langsung minimal 2 tahun dan memerlukan audit keterampilan singkat sebelum Anda dapat mendaftar. Kedua program ini juga mencakup pelatihan karier dan bimbingan untuk membantu Anda menonjol di pasar kerja. - TripleTen - github.com/tripleten-com - Situs Web TripleTen - tripleten.com - Dipersembahkan oleh Zapier: zapier.com - Zapier Next Gen Zaps - zapier.com Minggu ini, Andrew Warner ditemani oleh Adam Brakhane untuk menjelajahi repositori GitHub dan alat AI yang membuat pengembang lebih produktif. Mulai dari menjalankan aplikasi open-source di cloud hingga memberikan agen pengkodean AI kemampuan super dalam hal keamanan, desain, dan alur kerja, proyek-proyek ini menunjukkan apa yang mungkin terjadi ketika pengembang membangun di atas model AI terbaru. Episode dimulai dengan tiga penemuan bonus: cara untuk menghosting aplikasi open-source dengan murah, koleksi keterampilan pengkodean Theo, dan proyek yang memungkinkan pengembangan aplikasi iOS di Linux. Kemudian Andrew dan Adam menghitung mundur 10 repositori GitHub teratas minggu ini. Mereka mengeksplorasi kemampuan audit keamanan Cloudflare, Claude SEO, Impeccable, Superpowers, TradingAgents, kemampuan yang terinspirasi oleh Karpathy, alat desain dan aksesibilitas, T3 Code, dan Plugin Kerja Pengetahuan Anthropic. Sepanjang perjalanan, mereka membahas bagaimana proyek sumber terbuka dapat menjadi bisnis, mengapa kemampuan perlu disesuaikan dan diperbarui, bagaimana agen AI dapat dibujuk untuk mengikuti instruksi, dan mengapa alat pengkodean terbaik menjadi lebih dari sekadar asisten pengkodean. Repositori Unggulan: Pembuka Bonus - Every App: every-app/every-app - T3 Stack: t3dotgg/tstack - Omarchy Apple Dev: joshuaswarren/omarchy-apple-dev 10 Repositori GitHub Teratas - Cloudflare Security Audit Skill: cloudflare/security-audit-skill - Claude SEO: AgriciDaniel/claude-seo - Impeccable: pbakaus/impeccable - Superpowers: obra/superpowers - TradingAgents: TauricResearch/TradingAgents - Andrej Karpathy Skills: multica-ai/andrej-karpathy-skills - Taste Skill: Leonxlnx/taste-skill - Keterampilan Desain Antarmuka: jakubkrehel/skills - Kode T3: pingdotgg/t3code - Plugin Kerja Pengetahuan: anthropics/knowledge-work-plugins 00:00 - Repositori GitHub Terbaik Minggu Ini 00:27 - Every App: Host Aplikasi Open-Source seharga $5 00:57 - Mengapa Penyebaran yang Mudah Penting untuk Open Source 02:15 - T3 Stack: Keterampilan Pengkodean AI Theo 02:42 - Mengapa Anda Harus Menyesuaikan Keterampilan AI Anda 03:43 - Omarchy Apple Dev: Membangun Aplikasi iOS di Linux 04:29 - Mengapa Apple Membutuhkan Perangkat Keras Mac untuk Pengembangan iOS 05:43 - Keterampilan Audit Keamanan Cloudflare #1 07:00 - Claude SEO #2: Jalankan Audit SEO dengan Claude Kode 09:31 - #3 Sempurna: Berikan Agen AI Selera Desain yang Lebih Baik 10:35 - Bisakah Alat Desain Sumber Terbuka Menjadi Bisnis? 12:40 - Dipersembahkan oleh TripleTen 14:36 ​​- #4 Kekuatan Super: Berikan Agen Pengkodean Proses yang Lebih Baik 15:09 - Menggunakan Prinsip Persuasi untuk Meningkatkan Agen AI 17:11 - #5 TradingAgents: Simulasi Perdagangan AI 18:41 - #6 Keterampilan Andrej Karpathy 20:31 - Mengapa Agen AI Kesulitan Mengikuti Daftar Tugas 21:43 - Dipersembahkan oleh Zapier: Membuat Alur Kerja AI Lebih Murah 22:38 - #7 Keterampilan Selera: Berikan AI Anda Direktur Seni 24:36 - #8 Keterampilan Desain Antarmuka dan Aksesibilitas 26:20 - #9 Kode T3: Mengelola Beberapa Agen Pengkodean AI 27:20 - Menjalankan Agen AI di Beberapa Komputer 30:46 - #10 Plugin Kerja Pengetahuan Antropik 31:14 - Apa yang Harus Kita Bahas di Bintang Pertama? 31:45 - Lapisan Konteks Bisnis Bertenaga AI Adam 33:12 - Alur Kerja Pengeditan Video AI Otomatis Andrew 34:18 - Lebih Banyak Repositori GitHub untuk Dijelajahi 👉 Pertanyaan Media/Sponsor: thenextnewthing.ai 👉 Bergabunglah dengan kami: thenextnewthing.ai [it] Copia qualsiasi app, Claude fa SEO e le 10 migliori repository Link alla risorsa: thenextnewthing.ai - Questo video è sponsorizzato da TripleTen, una piattaforma di formazione professionale pensata per aiutare i professionisti del settore tecnologico e chi desidera cambiare carriera a trovare ruoli più remunerativi nel campo della tecnologia e dell'intelligenza artificiale. - Scopri due programmi TripleTen per ingegneri esperti: - Ingegneria dei sistemi di IA (22 settimane): tripleten.com - IA e apprendimento automatico (20 settimane): tripleten.com - Entrambi i programmi sono rivolti a ingegneri con almeno 2 anni di esperienza pratica recente in produzione e richiedono una breve valutazione delle competenze prima dell'iscrizione. Entrambi includono anche coaching e mentoring per aiutarti a distinguerti nel mercato del lavoro. - TripleTen - github.com/tripleten-com - Sito web di TripleTen - tripleten.com - Presentato da Zapier: zapier.com - Zapier Next Gen Zaps - zapier.com Questa settimana, Andrew Warner e Adam Brakhane esplorano i repository GitHub e gli strumenti di intelligenza artificiale che stanno rendendo gli sviluppatori più produttivi. Dall'esecuzione di app open source nel cloud al potenziamento di agenti di programmazione basati su IA con funzionalità avanzate di sicurezza, design e gestione dei flussi di lavoro, questi progetti dimostrano cosa è possibile realizzare quando gli sviluppatori si basano sui modelli di intelligenza artificiale più recenti. L'episodio inizia con tre scoperte bonus: un modo per ospitare app open source a basso costo, la raccolta di competenze di programmazione di Theo e un progetto che rende possibile lo sviluppo di app iOS su Linux. Infine, Andrew e Adam presentano la classifica dei 10 migliori repository GitHub della settimana. Esplorano la funzionalità di audit di sicurezza di Cloudflare, Claude SEO, Impeccable, Superpowers, TradingAgents, le competenze ispirate a Karpathy, gli strumenti di design e accessibilità, T3 Code e i plugin Knowledge Work di Anthropic. Nel corso della presentazione, discutono di come i progetti open source possano trasformarsi in attività commerciali, perché le competenze debbano essere personalizzate e costantemente aggiornate, come sia possibile persuadere gli agenti IA a seguire le istruzioni e perché i migliori strumenti di programmazione stiano diventando molto più che semplici assistenti alla programmazione. Repository in evidenza: Bonus Openers - Every App: every-app/every-app - T3 Stack: t3dotgg/tstack - Omarchy Apple Dev: joshuaswarren/omarchy-apple-dev I 10 migliori repository GitHub - Cloudflare Security Audit Skill: cloudflare/security-audit-skill - Claude SEO: AgriciDaniel/claude-seo - Impeccable: pbakaus/impeccable - Superpowers: obra/superpowers - TradingAgents: TauricResearch/TradingAgents - Andrej Karpathy Skills: Italiano: multica-ai/andrej-karpathy-skills - Competenza nel gusto: Leonxlnx/taste-skill - Competenze di progettazione dell'interfaccia: jakubkrehel/skills - Codice T3: pingdotgg/t3code - Plugin per il lavoro basato sulla conoscenza: anthropics/knowledge-work-plugins 00:00 - I migliori repository GitHub di questa settimana 00:27 - Every App: ospita app open source per 5 $ 00:57 - Perché la facilità di implementazione è importante per l'open source 02:15 - Stack T3: le competenze di programmazione AI di Theo 02:42 - Perché dovresti personalizzare le tue competenze AI 03:43 - Omarchy Apple Dev: crea app iOS su Linux 04:29 - Perché Apple richiede hardware Mac per lo sviluppo iOS 05:43 - #1 Competenza per l'audit di sicurezza di Cloudflare 07:00 - #2 Claude SEO: Eseguire audit SEO con Claude Code 09:31 - #3 Impeccable: Dare agli agenti IA un gusto migliore per il design 10:35 - Gli strumenti di design open source possono diventare aziende? 12:40 - Presentato da TripleTen 14:36 ​​- #4 Superpoteri: Migliorare il processo degli agenti di programmazione 15:09 - Utilizzare i principi della persuasione per migliorare gli agenti IA 17:11 - #5 TradingAgents: Simulazione di trading tramite IA 18:41 - #6 Competenze di Andrej Karpathy 20:31 - Perché gli agenti IA faticano a seguire le liste di attività 21:43 - Presentato da Zapier: Rendere i flussi di lavoro IA più economici 22:38 - #7 Abilità di gusto: Assegnare un direttore artistico alla tua IA 24:36 - #8 Competenze di progettazione dell'interfaccia e accessibilità 26:20 - #9 Codice T3: Gestire più agenti di programmazione IA 27:20 - Eseguire agenti IA su più computer 30:46 - #10 Plugin per il lavoro basato sull... [pt-BR] Copie qualquer aplicativo, Claude fazendo SEO e os 10 principais repositórios Link para o recurso: thenextnewthing.ai - Este vídeo é patrocinado pela TripleTen, uma plataforma de educação profissional criada para ajudar profissionais de tecnologia e pessoas em transição de carreira a ingressarem em cargos mais bem remunerados nas áreas de tecnologia e IA. - Explore dois programas da TripleTen para engenheiros experientes: - Engenharia de Sistemas de IA (22 semanas): tripleten.com - IA e Aprendizado de Máquina (20 semanas): tripleten.com - Ambos são destinados a engenheiros com mais de 2 anos de experiência prática recente em produção e exigem uma breve avaliação de habilidades antes da inscrição. Ambos também incluem coaching de carreira e mentoria para ajudar você a se destacar no mercado de trabalho. - TripleTen - github.com/tripleten-com - Site da TripleTen - tripleten.com - Apresentado por Zapier: zapier.com - Zaps de Próxima Geração do Zapier - zapier.com Nesta semana, Andrew Warner recebe Adam Brakhane para explorar os repositórios do GitHub e as ferramentas de IA que estão tornando os desenvolvedores mais produtivos. Desde a execução de aplicativos de código aberto na nuvem até a atribuição de superpoderes de segurança, design e fluxo de trabalho a agentes de codificação de IA, esses projetos mostram o que é possível quando os desenvolvedores criam com base nos modelos de IA mais recentes. O episódio começa com três descobertas bônus: uma maneira de hospedar aplicativos de código aberto a baixo custo, a coleção de habilidades de programação de Theo e um projeto que torna possível o desenvolvimento de aplicativos iOS no Linux. Em seguida, Andrew e Adam apresentam a lista dos 10 principais repositórios do GitHub da semana. Eles exploram a habilidade de auditoria de segurança da Cloudflare, o Claude SEO, o Impeccable, o Superpowers, o TradingAgents, habilidades inspiradas em Karpathy, ferramentas de design e acessibilidade, o T3 Code e os plugins de trabalho intelectual da Anthropic. Ao longo do caminho, discutem como projetos de código aberto podem se tornar negócios, por que as habilidades precisam ser personalizadas e mantidas atualizadas, como agentes de IA podem ser persuadidos a seguir instruções e por que as melhores ferramentas de programação estão se tornando mais do que apenas assistentes de codificação. Repositórios em destaque: Aberturas bônus - Every App: every-app/every-app - T3 Stack: t3dotgg/tstack - Omarchy Apple Dev: joshuaswarren/omarchy-apple-dev Top 10 Repositórios do GitHub - Cloudflare Security Audit Skill: cloudflare/security-audit-skill - Claude SEO: AgriciDaniel/claude-seo - Impeccable: pbakaus/impeccable - Superpowers: obra/superpowers - TradingAgents: TauricResearch/TradingAgents - Andrej Karpathy Skills: multica-ai/andrej-karpathy-skills - Habilidade de Paladar: Leonxlnx/taste-skill - Habilidades de Design de Interface: jakubkrehel/skills - Código T3: pingdotgg/t3code - Plugins de Trabalho Intelectual: anthropics/knowledge-work-plugins 00:00 - Os Melhores Repositórios do GitHub desta Semana 00:27 - Every App: Hospede Aplicativos de Código Aberto por US$ 5 00:57 - Por que a Facilidade de Implantação é Importante para o Código Aberto 02:15 - T3 Stack: As Habilidades de Programação de IA do Theo 02:42 - Por que Você Deve Personalizar Suas Habilidades de IA 03:43 - Omarchy Apple Dev: Crie Aplicativos iOS no Linux 04:29 - Por que a Apple Exige Hardware Mac para iOS Desenvolvimento 05:43 - #1 Habilidade de Auditoria de Segurança do Cloudflare 07:00 - #2 Claude SEO: Execute Auditorias de SEO com o Claude Code 09:31 - #3 Impecável: Dê aos Agentes de IA um Melhor Senso de Design 10:35 - Ferramentas de Design de Código Aberto Podem se Tornar Negócios? 12:40 - Apresentado pela TripleTen 14:36 ​​- #4 Superpoderes: Dê aos Agentes de Codificação um Processo Melhor 15:09 - Usando Princípios de Persuasão para Aprimorar Agentes de IA 17:11 - #5 TradingAgents: Simulação de Negociação por IA 18:41 - #6 Habilidades de Andrej Karpathy 20:31 - Por que os Agentes de IA Têm Dificuldade em Seguir Listas de Tarefas 21:43 - Apresentado pela Zapier: Torne os Fluxos de Trabalho de IA Mais Baratos 22:38 - #7 Habilidade de Paladar: Dê à Sua IA um Diretor de Arte 24:36 - #8 Habilidades de Design de Interface e Acessibilidade 26:20 - #9 Código T3: Gerencie Vários Agentes de Codificação de IA 27:20 - Execute Agentes de IA em Vários Computadores 30:46 - #10 Plugins de Trabalho de Conhecimento Antrópico 31:14 - O que Devemos Abordar no First Stars? 31:45 - Camada d...

3- Karpathy Skills: Modeller senin adına yanlış varsayımlar yapıyor ve kontrol etmeden onları takip ediyorlar. Kafalarını karıştırmıyorlar, açıklama aramıyorlar, tutarsızlıkları ortaya çıkarmıyorlar, takas sunmuyorlar, gerektiğinde karşı koymuyorlar. multica-ai/andrej-karpathy-skills

agent 干活进度看不见，交回来才发现跑偏？ multica 是人与 agent 同看板协作：派 issue 给 agent，跑在自有机器上，进度卡点都汇报，完事交回评审。 带 agent 团队干活的话，看板协作比私聊派活清楚。 先给 agent 派哪类活，修 bug 还是写功能？ #Agent协作 #开源 @grok multica-ai/multica

[개발자들의 PICK] 이번 주 깃허브 트렌딩 결산, 인기 오픈소스 TOP 10 이번 주 깃허브에서 가장 많은 별을 받은 오픈 소스 TOP 10을 10위부터 소개합니다. 소개한 저장소 링크는 고정 댓글에 있습니다. 00:00 이번 주 TOP 10 00:05 10위 안드로이드 스킬 00:51 9위 오르카 01:38 8위 오픈코드 02:15 7위 루플로 02:47 6위 그래피파이 03:32 5위 트레이딩 에이전츠 04:17 4위 카파시 스킬 04:56 3위 클로드 SEO 05:35 2위 지식 노동 플러그인 06:14 1위 클라우드플레어 시큐리티 오딧 스킬 06:55 마무리 📌 이번 주! 깃허브 트렌딩 TOP 10 저장소 10위 안드로이드 스킬 — android/skills 9위 오르카 — stablyai/orca 8위 오픈코드 — anomalyco/opencode 7위 루플로 — ruvnet/ruflo 6위 그래피파이 — Graphify-Labs/graphify 5위 트레이딩 에이전츠 — TauricResearch/TradingAgents 4위 카파시 스킬 — multica-ai/andrej-karpathy-skills 3위 클로드 SEO — AgriciDaniel/claude-seo 2위 지식 노동 플러그인 — anthropics/knowledge-work-plugins 1위 클라우드플레어 시큐리티 오딧 스킬 — cloudflare/security-audit-skill [en-US] [Developer's PICK] This Week's GitHub Trending Roundup, Top 10 Popular Open Source Projects Here are the Top 10 Open Sources with the Most Stars on GitHub this week, starting from 10th place. Links to the featured repositories are in the pinned comment. 00:00 This Week's Top 10 00:05 10th Android Skills 00:51 9th Orca 01:38 8th OpenCode 02:15 7th Luplo 02:47 6th Graphify 03:32 5th Trading Agents 04:17 4th Kapasi Skills 04:56 3rd Claude SEO 05:35 2nd Knowledge Work Plugin 06:14 1st Cloudflare Security Audit Skills 06:55 Wrap-up 📌 This Week! GitHub Trending Top 10 Repositories 10th Android Skills — android/skills 9th Orca — stablyai/orca 8th Open Code — anomalyco/opencode 7th Ruflo — ruvnet/ruflo 6th Graphify — Graphify-Labs/graphify 5th Trading Agents — TauricResearch/TradingAgents 4th Karpathy Skills — multica-ai/andrej-karpathy-skills 3rd Claude SEO — AgriciDaniel/claude-seo 2nd Knowledge Work Plugins — anthropics/knowledge-work-plugins 1st Cloudflare Security Audit Skills — cloudflare/security-audit-skill [id] [Pilihan Pengembang] Rekap Trending GitHub Minggu Ini, 10 Open Source Terpopuler Berikut adalah 10 Open Source Teratas dengan Bintang Terbanyak di GitHub minggu ini, dimulai dari peringkat ke-10. Tautan ke repositori unggulan ada di komentar yang disematkan. 00:00 10 Teratas Minggu Ini 00:05 Peringkat ke-10 Android Skills 00:51 Peringkat ke-9 Orca 01:38 Peringkat ke-8 OpenCode 02:15 Peringkat ke-7 Luplo 02:47 Peringkat ke-6 Graphify 03:32 Peringkat ke-5 Trading Agents 04:17 Peringkat ke-4 Kapasi Skills 04:56 Peringkat ke-3 Claude SEO 05:35 Peringkat ke-2 Knowledge Work Plugin 06:14 Peringkat ke-1 Cloudflare Security Audit Skills 06:55 Ringkasan 📌 Minggu Ini! 10 Repositori Terpopuler di GitHub 10 Android Skills — android/skills 9 Orca — stablyai/orca 8 Open Code — anomalyco/opencode 7 Ruflo — ruvnet/ruflo 6 Graphify — Graphify-Labs/graphify 5 Trading Agents — TauricResearch/TradingAgents 4 Karpathy Skills — multica-ai/andrej-karpathy-skills 3 Claude SEO — AgriciDaniel/claude-seo 2 Knowledge Work Plugins — anthropics/knowledge-work-plugins 1 Cloudflare Security Audit Skills — cloudflare/security-audit-skill

CLAUDE.md Best Practices: A Template for Unattended Agents Originally published at Unattended. CLAUDE.md best practices come down to three things. Keep the file short. Include only what Claude can't work out from the code. Write it so that a new agent with nobody around to ask can still finish the job. For unattended runs, that last point carries the most weight: the file should cover what to decide alone, what to stop on, and where to leave a note. If you only copy a style guide, the agent drifts, stalls, or keeps hitting permission denials overnight. Anthropic's own test for each line is: "Would removing this cause Claude to make mistakes?" Which option is right for you? Choose by how Claude will run, not by how big the project is. If you're unsure, start with the first row. | Your situation | What to pick first | First page to open | |---|---|---| | Beginner, no CLAUDE.md yet | Run /init, then cut the draft down to commands, rules, and "never" lines | Best practices: Write an effective CLAUDE.md | | Daily interactive work | A short project CLAUDE.md in git, plus skills for workflows you only need sometimes | Skills docs | | Nightly cron or claude -p job | The unattended template below, plus permission rules in settings | Headless mode | | CI with --bare | Pass the file yourself with --append-system-prompt-file | Bare mode section | | Tight token budget | Move rarely used content into skills or @ imports that only load when needed | /claude-code-how-to-see-token-usage-and-reduce-it/ | | Rules that must never break | A PreToolUse hook or a deny rule instead of a CLAUDE.md line | Hooks guide | Here's why, one line each. /init gives you a draft quickly, but the draft is too long. Skills keep everyday sessions lean. Unattended runs can't ask you anything. --bare doesn't load CLAUDE.md at all. Every line in CLAUDE.md costs tokens on every run. And a hook is the only kind of rule that can't be talked out of. What should a CLAUDE.md file example look like? A good CLAUDE.md example is short and scoped. It covers project context, commands, conventions, and explicit rules for what the agent must never do unattended. It isn't a full style guide. Anthropic's own sample is just two headings, "Code style" and "Workflow", with two bullets each. The docs say "There's no required format for CLAUDE.md files, but keep it short and human-readable." An unattended agent needs a few extra sections that an interactive user never thinks about, because the interactive user is the one answering questions. Here is an example for a scheduled job. It's illustrative, so replace the commands and paths with your own: # Project Static blog generator (Node 22). Content in /content, build output in /dist. # Commands - Install: npm ci - Test: npm test -- --run - Build: npm run build (must exit 0 before any commit) # Conventions - ES modules only. No new dependencies without a note in REPORT.md. # Autonomy: decide alone - Fix failing tests caused by your own change. - Retry a flaky network step once, then move on. # Autonomy: stop and report - Anything touching /migrations, .env, or CI config. - A task that needs more than ~10 files changed. - Any ambiguity you would normally ask about. # When blocked - Do not wait for an answer. Write the question and what you tried to BLOCKED.md, commit nothing, and end the run. # Done means - Tests pass, build exits 0, REPORT.md lists files changed and commands run. # Compaction - When compacting, always preserve the full list of modified files and any test commands. Each section maps to a documented behavior. "Done means" comes from Anthropic's advice to give Claude a check it can run. Their wording: it's "the difference between a session you watch and one you walk away from." The compaction line is close to word for word what the best practices page suggests for keeping critical context through summarization. The "When blocked" section exists because a run started with --permission-prompts none removes tools that need a person, such as AskUserQuestion, according to the headless docs. Without that section, Claude has no planned route when it wants to ask you something. Common pitfalls in examples: - Putting permissions in prose. "Never run rm -rf" in CLAUDE.md is a request, not a block. Use a deny rule or hook instead (see /claude-code-permissions-best-practices-headless-template/). - Describing the codebase file by file. Anthropic's exclude list names this directly, along with "Self-evident practices like 'write clean code'" (best practices). - No definition of done. Without one, "Claude stops when the work looks done." What are the core CLAUDE.md best practices? Keep CLAUDE.md concise, include only information that changes decisions, and write it as onboarding for an agent that can't ask questions. The reason is the context window. Anthropic says "Claude's context window fills up fast, and performance degrades as it fills," and CLAUDE.md loads at the start of every session. The memory docs add that Claude treats these files as "context, not enforced configuration." Every line costs tokens, and none of them is guaranteed to be followed. Anthropic's include/exclude table (source), rearranged for unattended use: | Include | Exclude | |---|---| | Bash commands Claude can't guess | Anything Claude can figure out by reading code | | Code style rules that differ from defaults | Standard language conventions | | Test instructions and preferred runner | Detailed API docs (link instead) | | Branch naming and PR conventions | Information that changes frequently | | Project-specific architecture decisions | Long explanations or tutorials | | Required env vars and environment quirks | File-by-file descriptions | | Non-obvious gotchas | "Write clean code" | Steps to write or rewrite yours: 1. In the project folder, run claude, then type /init to generate a starter file (best practices). 2. Delete every line that fails the test "Would removing this cause Claude to make mistakes?" 3. Add the autonomy, "when blocked", and "done means" sections from the example above. 4. Run /context to confirm the file loaded (best practices). 5. For a checked-in file, run /doctor, and Claude proposes cuts for content it can work out from the codebase (best practices). 6. Commit it to git. Two failure signals tell you which way to fix the file. Anthropic puts it this way: "If Claude keeps doing something you don't want despite having a rule against it, the file is probably too long." If Claude asks about something the file already answers, the wording is probably ambiguous (best practices). Use emphasis sparingly. Mark only the one rule Claude keeps skipping with "IMPORTANT", because "If you emphasize many lines, none of them stands out." What CLAUDE.md patterns do top GitHub repos use? Popular GitHub repos usually use three CLAUDE.md patterns: a command-and-convention sheet, a behavior-principles file, or a shared team file updated after every mistake. | Pattern | Public example | What it contains | Good for unattended runs? | |---|---|---|---| | Command and convention sheet | EDSL file in awesome-claude-code | Build/test commands, strict style (Black, mypy), testing requirements | Yes, if you add stop rules | | Behavior principles | multica-ai/andrej-karpathy-skills | Four principles about assumptions, simplicity, edit scope, goals | Partly; it has no project commands | | Living team file | Claude Code team's own file (Boris Cherny on X) | Corrections added whenever Claude gets something wrong | Yes; it grows from real failures | A few structural habits show up again and again. One is splitting rules by directory or file type: Anthropic documents nested CLAUDE.md files and a .claude/rules/ folder for rules scoped to specific files (memory docs). Another is pointing to longer docs with @path/to/import instead of pasting them in (best practices). Imports help with organization. They don't make the file cheaper if the imported file loads every session anyway. If your repo already has an AGENTS.md, the memory docs say Claude "can also read a repository's AGENTS.md files in place of CLAUDE.md," so you don't need two copies. Limitation: most public examples are written for someone sitting at the terminal. Hardly any say what the agent should do when it's stuck at 3 a.m., and that's the section you'll need to write yourself. What does Anthropic's official guidance say about CLAUDE.md? Anthropic's official guidance treats CLAUDE.md as a short, advisory memory file that loads every session. It covers commands, conventions, and gotchas, and you refine it like code. The points that matter most for automation, all from Anthropic's docs as of October 2026: - Scopes. CLAUDE.md can live at managed-policy, user (~/.claude/CLAUDE.md), project (./CLAUDE.md or ./.claude/CLAUDE.md), and local (./CLAUDE.local.md) levels (memory docs). On a VPS, the user-level file belongs to whatever account the cron job runs as. That may not be the account you set up interactively. - Auto memory is separate. Claude writes its own notes, and only the first 200 lines or 25KB load each session (memory docs). - Hooks vs instructions. "Unlike CLAUDE.md instructions which are advisory, hooks are deterministic and guarantee the action happens". To block an action no matter what Claude decides, the memory docs point to a PreToolUse hook (memory docs). - Treat it like code. "Review it when things go wrong, prune it regularly, and test changes by observing whether Claude's behavior actually shifts" (best practices). The headless page is where unattended setups most often go wrong. Adding --bare skips auto-discovery of "hooks, skills, custom commands, subagents, installed plugins, MCP servers, auto memory, and CLAUDE.md" (headless docs). The same page warns that "--bare is the recommended mode for scripted and SDK calls, and will become the default for -p in a future release." So a cron job that depends on CLAUDE.md being picked up automatically may stop following it after an update, with no error to tell you. To make the file load whatever the default turns out to be: 1. Keep the unattended rules in a file, for example CLAUDE.md or a separate AGENT_RULES.md. 2. Pass it explicitly. The headless docs list --append-system-prompt-file for adding to the system prompt in bare mode: claude --bare -p "Run the nightly content check" \ --append-system-prompt-file ./CLAUDE.md \ --allowedTools "Read,Edit,Bash(npm test *)" \ --permission-prompts none 3. Remember that bare mode doesn't use your subscription login. Set ANTHROPIC_API_KEY first (headless docs). Two more defaults matter. With Claude Code v2.1.302 or later, claude -p starts in Manual mode when nothing sets a permission mode, and a call that needs approval is denied unless something approves it. --permission-prompts requires v2.1.259 or later (headless docs). This is why CLAUDE.md has to tell the agent what to do after a denial. The full flag set is covered in /claude-code-headless-mode-flags-auth-and-cost-2026/. What do Reddit discussions say works (and fails) in CLAUDE.md? The usual complaint on Reddit and developer forums is that long CLAUDE.md files get ignored. The fix people recommend is to trim the file to essentials and prune it regularly. Plenty of community posts carry titles like "Why your long CLAUDE.md gets ignored, and how to fix it", and forum threads ask whether popular rule files actually help (Devtalk). Treat these as anecdotes, not measurements. The pattern does line up with Anthropic's own failure list, which names "The over-specified CLAUDE.md" and says important rules "get lost in the noise". What community advice tends to get right: - Add rules only after a real mistake. That matches the practice used by the Claude Code team (next sections). - Turn rules into hooks. Anthropic's own fix for an over-specified file: "If Claude already does something correctly without the instruction, delete it or convert it to a hook." What it tends to get wrong for unattended work: - Shouting. Files full of "ALWAYS" and "NEVER" in capitals cancel themselves out, since emphasis only works on one line. - Language and tone rules at the top. In a nightly job nobody reads the tone. Spend those lines on stop conditions instead. For example, say your overnight agent keeps editing CI config even though a rule forbids it. Community advice would tell you to move the rule higher in the file. The more reliable fix is a deny rule or a PreToolUse hook on that path, with one line left in CLAUDE.md explaining why. Copy-paste hook scripts for this are in /claude-code-hooks-examples-7-copy-paste-scripts/. What does Andrej Karpathy say about CLAUDE.md style files? Karpathy didn't write a CLAUDE.md himself. The popular "Karpathy CLAUDE.md" is a community file built from his posts on how LLMs fail at coding, and it focuses on behavior rather than style. The repo multica-ai/andrej-karpathy-skills showed about 223.8k stars when checked in October 2026. It turns his observations into four principles: Think Before Coding, Simplicity First, Surgical Changes, and Goal-Driven Execution. The line that matters most for unattended runs is the one the repo quotes from Karpathy: "Don't tell it what to do, give it success criteria and watch it go." How to use it in a scheduled setup: 1. Keep "Surgical Changes" and "Goal-Driven Execution". They limit drift and define when the agent is finished. 2. Rewrite "Think Before Coding" for unattended runs. The original tells the agent to ask for clarification when a request is ambiguous. Overnight, nobody is there to answer, so change that to "write the question to BLOCKED.md and stop." 3. Add your own commands. The file has no project-specific build or test lines. Who it's not for: if your CLAUDE.md is already near the size where rules get lost, pasting in a generic principles file makes that worse. Take the two principles you need and leave the rest. How is CLAUDE.md best practice evolving for 2026 (skills, headless agents)? In 2026, CLAUDE.md is shrinking. Workflows are moving into skills, hard rules into hooks and permission settings, and the core file is turning into a statement of autonomy boundaries. The split, according to Anthropic's docs as of October 2026: | Content | Put it in | Why | |---|---|---| | Always-true commands, conventions, stop rules | CLAUDE.md | "CLAUDE.md is loaded every session, so only include things that apply broadly" | | Occasional domain knowledge or workflows | .claude/skills/<name>/SKILL.md | Claude loads skills on demand (best practices) | | Workflows with side effects | Skill with disable-model-invocation: true | Only runs when you trigger it (best practices) | | Must-never-happen actions | Hooks, deny rules | Deterministic (hooks guide) | | "Keep going until tests pass" | /goal condition or a Stop hook | Lets "an unattended run finish correctly without you" | Skills also work headless. User-invoked skills run in -p mode if you include /skill-name in the prompt string (headless docs). A clean nightly setup looks like this: 1. CLAUDE.md, about a screen long: commands, autonomy rules, "when blocked", "done means". 2. A skill such as /nightly-report that holds the step-by-step procedure. 3. Cron calls claude -p "/nightly-report" --permission-mode dontAsk --allowedTools "...". Under dontAsk, Claude Code "denies every call that would otherwise prompt" (headless docs). 4. A Stop hook runs your test command, so the run can't end while tests are red. Auto mode is the other 2026 change. From v2.1.283 it's the starting permission mode for interactive terminal and VS Code sessions, with a classifier that blocks actions like scope escalation (best practices). It reduces prompts, but it doesn't replace stop rules. The classifier judges risk. It doesn't know your business rules, such as "never publish on Sundays". Scheduling details are in /claude-code-cron-job-setup-run-automations-overnight/. What do practitioner write-ups (Medium, Boris Cherny) add to best practices? Practitioners, including Claude Code creator Boris Cherny, add one habit: update CLAUDE.md after every mistake you catch. Keep the file small enough that the new rule doesn't get lost. Cherny described the Claude Code team's setup on X in January 2026. They share one CLAUDE.md for the repo, check it into git, and "the whole team contributes multiple times a week". Whenever Claude gets something wrong, they add it to the file. In a follow-up post, he said he tags @.claude on coworkers' PRs to add CLAUDE.md lines through the Claude Code GitHub Action. For very long tasks, he mentioned having Claude verify its work or using a Stop hook (thread). Adapted for unattended runs: 1. Each morning, read the run's REPORT.md and BLOCKED.md. 2. For each real mistake, decide on the fix. A one-line CLAUDE.md rule works if the mistake was a judgment call. A hook or deny rule is better if it must never happen again. 3. Add the line in the same commit as the fix. 4. Once a month, delete rules the agent now follows without being told. Pitfall: a file that only grows. Adding a line after every mistake without ever pruning produces exactly the bloated file Anthropic warns about. A related mistake is reacting to a stalled run by adding more instructions. Stalls usually come from permissions or context, not missing rules. Check /why-claude-code-stops-mid-automation-and-how-to-fix-it/ before you edit the file. Next steps Do today (DIY): 1. Open your project, run /init (or open your existing file with /memory), and cut every line that fails "Would removing this cause Claude to make mistakes?" 2. Add four sections from the example: "Autonomy: decide alone", "Autonomy: stop and report", "When blocked", and "Done means". 3. If your cron job uses -p, add --append-system-prompt-file ./CLAUDE.md now, so the coming --bare default doesn't silently drop the file. If you'd rather not do it yourself: Anthropic's paid plans include Claude Code. Pro is $20/month ($17/month billed annually), and Max starts at $100/month (as of October 2026; check the official pricing page). Pro suits one person running light scheduled jobs. It's not a good fit if you need many long runs per night, because Claude Code "shares the same usage limits as the rest of your plan." Bare-mode runs bill through an API key instead, so they don't draw on a subscription. Anthropic also publishes a Claude Code 101 course on CLAUDE.md; check the page for current access terms. It suits beginners, but it won't write your stop rules for you. Related guides on this site: - /claude-code-permissions-best-practices-headless-template/ - /claude-code-hooks-examples-7-copy-paste-scripts/ - /claude-code-headless-mode-flags-auth-and-cost-2026/ - /claude-code-cron-job-setup-run-automations-overnight/ - /why-claude-code-stops-mid-automation-and-how-to-fix-it/ FAQ Where can I find CLAUDE.md best practices examples? Anthropic's best practices page has a short official sample and an include/exclude table. Community examples are collected in awesome-claude-code. The template in this guide adds the unattended sections most public examples leave out. What are Anthropic's CLAUDE.md best practices? Keep it short. Include only commands, rules that differ from defaults, and gotchas. Check it into git and prune it regularly. Anthropic also says CLAUDE.md is advisory, so anything that must always happen belongs in a hook (best practices). Are there good CLAUDE.md best practices on GitHub? Yes. The most-starred example is multica-ai/andrej-karpathy-skills, a behavior-principles file. Language-specific files such as the EDSL example in awesome-claude-code show the command-sheet pattern. Either way, add your own commands and stop rules. What are the Karpathy CLAUDE.md best practices? They're four principles taken from Karpathy's posts: Think Before Coding, Simplicity First, Surgical Changes, and Goal-Driven Execution. Karpathy didn't publish the file himself; the multica-ai repo did. For unattended runs, change "ask for clarification" to "log the question and stop". Sources - Best practices for Claude Code (Anthropic) - How Claude remembers your project (Anthropic memory docs) - Run Claude Code programmatically (headless docs) - Hooks guide (Anthropic) - Skills (Anthropic) - Claude pricing - Claude Code 101: The CLAUDE.md file (Anthropic Academy) - Boris Cherny on X: shared CLAUDE.md - Boris Cherny on X: tagging @.claude in PRs - Boris Cherny on X: long-running tasks - multica-ai/andrej-karpathy-skills (GitHub) - awesome-claude-code EDSL CLAUDE.md (GitHub) - Why your long CLAUDE.md gets ignored (DEV Community) - Has anyone tried the Karpathy CLAUDE.md rules? (Devtalk) - Writing the Best CLAUDE.md (DataCamp) - Creating the Perfect CLAUDE.md (Dometrain) --- This article was created with the help of AI. Sources are linked inline; check the official documentation for the latest details.

Karpathy's 4 Rules That Fix AI Coding | Repos You Should Know Day 1 This GitHub repo doesn't have a single line of code, and it has over 220,000 stars. Day 1 of finding repositories you should know about: Karpathy Skills. It's one CLAUDE.md file, built from Andrej Karpathy's observations on where AI coding goes wrong. Once it's in your project, Claude Code follows 4 principles: 1. Think Before Coding: don't assume, surface tradeoffs, ask when unsure 2. Simplicity First: minimum code that solves the problem, nothing speculative 3. Surgical Changes: touch only what you must, clean up only your own mess 4. Goal-Driven Execution: define success criteria, loop until verified 🔗 Repo: multica-ai/andrej-karpathy-skills ⚡ Install in Claude Code: /plugin marketplace add forrestchang/andrej-karpathy-skills /plugin install andrej-karpathy-skills@karpathy-skills Using Cursor? The repo ships a ready-made rule: .cursor/rules/karpathy-guidelines.mdc Subscribe so you don't miss Day 2 🔔 📲 Follow me: Instagram: instagram.com X: x.com YouTube: youtube.com #Shorts #ClaudeCode #AICoding #GitHub #AndrejKarpathy #Cursor #AIAgents

What an Empty Prompt Costs: Measure the Baseline Tokens of 4 Coding Agent CLIs - Book: LLM Observability Pocket Guide - Also by me: Thinking in Go (2-book series) — Complete Guide to Go Programming + Hexagonal Architecture in Go - My project: Hermes IDE | GitHub — an IDE for developers who ship with Claude Code and other AI coding tools - Me: xgabriel.com | GitHub --- OpenAI's docs for codex exec --json print a sample usage line: 24,763 input tokens, 24,448 of them cached, and 122 output tokens (Codex non-interactive docs). The page doesn't say what prompt produced it. Whatever it was, the model read about 200 tokens for every token it wrote. Every coding agent CLI sends a prefix before your task says a word: its system prompt, the definitions of its built-in tools, the tools of the MCP servers you configured, and every instruction file it found on the way in. A fresh session sends that prefix again, and so does every headless run in CI. Don't copy anyone's number for this, including the two samples quoted here. The prefix moves with the CLI version, the model, your global config and the folder you start in. So measure it: a one-word prompt through Claude Code, Codex, GitHub Copilot CLI and OpenCode, then the same prompt after adding one MCP server and one AGENTS.md. Every number in your table comes from your own machine. What rides along before your prompt The system prompt and built-in tools belong to the vendor. MCP servers and instruction files are yours. MCP servers behave differently per CLI. Claude Code defers MCP tool definitions by default: "Only tool names and server instructions load at session start" (MCP docs). OpenCode's docs warn the opposite way: an MCP server "adds to the context," and some, like the GitHub MCP server, "tend to add a lot of tokens" (OpenCode MCP docs). Copilot CLI ships with the GitHub MCP server built in (GitHub docs). Codex's MCP page doesn't describe deferral either way (Codex MCP docs), which is one more reason to measure. Instruction files are where an "empty" prompt stops being empty. Each CLI reads more files than its own: | CLI | Instruction files it loads at startup | |---|---| | Claude Code | CLAUDE.md files, or AGENTS.md when no project CLAUDE.md exists (v2.1.277+), plus auto memory (docs) | | Codex | ~/.codex/AGENTS.md, then AGENTS.md from the git root down to your folder, up to 32 KiB (docs) | | Copilot CLI | AGENTS.md, .github/copilot-instructions.md, ~/.copilot/copilot-instructions.md, CLAUDE.md and GEMINI.md, combined (docs) | | OpenCode | AGENTS.md (falling back to CLAUDE.md), plus ~/.config/opencode/AGENTS.md, or ~/.claude/CLAUDE.md if that one is missing (docs) | Read the last two rows twice. Without a global OpenCode file, your personal ~/.claude/CLAUDE.md rides along in OpenCode too, and a repo's CLAUDE.md lands in Copilot CLI's prefix as well as Claude's. !A tiny blank speech bubble perched on top of a tall stack of crates, gears and folded documents on a conveyor belt Your one-word prompt is the small part of the first request. Run 1: one word, four CLIs Start in an empty git repo, so no project file sneaks in. Your global files still load, and that's the point: this is the baseline you pay for today. mkdir -p ~/tmp/baseline && cd ~/tmp/baseline git init -q Claude Code and OpenCode also search parent folders, so check that ~ and ~/tmp hold no CLAUDE.md or AGENTS.md (Claude Code memory docs, OpenCode rules). Codex stops at the git root. Claude Code. --output-format json returns usage, total_cost_usd and a per-model modelUsage breakdown (headless docs, cost tracking). The cost is a client-side estimate, not your bill. The commands below use jq to read the JSON. claude -p "ping" --output-format json > claude-1.json jq '.usage, .total_cost_usd' claude-1.json jq '[.usage.input_tokens, .usage.cache_creation_input_tokens, .usage.cache_read_input_tokens] | add' claude-1.json That second command matters. In Anthropic's usage object, input_tokens counts only what came after the last cache breakpoint, so the full input is the sum of all three fields (prompt caching docs). Read input_tokens alone and the baseline looks tiny. Also check .num_turns: if the agent ran a tool to answer ping, the usage covers more than one model request and counts the prefix more than once. For a floor without your hooks, plugins, MCP servers and CLAUDE.md, add --bare. It needs ANTHROPIC_API_KEY, because bare mode skips your subscription login (headless docs). Codex. codex exec --json streams JSONL, and turn.completed carries input_tokens, cached_input_tokens, output_tokens and reasoning_output_tokens (docs). codex exec --json "ping" > codex-1.jsonl jq -c 'select(.type=="turn.completed") | .usage' \ codex-1.jsonl Here the cached count sits inside the input count: OpenAI's caching guide subtracts cached tokens from input to get the uncached part. So input_tokens is your prefix plus the one-word prompt, as long as the turn made a single model call. Copilot CLI. The programmatic reference documents --output-format json but no usage fields. Its -s flag suppresses stats, so leave it off and the stats print after the reply (reference). copilot -p "ping" 2>&1 | tee copilot-1.txt GitHub's changelog says /usage, and the summary printed when a session ends, include "a break down of token usage for each model" (changelog). Check whether your -p output prints the same breakdown. For the split by layer, start copilot and type /context. It lists System Prompt, Custom Instructions, System Tools, MCP Tools and Messages separately (GitHub docs). OpenCode. opencode run --format json emits raw JSON events (CLI docs). OpenCode records each step's tokens as input, output, reasoning and cache.read/cache.write. In a session posted to OpenCode's tracker, the stored step showed input 209 and cache read 18,944, so add the cache fields to input. opencode run "ping" --format json > opencode-1.jsonl jq -c 'select(.type=="step_finish") | .part.tokens' \ opencode-1.jsonl If that prints nothing, you may be on a v2 build where step_finish never reaches stdout (open report). opencode stats --days 1 reads token usage from your stored sessions instead. Run each command twice within a few minutes. The first run tends to write the prefix to the cache, and the second reads it back. On Claude models, cache reads bill at 0.1x the input rate or less (Anthropic), so a cold first session costs more than a warm repeat. Here is which number to write down from each CLI: | CLI | Full input for the turn | Where to read it | |---|---|---| | Claude Code | input_tokens + cache_creation_input_tokens + cache_read_input_tokens | .usage in the --output-format json result | | Codex | input_tokens (cached tokens are already inside it) | turn.completed event | | Copilot CLI | the per-layer lines | /context in an interactive session | | OpenCode | input + cache.read + cache.write | step_finish events, or opencode stats | Run 2: one MCP server and one AGENTS.md Now add the two things most repos grow first. For the MCP server, use the reference "everything" server. Its README calls it "a test server for builders of MCP clients." Or use a server you actually run. For the memory file, write a short AGENTS.md. Better still, copy the one from a work repo. In this repo all four CLIs read it, since there's no CLAUDE.md here or above it (Claude Code needs v2.1.277 or later for that). cat > AGENTS.md <<'EOF' # Project rules - Run `npm test` before you say a task is done. - Never edit files under migrations/ without asking. EOF Claude Code takes a one-off server with --mcp-config, and with -p it waits for the server to connect before the first turn (headless docs). Save this as mcp.json: { "mcpServers": { "everything": { "type": "stdio", "command": "npx", "args": ["-y", "@modelcontextprotocol/server-everything"] } } } Then run it twice: once as shipped, once with tool search off, so you see what deferral saves you. claude -p "ping" --output-format json \ --mcp-config mcp.json > claude-2.json ENABLE_TOOL_SEARCH=false claude -p "ping" \ --output-format json --mcp-config mcp.json \ > claude-2-upfront.json ENABLE_TOOL_SEARCH=false loads every MCP tool upfront (MCP docs). The gap between those two files is the price of one server's tool definitions on every fresh session. Claude Code skips that cost by default. Codex: register the server, then rerun the Run 1 command into codex-2.jsonl. Delete the [mcp_servers.everything] table from ~/.codex/config.toml when you're done. codex mcp add everything -- \ npx -y @modelcontextprotocol/server-everything Copilot CLI: add the server with /mcp add, or put the same mcpServers block, with "type": "local", in ~/.copilot/mcp-config.json (GitHub docs). Rerun, then check /context to see the MCP Tools line grow. Remove the everything entry from ~/.copilot/mcp-config.json when you're done. OpenCode: put the server in the repo's opencode.json and rerun. { "$schema": "https://opencode.ai/config.json", "mcp": { "everything": { "type": "local", "command": ["npx", "-y", "@modelcontextprotocol/server-everything"], "enabled": true } } } Turn the delta into money The gap between Run 1 and Run 2 is what one MCP server and one memory file add to every session start. Multiply it the way your bill does: prefix tokens, times fresh sessions per day, times your model's input or cache-write rate, plus cache reads for every later turn that re-sends the prefix. Headless runs count, and so do parallel agents. Claude Code's docs say agent-team teammates "load CLAUDE.md, MCP servers, and skills automatically" (costs docs). How that turns into money depends on the plan. On a Claude Pro or Max subscription, the dollar figure in /usage "isn't relevant for billing purposes," but the tokens still draw down your plan limits (costs docs). On Copilot, plans moved to GitHub AI Credits on June 1, 2026, "calculated based on token consumption, including input, output, and cached tokens" (GitHub changelog). Annual Pro and Pro+ plans stay on premium requests until they expire (same changelog). On an API key, every token is billed. If the delta surprises you, the fixes are boring. Claude Code's docs suggest moving workflow-specific instructions out of CLAUDE.md into skills that load on demand, and keeping CLAUDE.md under 200 lines (costs docs). Turn off MCP servers you aren't using. Check your global files, because they ride along in every repo. Watching the baseline in Hermes IDE I build Hermes IDE, so weigh this section with that in mind. Hermes runs each agent in a real terminal by default, and for every session it shows spend: the cost the agent reports, or an estimate from its transcript, marked as estimated, with token totals (README). Since 2.3.4, released October 15, 2026, there's also a context meter: it shows how full the conversation is and offers Compact when it gets close to full. Send the same ping as the first message of a session, and the context meter, the token total and the cost show your baseline before you've given the agent any real work. In the screenshot below, pressing the Context chip above a Claude Code session opened Claude Code's own /context breakdown, and each session's context gauge sits in the sidebar. !Hermes IDE with a Claude Code session showing its /context breakdown, per-session context gauges in the sidebar, and a status strip with token total and estimated cost Hermes 2.3.4: Claude Code's /context breakdown (44.4k of 1M tokens, 4%) after pressing the Context chip, with per-session context gauges in the sidebar and the token total and estimated cost in the status strip. Hermes adds a layer of its own. Its project scan points the agent at a context file, kept within a token budget you can set per project, and saved memory facts and pinned files reach every later session (README). Run the ping in a Hermes session and in a plain terminal on the same repo, and the difference shows what Hermes adds to the first request, if anything. !Hermes IDE with Codex and Claude Code sessions side by side, each in its own hermes/ worktree, and a sidebar listing each session's status, context use, cost and branch Hermes 2.3.4: Codex and Claude Code side by side, each in its own hermes/ worktree, with each session's status, context use, cost and branch in the sidebar. Hermes is free, source-available under BSL 1.1, and runs on macOS, Windows and Linux. It isn't affiliated with Nous Research, Anthropic, OpenAI or any agent vendor. Download it or read the source on GitHub. --- If this was useful Measuring the baseline is the first token-cost question. The next ones are which session, tool call or teammate spent the rest, and whether the spend bought better answers. That's what the LLM Observability Pocket Guide covers: picking the tracing and evals tools that answer those questions for your team. If you'd rather watch the gauges than parse JSON, Hermes IDE puts them next to each agent.