Advanced AI Onboarding | AI Models Made Useful

Start hereEmpieza aquí

The advanced map: context, access, actionEl mapa avanzado: contexto, acceso y acción

Advanced AI work is not about memorizing every product name. Know where the useful context lives, which AI can see it, and whether the AI should answer, connect, or act.El trabajo avanzado con IA no consiste en memorizar productos. Identifica dónde vive el contexto útil, qué IA puede verlo y si debe responder, conectarse o actuar.

Product reading key: company, model, app, tool, agent, ADEClave de lectura: empresa, modelo, app, herramienta, agente y ADE

The starting guide defines the terms. Here, use them to read each company map without mixing the reasoning engine with the product surface or its permissions.La guía inicial define los términos. Aquí úsalos para leer cada mapa empresarial sin confundir el motor de razonamiento con el producto o sus permisos.

CompanyEmpresa

Owns or hosts a family of models and products.Crea o aloja una familia de modelos y productos.

ModelModelo

Generates the answer, decision, or tool call.Genera la respuesta, decisión o llamada de herramienta.

AppApp

Adds the interface, account, context, and product policy.Añade interfaz, cuenta, contexto y política del producto.

ToolHerramienta

Reads context or performs one permissioned action.Lee contexto o realiza una acción con permisos.

AgentAgente

Chooses steps and tools inside a goal-and-stop loop.Elige pasos y herramientas dentro de un bucle con meta y detención.

ADE

Coordinates agent tasks, workspaces, review, and delivery.Coordina tareas de agentes, espacios, revisión y entrega.

1. Locate1. Ubica

Where is the context?¿Dónde está el contexto?

Email, Teams, SharePoint, Drive, local files, notes, source packets, screenshots, or a repository.Correo, Teams, SharePoint, Drive, archivos locales, notas, fuentes, capturas o un repositorio.

2. Match2. Conecta

Which AI can see it?¿Qué IA puede verlo?

Choose the surface already connected to the context instead of copying everything into a random chat.Elige el espacio que ya tiene acceso al contexto en vez de copiarlo todo a un chat cualquiera.

3. Choose3. Elige

Answer or action?¿Respuesta o acción?

A chat answer is different from an agent that can edit files, run commands, create branches, or call tools.Una respuesta de chat no es lo mismo que un agente capaz de editar archivos, ejecutar comandos, crear ramas o usar herramientas.

4. Check4. Revisa

Review the real thingsRevisa lo que realmente importa

Check source files, recipients, branch diffs, dates, numbers, permissions, and anything the AI had to infer.Comprueba fuentes, destinatarios, cambios de rama, fechas, cifras, permisos y todo lo que la IA haya tenido que inferir.

The bridge to buildingEl puente hacia construir

From an advanced AI tool to a trustworthy AI systemDe una herramienta de IA avanzada a un sistema de IA confiable

Once you can route context and permissions, the next step is understanding the small system around the model. This map keeps the path compact: open the parts you need, then prove the boundary before you add more autonomy.Cuando puedes enrutar contexto y permisos, el siguiente paso es entender el pequeño sistema que rodea al modelo. Este mapa mantiene el camino compacto: abre solo las partes que necesitas y demuestra el límite antes de añadir más autonomía.

1 · Job1 · Trabajo

User + taskUsuario + tarea

Name the outcome, inputs, and what “done” means.Nombra el resultado, las entradas y qué significa «terminado».

2 · Call2 · Llamada

API + modelAPI + modelo

Send a bounded request; handle timeouts, retries, and state explicitly.Envía una solicitud acotada; maneja explícitamente tiempos de espera, reintentos y estado.

3 · Contract3 · Contrato

Output + toolsSalida + herramientas

Use a schema and narrow tool arguments, not a hopeful paragraph.Usa un esquema y argumentos de herramientas acotados, no un párrafo esperanzador.

4 · Context4 · Contexto

Retrieve + groundRecuperar + fundamentar

Find trusted evidence, show its boundary, and keep access rules intact.Encuentra evidencia confiable, muestra su límite y conserva las reglas de acceso.

5 · Evaluate5 · Evaluar

Measure the jobMide el trabajo

Test correctness, groundedness, safety, latency, and cost against examples.Prueba corrección, fundamentación, seguridad, latencia y costo con ejemplos.

6 · Guard6 · Proteger

Permissions + stopPermisos + detención

Keep humans at consequential boundaries and make failure safe.Mantén a las personas en los límites importantes y haz que el fallo sea seguro.

7 · Loop7 · Bucle

Workflow or agent?¿Flujo o agente?

Add iteration only when the job needs it and the boundary is observable.Añade iteración solo cuando el trabajo la necesite y el límite sea observable.

You can stop here when…Puedes detenerte aquí cuando…

You can describe the job, supply trusted context, validate the output, and keep a person or a safe stop before consequential actions. That is already enough to build a useful AI feature without calling it an agent.Puedes describir el trabajo, aportar contexto confiable, validar la salida y mantener a una persona o una detención segura antes de acciones importantes. Eso basta para construir una función útil de IA sin llamarla agente.

Continue when…Continúa cuando…

The system must choose among tools, carry state across steps, recover from partial failure, or work toward a goal with explicit stopping conditions. Continue to Agentic Development; keep Tool Atlas as an optional product reference.El sistema debe elegir entre herramientas, mantener estado entre pasos, recuperarse de fallos parciales o trabajar hacia un objetivo con condiciones explícitas de detención. Continúa a Desarrollo agéntico; usa el Atlas de herramientas como referencia opcional de productos.

Open the durable conceptsAbrir los conceptos duraderos

API call, structured output, and tool contractLlamada API, salida estructurada y contrato de herramienta

An API call is the boundary between your application and a model. Structured output turns an answer into data your program can validate. A tool contract names the action, arguments, permissions, and result shape. Validate every boundary and treat model-proposed tool calls as requests—not authorization.Una llamada API es el límite entre tu aplicación y un modelo. La salida estructurada convierte una respuesta en datos que tu programa puede validar. Un contrato de herramienta nombra la acción, los argumentos, los permisos y la forma del resultado. Valida cada límite y trata las llamadas propuestas por el modelo como solicitudes, no como autorización.

Context and retrievalContexto y recuperación

Separate ingestion from serving: clean, permission, chunk, index, and refresh sources before a request arrives; then retrieve a small, relevant set and show evidence. Debug retrieval failure separately from generation failure. Bigger context is not automatically better context.Separa la ingesta del servicio: limpia, asigna permisos, divide, indexa y actualiza las fuentes antes de que llegue una solicitud; luego recupera un conjunto pequeño y relevante y muestra la evidencia. Depura el fallo de recuperación por separado del fallo de generación. Más contexto no es automáticamente mejor contexto.

Evaluation is a loop, not a launch checkboxLa evaluación es un ciclo, no una casilla de lanzamiento

Keep a small representative set of tasks. Grade answer quality, groundedness, tool choice, trajectory, safety, and latency. Combine automated checks with human review, then rerun the set after changing prompts, models, retrieval, or tools. A passing demo is not a regression suite.Mantén un conjunto pequeño y representativo de tareas. Evalúa calidad de respuesta, fundamentación, elección de herramientas, trayectoria, seguridad y latencia. Combina comprobaciones automáticas con revisión humana y vuelve a ejecutar el conjunto después de cambiar prompts, modelos, recuperación o herramientas. Una demo que funciona no es una suite de regresión.

Workflow vs agentFlujo de trabajo frente a agente

A workflow follows steps you design; an agent chooses the next step inside a boundary. Start with the smallest predictable pattern. Add an agent loop only when tool selection, state, or recovery creates real value—and define a step limit, approval gate, timeout, and safe exit.Un flujo sigue pasos que diseñas; un agente elige el siguiente paso dentro de un límite. Empieza con el patrón predecible más pequeño. Añade un bucle de agente solo cuando la selección de herramientas, el estado o la recuperación aporten valor real—y define un límite de pasos, una aprobación, un tiempo de espera y una salida segura.

Safety and permissionsSeguridad y permisos

Treat retrieved text, tool output, files, and web pages as untrusted input. Use least privilege, audience-bound credentials, consent before sensitive actions, isolated execution, audit trails, and rollback. Design for goal hijacking, tool misuse, privilege abuse, supply-chain risk, and unexpected code execution.Trata el texto recuperado, las salidas de herramientas, los archivos y las páginas web como entradas no confiables. Usa mínimo privilegio, credenciales limitadas a su audiencia, consentimiento antes de acciones sensibles, ejecución aislada, auditoría y reversión. Diseña para secuestro de objetivos, abuso de herramientas, abuso de privilegios, riesgo de cadena de suministro y ejecución inesperada de código.

Human approval belongs at the boundaryLa aprobación humana pertenece al límite

A person should approve external messages, financial or access changes, destructive writes, production releases, and anything with unclear scope. Show the proposed action, inputs, affected resources, and expected rollback before approval.Una persona debe aprobar mensajes externos, cambios financieros o de acceso, escrituras destructivas, lanzamientos a producción y cualquier cosa con alcance incierto. Muestra la acción propuesta, las entradas, los recursos afectados y la reversión esperada antes de aprobar.

Mini proof: write the system boundary cardMini prueba: escribe la tarjeta de límites del sistema

Before wiring another tool, fill this out in plain language. If a field is vague, the system boundary is not ready for more autonomy.Antes de conectar otra herramienta, completa esto en lenguaje sencillo. Si un campo es vago, el límite del sistema no está listo para más autonomía.

System boundary card

User / job:
Trusted inputs and source boundary:
Model responsibility (and what it must not decide):
Tools / actions + least privilege:
Evidence and evaluation set:
Human approval boundary:
Timeout, failure path, rollback, or safe stop:

Proof: run three representative tasks, inspect the evidence,
and record what happens when a tool or source is unavailable.

Tarjeta de límites del sistema

Usuario / trabajo:
Entradas confiables y límite de fuentes:
Responsabilidad del modelo (y lo que no debe decidir):
Herramientas / acciones + mínimo privilegio:
Evidencia y conjunto de evaluación:
Límite de aprobación humana:
Tiempo de espera, fallo, reversión o detención segura:

Prueba: ejecuta tres tareas representativas, inspecciona la evidencia
y registra qué ocurre cuando una herramienta o fuente no está disponible.

Read the primary patternsLee los patrones primarios

OpenAI function callingFunction calling de OpenAI OpenAI structured outputsSalidas estructuradas de OpenAI Anthropic: building effective agentsAnthropic: construir agentes efectivos Google Cloud: RAG architectureGoogle Cloud: arquitectura RAG MCP authorizationAutorización MCP OWASP: agentic threatsOWASP: amenazas agénticas NIST AI Risk Management FrameworkMarco de gestión de riesgos de IA de NIST

Try a workflow: n8n automation · Dify knowledge app · Supabase permission-aware retrieval.Prueba un flujo: automatización n8n · app Dify · recuperación con permisos en Supabase.

Context routing board

Use the room that already has the context

Most weak AI results happen because the model is missing the real situation. Start by choosing the room where the useful information already lives.

→

Microsoft work context

Outlook, Teams, Word, Excel, PowerPoint, OneNote, OneDrive, and SharePoint usually belong in Microsoft 365 Copilot first.

Microsoft 365 Copilot Copilot Chat Copilot agents

→

Google and source packets

Gmail, Drive, Docs, Sheets, Slides, Android, YouTube, and search-connected work fit Gemini. Defined source packets fit Gemini Notebook.

Gemini Google Workspace Gemini Notebook

→

Local files, notes, and personal knowledge

For Apple Notes, screenshots, local files, Obsidian, Markdown, or personal folders, choose the app that can receive that context with the least friction.

Claude Desktop ChatGPT macOS/Windows/mobile Obsidian/Markdown

→

Repos, branches, and project files

Code work belongs in an agent lane when files, tests, diffs, branches, or terminal commands matter.

Codex Agent Claude Code Cursor Grok Build

Useful habit

Before pasting work context into another AI, ask: does this tool already have approved access to the same context, or am I about to move information into a worse room?

Agent location

Know where the AI is working

A chat response, a connected app, a local agent, and a remote cloud agent have different powers. The more an AI can touch, the more deliberately you should review its work.

Lane 1

Chat answer

The AI answers from what you type, attach, or connect. Good for reasoning, rewriting, learning, and comparison.

Examples: ChatGPT, Claude, Gemini, Grok.

Lane 2

Connected app

The AI can search or reference approved services, and sometimes take narrow actions through apps, connectors, or MCP.

Examples: ChatGPT apps, Claude connectors, MCP servers.

Lane 3

Local agent

The AI works in your local project or desktop flow, can inspect files, suggest edits, and may ask before running commands.

Examples: Codex app, Claude Code local, Cursor foreground.

Lane 4

Remote/work agent

The AI works in a cloud, GitHub, or work-app environment. It can create branches, files, docs, decks, or agent outputs for review.

Examples: Copilot agents, Codex remote, Claude Code web, Cursor background, Antigravity, Grok Build.

Try a local workflow: LM Studio for a desktop start, or Ollama for API integration.Prueba un flujo local: LM Studio para empezar en escritorio o Ollama para integrar APIs.

Microsoft

Microsoft Copilot for web, work chat, and Microsoft 365 context

Microsoft now has several Copilot lanes: free/personal Copilot for web and general help, Copilot Chat Basic for secure work chat, and Microsoft 365 Copilot for full work-data grounding inside Microsoft apps.

Open free Copilot Open Copilot Copilot Chat overview Copilot in apps

Free and personal Copilot

Microsoft Copilot

General Copilot for web-grounded questions, writing help, image creation, Edge page help, and everyday learning. Signing in adds history, longer conversations, voice, image creation, and other features.

Conversation modes

Quick response Think Deeper Study and learn Smart Search

Best first use

Use it for everyday questions, web summaries, learning, brainstorming, writing, and trying the model-depth controls without company context.

When not enough

Move to Microsoft 365 Copilot when the answer must use work files, meetings, emails, chats, or app-specific editing inside Word, Excel, PowerPoint, Outlook, Teams, or OneNote.

Microsoft work chat

Copilot Chat

Secure work chat included for eligible Microsoft 365 business users. It is web-grounded by default, includes standard access to file upload, image generation, and model choices, and can use open or uploaded context.

What Basic can use

Web grounding Uploaded files/images Open Outlook/Teams context Copilot Pages Pay-as-you-go agents

Model selector

Use Auto for normal work. Pick Quick response for speed. Pick Think deeper when the task needs more reasoning. Some tenants also offer Claude and current GPT-family choices in the selector.

Best first use

Ask it to research, draft, summarize an uploaded file, create a page, or work beside an open Outlook or Teams item before you reach for a broader chatbot.

Quick label check

Look at your Copilot label: Copilot Chat (Basic) means secure chat without Copilot in Word/Excel/PowerPoint/OneNote; M365 Copilot (Basic) means standard access in those apps; M365 Copilot (Premium) means the full add-on experience with priority access.

Microsoft 365 apps

Microsoft 365 Copilot

The Microsoft 365 Copilot add-on adds priority access, full work-data grounding, deeper app integration, and advanced skills across Word, Excel, PowerPoint, Outlook, Teams, OneNote, Loop, OneDrive, and SharePoint.

Where it fits

Word, PowerPoint, and OneNote for drafting, rewriting, and document Q&A.
Excel for formulas, tables, charts, Python-assisted analysis, and deeper reasoning where enabled.
Teams and Outlook for meetings, threads, follow-ups, calendar, and email context.

Why people pick it

It can answer from meetings, emails, chats, files, and people your account can access instead of making you upload or paste everything by hand.

Use another model when

You need outside critique, broader brainstorming, non-Microsoft context, or a second read with zero company background.

Work app agents

Agents, Researcher, Analyst, and Office creation

Microsoft’s agent path ranges from Basic pay-as-you-go agents to Premium access for advanced reasoning agents like Researcher and Analyst, plus custom agents grounded in company data where enabled.

What to know

Basic users may see limited agents and pay-as-you-go options. Premium users get broader access to advanced agents, Researcher, Analyst, prebuilt Microsoft agents, and custom agents. Eligible tenants can schedule recurring prompts for some agents.

Good use cases

Use Researcher for sourced briefs, Analyst for data-heavy work, app agents for Word/Excel/PowerPoint creation, and custom agents for repeatable internal workflows.

Review habit

Open the source files it used and the generated file it created. Check structure, formulas, citations, inherited sensitivity labels, and any required AI-media watermark before routing it onward; controls vary by platform and admin policy.

OpenAI

Chat, Work, and Codex serve different jobs—and have different model choices.

Use Chat for conversation, Work for longer connected deliverables, and Codex for software work. In Work and Codex, consider GPT-6 Luna for focused tasks and GPT-6.1 Sol for complex work. These are not Chat defaults; availability depends on plan, client, and workspace settings.

ChatGPT Codex Agent

OpenAI app

ChatGPT app

General AI workspace for Chat, Work, files, images, voice, research, projects, memory, apps, and custom GPTs.

Open ChatGPT Download ChatGPT apps Apps

What it does

Search Deep Research Files Data Analysis Voice Writing/code blocks Projects Memory Apps ChatGPT Work GPT-6.1 Sol / GPT-6 Luna

Set it up

Open ChatGPT on web for the full workspace.
Install macOS or Windows when local computer context matters.
Create Projects for recurring work and shared context.
Where available, choose Luna for speed, GPT-6.1 Terra for writing, visual understanding, and supported computer use, and Sol when the task needs more reasoning.

Use it for

Conversation, research, writing, files, data, image/audio context, and longer connected deliverables through Work.

Web app

Best full workspace. Use it for long sessions, uploaded files, Projects, research, data, images, writing/code blocks, and recurring work context.

Mobile app

Best capture surface. Use voice, photos, screenshots, quick questions, and steering Codex work when that workflow is enabled.

ChatGPT desktop

One app, three lanes. On macOS and Windows, ChatGPT desktop brings Chat, Work, and Codex together beside files, windows, screenshots, and focused work.

Apps

Best for connected context. Apps can search, reference, sync, run deep research, or take approved actions through connected tools where available.

ChatGPT vs Microsoft Copilot

Use ChatGPT when you need flexible reasoning, drafting, critique, files, research, images, or a second opinion. Use Microsoft 365 Copilot first when the answer depends on Outlook, Teams, SharePoint, OneDrive, Office files, or company permissions.

Choose the lane before the model

Use Chat for an answer, Work for a connected multi-step deliverable, and Codex when the outcome changes software. Where the tiers appear, choose Luna for speed, GPT-6.1 Terra for writing and visual work, or Sol for complex reasoning, and raise effort only when the stakes justify it.

OpenAI developer agent

Codex agent in ChatGPT desktop + CLI

Software agent for real project files: inspect repos, edit code, run commands, create diffs, test, work remotely, and now operate inside the unified ChatGPT desktop app.

Codex Agent Plan access

Set it up

npm i -g @openai/codex
codex

brew install --cask codex

What it offers

Desktop app CLI IDE extension Cloud tasks GitHub workflows Mobile steering Model + effort controls Multi-repo projects Task lists and goals Agents dashboard Queued messages Shared thread snapshots

Use it for

Code changes, debugging, tests, PR work, repo review, and turning a software task into tracked file edits.

macOS desktop

Most complete desktop host. Use ChatGPT desktop for multi-agent threads, worktrees, Appshots, locked-computer remote work, and ChatGPT mobile steering.

Windows desktop

Desktop command center. Use it for parallel agents, inline diff editing, pull-request review, and moving between ChatGPT desktop, CLI, and IDE across multiple repositories.

Remote + mobile

Use your phone to keep work moving. From ChatGPT mobile you can review outputs, approve commands, change models, start new work, and follow screenshots, terminal output, diffs, tests, and approvals from a connected Mac or remote environment.

Remote environments

Use for team/devbox workflows. Codex can connect through Remote SSH, run inside managed remote machines, use cloud environments, and keep project context synced across authorized ChatGPT devices.

Supervise longer Codex work and share carefully

In current Codex CLI builds, use `codex agents` to manage tasks, `codex queue` to send follow-ups, `/cd` or `/cwd` to confirm the working directory, and `codex doctor` when setup fails. On macOS, shared thread links are read-only snapshots; review every link for sensitive content and revoke it from ChatGPT data controls when finished.

Anthropic

Claude for careful work. Sonnet 5.5 for efficient execution. Opus 5.5 for proactive work. Fable 5.1 and Mythos 5.1 for the frontier.

Anthropic’s stack is strongest when the job needs long context, thoughtful writing, reusable artifacts, browser or desktop work, or repo-aware coding. Sonnet 5.5 is the efficient serious-work lane; Opus 5.5 is the thoughtful proactive lane; Fable 5.1 and Mythos 5.1 are the newest advanced coding and knowledge-work options. Access, usage, and effort controls vary by plan and surface.

Claude Cowork Claude Code

Anthropic app

Claude app

Careful writing, long-context review, project knowledge, artifacts, connectors, browser work, mobile actions, and current Sonnet 5.5 / Opus 5.5 / Fable 5.1 / Mythos 5.1 model lanes.

Open Claude Install desktop iOS actions Android actions

What it offers

Projects Project knowledge Artifacts Shared artifacts MCP in artifacts Web connectors Mobile app actions Desktop extensions Editable memory topics Microsoft 365 write tools Haiku 5.5 Sonnet 5.5 Opus 5.5 Fable 5.1 Mythos 5.1

Set it up

Use Projects for repeated context.
Use Artifacts for outputs you will edit or reuse.
Install mobile when the next step is a message, email, calendar item, map, reminder, alarm, timer, or health summary.
Install desktop when local files, desktop extensions, Cowork, or Mac quick entry matter.

Use it for

Nuanced writing, long documents, tone review, research synthesis, editable artifacts, and agentic knowledge work.

iOS app

Action layer for Apple mobile. Claude can draft messages and email, create calendar events, show map locations, manage reminders, and analyze Apple Health data when eligible.

Android app

Action layer for Android. Claude can draft messages and email, create calendar events, set alarms and timers, use location/maps, and analyze Health Connect data when eligible.

macOS Desktop

Best for staying in flow. Quick entry can open Claude from anywhere, capture screenshots, attach app windows, and use voice dictation. Desktop extensions add local files, apps, calendars, email, and messaging.

Windows Desktop

Best for local desktop context. Use Claude Desktop for extensions, local files, Cowork, and connected workflows. Mac quick entry is not available on Windows.

Choose the smallest Claude lane that can prove the work

Start with Sonnet 5.5 for efficient everyday analysis, writing, tools, and coding. Move to Opus 5.5 for more proactive handling, or to Fable 5.1 or Mythos 5.1 when a difficult coding, research, or knowledge task earns the frontier lane. In any lane, state the goal, what may be touched, and what you will review.

Claude Duet: a proactive agent in research previewClaude Duet: agente proactivo en preview de investigación

Duet is a separate app on web, iOS and Android. Its waitlist starts with U.S. Pro and Max subscribers and shares their Claude usage limits. It can prepare a morning brief and continue delegated work on a cloud computer while you are away.Duet es una app separada para web, iOS y Android. Su lista de espera empieza con suscriptores Pro y Max de EE. UU. y comparte sus límites de Claude. Puede preparar un resumen matinal y seguir trabajo delegado en un ordenador en la nube mientras estás fuera.

First workflow: connect one work source, ask for a brief with links, and check urgency and facts. Before allowing external actions, inspect the destination, recipients and content. Disconnecting an app stops new access; reset Duet memory separately to remove what it learned.Primer flujo: conecta una fuente laboral, pide un resumen con enlaces y comprueba urgencia y hechos. Antes de permitir acciones externas, inspecciona destino, destinatarios y contenido. Desconectar una app detiene acceso nuevo; restablece la memoria de Duet por separado para eliminar lo aprendido.

Claude Duet ↗ · Privacy + controls ↗Privacidad y controles ↗

Claude knowledge-work agent

Claude Cowork

Knowledge-work agent for files, apps, browser work, deliverables, plugins, and scheduled tasks. On web and mobile, remote Cowork runs in Anthropic’s cloud in beta; availability varies by plan, and its sessions and files are saved to your Claude account.

Cowork setup Computer use

What it offers

Desktop, web, and mobile Remote sessions Local files Built-in separate browser Claude in Chrome Editable memory topics Plugins Parallel subtasks Scheduled tasks

Set it up

Start local work in Claude Desktop when it needs your computer. Where your plan includes the beta, use remote Cowork on web or mobile when work should continue without the laptop online; review what will be stored with your Claude account first.

Use it for

Folder organization, reports, slide prep, browser tasks, multi-step knowledge work, and scheduled deliverables you want to continue across devices.

Enterprise plugin safety is available in beta

Claude Enterprise admins can enable security scanning for third-party skills and plugins when they are uploaded or edited. Treat scanning as another check—not a substitute for reviewing permissions, source, and the actions a plugin can take.

Choose which browser boundary the job needs

Cowork’s built-in browser is separate from your everyday browser and fits delegated web tasks. Claude in Chrome works in the page and signed-in session you already have open. Both can encounter prompt injection, so use trusted sites, narrow accounts, and review consequential actions.

Anthropic code agent

Claude Code

Code agent for terminal, IDE, Git, GitHub, web/remote tasks, MCP, AWS, and Google Cloud setups.

Setup docs Web tasks Remote control GitHub Actions

Set it up

npm install -g @anthropic-ai/claude-code
cd your-project
claude

curl -fsSL claude.ai/install.sh | bash
claude doctor

What it offers

Terminal agent IDE integrations Git/GitHub Web remote tasks Remote Control GitHub Actions Remote MCP AWS/GCP enterprise

Use it for

Repo inspection, implementation, code review, terminal workflows, PRs, and remote coding tasks.

Terminal / IDE

Use when you need tight control. Claude Code runs in your local project, can inspect files, run commands, use local MCP servers, and keep the workflow visible in your terminal or IDE.

Remote Control

Use when you need to steer local work from another device. The session keeps running on your machine while claude.ai/code or the Claude mobile app acts as the remote window.

Claude Code on web

Use for async GitHub work. Pick a repo, describe the task, let Claude work in an isolated remote environment, then review the branch or pull request when it finishes.

Automation routes

Use for team workflows. GitHub Actions can respond to `@claude` in issues or PRs, while remote MCP connectors let Claude reach cloud tools from Claude, Cowork, Desktop, and mobile.

Google

Gemini for Google context. Gemini Notebook for source packets. Antigravity for agent coding.

Google’s stack is most useful when the work already lives in Gmail, Drive, Docs, Sheets, Android, source documents, or developer tooling.

Gemini Gemini Notebook CLI tools

Google app

Gemini app + Workspace

Google AI assistant for Search-connected work, Workspace context, Android/mobile, Deep Research, Canvas, Gems, Gemini 4 Argon, Deep Think, and Omni-style video/media tasks.

Open Gemini Deep Research

What it offers

Gmail/Drive context Workspace Gemini 4 Argon Deep Research Canvas Gems Gemini Omni Android/mobile Computer use for builders

Set it up

Open Gemini with the Google account that owns the context.
Use Workspace features when Gmail, Drive, Docs, Sheets, or Slides matter.
Create Gems for repeated role/task setups.

Use it for

Google-native work, research with Search context, Android help, Drive/doc synthesis, Workspace productivity, and builder/enterprise agent work with Gemini 3.5 Flash computer use.

Google research + learning workspace

Gemini Notebook

Gemini Notebook is Google’s standalone, source-grounded research and thinking workspace. It syncs with notebooks in Gemini while keeping Notebook-only Studio artifacts in the dedicated app.

Open Gemini Notebook Notebook actions + limitsAcciones y límites de notebooks Rename + updates

What it offers

Source-grounded citations Fast + Deep Research Audio + video overviews Mind maps Reports + data tables Flashcards + quizzes Slide decks + infographics Gemini app sync Cloud computer + code*

Set it up

Create a notebook, add or research the sources you trust, select the subset for each question, follow inline citations, then generate the Studio artifact that fits the job. Cloud-computer analysis and flexible compute-specific limits depend on account, plan, and rollout.

Use it for

Deep learning, research synthesis, document onboarding, cited briefings, study and training materials, visual or audio explanation, and—on eligible plans—code-backed analysis grounded in the notebook.

Gemini API CLI: call models from your existing agentGemini API CLI: llama modelos desde tu agente actual

The new gemini-api beta is a stateless shell client for text, media, file analysis and managed agents through the Interactions API. It is a separate tool from Antigravity CLI and the older interactive Gemini CLI. Use it when an existing agent or CI job needs one API operation.La nueva beta gemini-api es un cliente de shell sin estado para texto, medios, análisis de archivos y agentes gestionados mediante Interactions API. Es una herramienta distinta de Antigravity CLI y de la antigua CLI interactiva Gemini. Úsala cuando un agente existente o CI necesite una operación API.

Pin a release in automation, inspect --schema and --dry-run output, keep credentials in the environment or secret store, and check remote request-storage settings. Stateless local commands can still create server-side history.Fija una versión al automatizar, inspecciona --schema y --dry-run, guarda credenciales en el entorno o gestor de secretos y revisa el almacenamiento remoto. Comandos locales sin estado aún pueden crear historial en el servidor.

Gemini API CLI ↗

Google developer agents

Antigravity CLI (consumer Gemini CLI transition complete)

Google’s current developer-agent path for terminal work, skills, hooks, subagents, plugins, and background multi-agent work. Check Antigravity’s current model picker and plan; Gemini CLI is now a migration reference, not the starting point for new consumer setups.

Antigravity CLI Install docs

Gemini CLI migration

The consumer transition date passed on June 18, 2026. If you still have Gemini CLI, bring its useful skills, hooks, subagents, and extensions into Antigravity. For a new setup, start directly with Antigravity CLI.

Migration guide

Antigravity CLI

curl -fsSL https://antigravity.google/cli/install.sh | bash
agy

irm https://antigravity.google/cli/install.ps1 | iex
agy

agy
agy plugin import gemini

Use it for

Terminal coding, background agent tasks, plugin workflows, and Google’s current agent-first developer tooling. Gemini 3.7 Flash is positioned for coding, knowledge work, web development, and multi-step tool use.

Choose an Antigravity model for the task

Use the current Antigravity picker for coding and tool-heavy work. Gemini 4 Argon expands Google’s complex-task options, and the Gemini macOS app now includes Antigravity’s coding interface. Confirm access before standardizing a team workflow.

Custom agents

Use specialized project roles. Antigravity 2.0 and CLI load Markdown plus YAML agent files from `.agents/agents/`. Commit only instructions safe for teammates, and scope tools and command permissions to the role.

IDE extensions

Bring Antigravity into an editor. Extensions are available for VS Code, Visual Studio, JetBrains 2026.2.1+, and Zed. Visual Studio and some Enterprise integrations remain in Preview.

Browser Remote Control

Steer sessions on another machine. Remote Control can reach that machine’s files, tools, credentials, and environment variables. Protect the account and device, narrow permissions, and end access when the session is done.

Cursor

Cursor IDE + Agent for learning by building

Remote Control follows local agents from iOS without moving them to the cloud: keep the desktop awake and online. Projects beta instead coordinates cloud work. Teams and Enterprise can use Rollouts and Security Review; Rollouts can suggest a revert PR, but does not merge or roll back on its own.Remote Control sigue agentes locales desde iOS sin moverlos a la nube: mantén el escritorio despierto y conectado. Projects beta coordina trabajo en la nube. Teams y Enterprise pueden usar Rollouts y Security Review; Rollouts puede proponer un PR de reversión, pero no lo fusiona ni revierte por sí solo. Official guide ↗Guía oficial ↗

Cursor is the AI-native editor path: start by asking questions about code, move into visible file edits, then use Agent, rules, models, and background agents as your confidence grows.

Download Cursor Cursor CLI Background agents Bugbot Cursor updates

Why people pick it

It keeps the AI beside the code. You can see files, diffs, terminal output, and model choices in one workspace, which makes it easier to learn which model works for a task.

IDE editing Ask mode Agent mode Model picker Cursor Router Rules Background agents iPhone + iPad review Google Workspace plugins GitHub app Bugbot Cursor CLI Cloud subscriptions Origin early beta Start without a repo Self-hosted machines

Your progression path

1. Learn

Use Ask

Ask how files connect before changing anything.

2. Edit

Use IDE help

Make small visible changes and review diffs.

3. Agent

Use Agent mode

Let it explore, edit multiple files, and run commands.

4. Route

Route with intent

Use Auto with Cost, Balance, or Intelligence, then override the model only when you have a reason.

Set up rules

Use Cursor Settings → Rules or files in `.cursor/rules`. For simple projects, `AGENTS.md` can act as a plain markdown instruction file.

Set up remote/background work

Cloud agents can use Cursor-hosted sandboxes or self-hosted machines in your network. Choose the execution boundary first, then keep repository, secret, and review permissions narrow.

Connect Google Workspace carefully

Cursor plugins can read and act in Gmail, Drive, and Calendar. Grant only the permissions the task needs, review drafts or changes before approval, and disconnect access when the workflow no longer needs it.

Set up the CLI

curl https://cursor.com/install -fsS | bash
cursor-agent

GitHub app

Why it matters: background agents and Bugbot need repository access so Cursor can clone repos, open branches, and push reviewable changes.

Bugbot / PR review

Use for review: Bugbot can review pull requests and follow repo-specific guidance such as `.cursor/BUGBOT.md` where configured.

Review habit

Do not skip the handoff: read the branch diff, run tests locally when possible, and check whether the background environment had the same setup you use.

iPhone + iPad

Use mobile as a full review surface: launch and steer cloud agents, read complete PRs, checks, comments, and approvals, then create, review, or merge from the inbox. Availability depends on plan.

Cloud subscriptions are for always-on agent follow-through

Cloud agents can monitor pull requests, watch a Slack thread, or run scheduled tasks, then wake when the subscribed event changes. This is cloud-agent-only for now; limit repository and Slack access, define the goal, and keep review gates around CI fixes and replies.

Origin code hosting is an early beta

Paid plans are beginning to receive Cursor-hosted repos, pull requests, code browsing, and GitHub sync. A cloud agent can now start from a prompt without a connected repo, preview the app in the browser, then save it to Origin or publish through Vercel. Decide which host is the source of truth before work branches.

Build a reusable agent with Cursor Agent SDKCrea un agente reutilizable con Cursor Agent SDK

The Agent SDK packages instructions, typed tools, skills, MCP connections and event channels as code. It is generally available and can deploy to Cursor or your own infrastructure. Cloud agent canvases can show interactive results; the Environments API remains beta.Agent SDK reúne instrucciones, herramientas tipadas, skills, conexiones MCP y canales de eventos como código. Está disponible de forma general y se despliega en Cursor o tu infraestructura. Los canvases de agentes cloud muestran resultados interactivos; Environments API sigue en beta.

First lab: build an issue-triage bot with read access and a proposed label as its output. Run evals that assert which tools were called, replay a saved event, and test a duplicate event and a forbidden action before enabling writes.Primer lab: crea un bot de clasificación de issues con acceso de lectura y una etiqueta propuesta como salida. Ejecuta evaluaciones de las herramientas llamadas, reproduce un evento guardado y prueba un duplicado y una acción prohibida antes de habilitar escrituras.

Agent SDK ↗ · Cloud environments ↗Entornos en la nube ↗

xAI

Grok for current context. Grok Build for code. Grok Bot for demonstrated routines.

xAI’s stack spans the Grok assistant and APIs, agentic development in Grok Build, and browser routines taught by demonstration in the Grok Bot beta.

Grok Grok Build Grok Bot

xAI app

Grok app + xAI API

xAI assistant across Grok.com, mobile apps, X, web search, files, connectors, APIs, voice, image generation, and Grok 4.7 where available.

Open Grok xAI docs

What it offers

Grok.com Mobile apps Grok in X X Search Web Search Voice API Imagine API xAI API Grok 4.7 File uploads Connectors

Set it up

Use Grok.com for the standalone app.
Use X when the context is social or from X.
Use xAI docs when building with the API.

Use it for

Web-aware questions, X/social pulse, files, xAI API builds, voice/image experiments, and current-context review.

xAI code agent

Grok Build

Grok 4.7 is available in Grok Build for longer agent, coding, knowledge, and visual work. Use the terminal agent interactively or headlessly, or create and publish apps from Grok on web and mobile. Availability and included usage vary by plan and surface.

Interactive TUI Headless scripting ACP Skills/plugins/hooks MCP discovery Agent Dashboard Plugin marketplace Web + mobile app building Publishing + GitHub export

Grok Build docs Headless scripting

Install

Start with the official installer, then launch from a project folder.

curl -fsSL https://x.ai/cli/install.sh | bash

1. Launch

Open a repo session

Use the TUI when you want to watch and steer the agent in the project.

cd your-project
grok

2. Inspect

See what it detected

Review config, instructions, skills, plugins, hooks, and MCP servers before deeper work.

grok inspect

3. Automate

Run headlessly

Use headless mode for scripts, checks, bots, or quick repo explanations.

grok -p "Explain this codebase"

xAI persistent browser agent · beta

Grok Bot

Team Bots separate shared team memory from private notes about each person. Owners and managers control shared skills and setup. A secret available to a Team Bot can be used in any teammate’s conversation: treat it as team-wide access, not a private credential.Team Bots separa memoria del equipo de notas privadas de cada persona. Propietarios y gestores controlan skills y configuración compartidas. Un secreto del Team Bot puede usarse en conversaciones de cualquier compañero: trátalo como acceso del equipo, no credencial privada. Official guide ↗Guía oficial ↗

Show Grok Bot a browser workflow by performing it while the bot follows along. It can save the steps as a routine and run that routine later on demand or on a schedule. Its X connector can now search posts, read timelines, check mentions, and manage bookmarks.

Demonstrated workflows Reusable routines Scheduled runs Persistent bot memory Cloud computer X connector

Grok Bot docs Bots and memory Compare the agent layers

Important boundary

Bot identity is not computer isolation.

Named bots keep their own conversation context and memory, but the bots share one cloud computer. Files, browser sign-ins, and computer state can therefore cross bot roles. Use narrow accounts and review every routine before scheduling it.

Before AI CLIs

CLI setup and prerequisites

Most AI coding tools need the same basics: a modern terminal, Node.js with npm, Git, and a shell that handles commands predictably.

Node.js LTS PowerShell 7

Mac

macOS baseline

Use Terminal or iTerm, install Homebrew, then install the common command-line dependencies.

Install Homebrew

/bin/bash -c "$(curl -fsSL https://raw.githubusercontent.com/Homebrew/install/HEAD/install.sh)"

Install Node.js and Git

brew install node git

Verify

node -v
npm -v
git --version

Win

Windows baseline

Use Windows Terminal with PowerShell 7. It installs side-by-side with Windows PowerShell 5.1 and launches with `pwsh`.

Install PowerShell 7

winget install --id Microsoft.PowerShell --source winget

Install Node.js LTS and Git

winget install --id OpenJS.NodeJS.LTS --exact
winget install --id Git.Git --exact --source winget

Open a new PowerShell 7 tab and verify

pwsh --version
node -v
npm -v
git --version

WT

Windows Terminal setup

Install Windows Terminal from the Microsoft Store if you want the easy path. Then make PowerShell 7 your default profile and create separate profiles for the AI tools you use most.

Install Windows Terminal

Easiest path

Open the Microsoft Store, search “Windows Terminal,” and click Install. This is the simplest route for most Windows users.

Open in Microsoft Store

Command-line path

winget install --id Microsoft.WindowsTerminal --source winget

Profile pattern

Use one profile per working lane: Codex Agent, Claude Code, Gemini or Antigravity, Grok Build, Cursor Agent. Set the profile name, command line, starting folder, color, and icon so you always know where you are.

A clean profile starter

Open Windows Terminal settings, add a profile, and use this shape as the mental model. The important fields are `name`, `commandline`, and `startingDirectory`.

{
  "name": "Codex Agent",
  "commandline": "pwsh.exe",
  "startingDirectory": "%USERPROFILE%/Documents/Projects",
  "tabTitle": "Codex Agent"
}

Start folder

Point each agent at the folder where you actually keep repos.

Shell

Use `pwsh.exe` for PowerShell 7 so modern commands and profiles behave consistently.

Identity

Give each AI lane a clear name and color so you do not paste commands into the wrong session.

Why Node/npm?

Many AI CLIs install through npm, including several tools on this page.

Why Git?

Coding agents need repo context, branch history, diffs, and a clean way to review file changes.

Why restart the terminal?

Installers often update PATH. A new terminal session makes `node`, `npm`, `git`, and `pwsh` visible.

If setup breaks Common CLI troubleshooting moves

Wrong PowerShell

Run `pwsh --version`. If that fails, you are probably in Windows PowerShell 5.1 or PATH has not refreshed.

PATH did not refresh

Close every terminal window, reopen Windows Terminal or Terminal, then verify `node`, `npm`, `git`, and the AI CLI again.

winget missing

Install from Microsoft Store when possible. Store installs are often easier for Windows Terminal and App Installer updates.

npm global install fails

Check Node LTS, permissions, and your shell. Avoid mixing admin and non-admin installs unless you know why.

Git is not identified

Set your Git name/email and confirm your repo branch before asking an agent to commit or push.

Native vs WSL

Pick one lane per project. Mixing Windows paths, WSL paths, and different Node installs makes agents harder to reason about.

Connected context and actionContexto y acción conectados

MCP: the connection layerMCP: la capa de conexión

Some MCP integrations now push events rather than waiting for a question. Cloudflare can deliver signed Workers Issues events to compatible client webhooks. Verify client support, signature checks, subscription scope, duplicate handling, and who may turn an event into an action.Algunas integraciones MCP ahora envían eventos sin esperar una pregunta. Cloudflare puede entregar eventos firmados de Workers Issues a webhooks compatibles. Verifica soporte del cliente, firmas, alcance, duplicados y quién puede convertir el evento en una acción. Official guide ↗Guía oficial ↗

Model Context Protocol is an open standard that lets an AI host connect to outside data, tools, and workflows through focused MCP servers. It standardizes the interface; it does not make every server, permission, or action safe.Model Context Protocol es un estándar abierto que permite a un host de IA conectarse con datos, herramientas y flujos externos mediante servidores MCP enfocados. Estandariza la interfaz; no vuelve seguros todos los servidores, permisos ni acciones.

MCP 2026-07-28 MCP changelog

MCP in one pictureMCP en una imagen

1 · Host

The AI applicationLa aplicación de IA

Claude, ChatGPT, Codex, Cursor, an IDE, or another agent host. It owns consent, policy, and the final context sent to the model.Claude, ChatGPT, Codex, Cursor, un IDE u otro host de agentes. Controla consentimiento, política y el contexto final enviado al modelo.

2 · Client

One server connectionUna conexión a servidor

The host manages a separate client for each server, carries version and capability information, and keeps servers from seeing one another directly.El host administra un cliente separado por servidor, lleva versión y capacidades y evita que los servidores se vean directamente.

3 · MCP server

Focused capabilityCapacidad enfocada

A local process or remote service that exposes prompts, resources, tools, or supported extensions for one clear job.Un proceso local o servicio remoto que expone prompts, recursos, herramientas o extensiones compatibles para un trabajo claro.

4 · System

The real data or actionEl dato o acción real

A repository, database, browser, design file, calendar, notes vault, business app, API, or local filesystem.Un repositorio, base de datos, navegador, diseño, calendario, bóveda de notas, app empresarial, API o sistema local de archivos.

What a server can offerQué puede ofrecer un servidor

The three core primitives differ by who chooses when they enter the workflow.Las tres primitivas principales se distinguen por quién decide cuándo entran al flujo.

P

Prompts · user-controlledPrompts · controlados por la persona

Templates or guided workflows the user explicitly selects.Plantillas o flujos guiados que la persona elige explícitamente.

R

Resources · application-controlledRecursos · controlados por la aplicación

Structured context such as files, schemas, repository history, logs, or knowledge records.Contexto estructurado como archivos, esquemas, historial del repositorio, logs o registros de conocimiento.

T

Tools · model-invokedHerramientas · invocadas por el modelo

Functions that retrieve information or take actions. The host still decides which tools exist, which require approval, and what the user may authorize.Funciones que recuperan información o realizan acciones. El host aún decide cuáles existen, cuáles requieren aprobación y qué puede autorizar la persona.

A safe first connectionUna primera conexión segura

1. Name the job1. Nombra el trabajo

Connect a server for one real need, not because its catalog looks impressive.Conecta un servidor por una necesidad real, no porque su catálogo se vea impresionante.

2. Verify provenance2. Verifica la procedencia

Confirm who maintains it, where the code or service lives, how it updates, and what it can execute.Confirma quién lo mantiene, dónde vive el código o servicio, cómo se actualiza y qué puede ejecutar.

3. Start read-only3. Empieza en solo lectura

Expose one account, folder, repository, or project and the smallest useful set of resources or tools.Expón una cuenta, carpeta, repositorio o proyecto y el conjunto mínimo útil de recursos o herramientas.

4. Run a harmless proof4. Ejecuta una prueba inocua

List available tools, read a known item, confirm the returned source, and inspect logs or activity before enabling writes.Lista herramientas, lee un elemento conocido, confirma la fuente devuelta e inspecciona logs o actividad antes de habilitar escritura.

5. Add actions one boundary at a time5. Agrega acciones un límite a la vez

Enable only the next needed mutation, keep destructive or external communication behind explicit approval, and know how to disconnect or revoke access.Habilita solo la siguiente mutación necesaria, mantén acciones destructivas o comunicaciones externas tras aprobación explícita y sabe cómo desconectar o revocar el acceso.

stdio

Local process boundaryLímite de proceso local

The host launches a subprocess and exchanges newline-delimited messages over standard input and output. This is useful for local files and developer tools, but the process may inherit environment variables, filesystem access, and the privileges of the user who started it.El host inicia un subproceso e intercambia mensajes delimitados por líneas mediante entrada y salida estándar. Sirve para archivos locales y herramientas de desarrollo, pero el proceso puede heredar variables de entorno, acceso a archivos y privilegios de quien lo inició.

Streamable HTTP

Remote service boundaryLímite de servicio remoto

Each request goes to one HTTP endpoint and returns JSON or a request-scoped stream. This fits shared cloud systems; authentication, network policy, server identity, least-privilege scopes, and token storage become part of the trust boundary.Cada solicitud va a un endpoint HTTP y devuelve JSON o un stream limitado a esa solicitud. Es adecuado para sistemas cloud compartidos; autenticación, política de red, identidad del servidor, scopes mínimos y almacenamiento de tokens forman parte del límite de confianza.

Current protocol depthProfundidad del protocolo actualWhat changed in MCP 2026-07-28?¿Qué cambió en MCP 2026-07-28?

Stateless requestsSolicitudes sin estado

Every request carries its own version, client identity, and capabilities. The old required initialization handshake and protocol session ID are gone for the current revision.Cada solicitud lleva su versión, identidad del cliente y capacidades. El antiguo handshake obligatorio y el ID de sesión del protocolo desaparecen en la revisión actual.

Discovery when neededDescubrimiento cuando se necesita

A client may call server discovery to learn supported versions and capabilities, but it can also send a self-contained request and handle compatibility directly.Un cliente puede consultar el servidor para conocer versiones y capacidades, o enviar una solicitud autocontenida y manejar compatibilidad directamente.

Input and long workEntrada y trabajo largo

Multi Round-Trip Requests let a server ask for confirmation or missing input inside the request flow. The Tasks extension can represent deferred or long-running work where supported.Multi Round-Trip Requests permite pedir confirmación o datos faltantes dentro del flujo. La extensión Tasks puede representar trabajo diferido o largo donde exista soporte.

Extensions and compatibilityExtensiones y compatibilidad

MCP Apps can add sandboxed interactive interfaces. Older hosts may still use earlier sessions, transports, or capabilities, so follow the host and server documentation instead of copying a random configuration.MCP Apps puede agregar interfaces interactivas en sandbox. Hosts antiguos pueden usar sesiones, transportes o capacidades previas; sigue la documentación del host y servidor en vez de copiar una configuración al azar.

You do not need to memorize wire-level changes to use MCP. You do need to know which host, server, version, transport, and permissions are actually in the path.No necesitas memorizar cambios del protocolo para usar MCP. Sí necesitas saber qué host, servidor, versión, transporte y permisos están realmente en la ruta.

Trust boundaryLímite de confianzaMCP security and readiness checklistChecklist de seguridad y preparación de MCP

Server sourceOrigen del servidor

Who publishes, signs, hosts, and updates it? A registry entry or polished tool description is not a security review.¿Quién lo publica, firma, aloja y actualiza? Una entrada de registro o descripción pulida no equivale a una revisión de seguridad.

Data boundaryLímite de datos

Which exact folders, accounts, repositories, projects, messages, or records can it read?¿Qué carpetas, cuentas, repositorios, proyectos, mensajes o registros exactos puede leer?

Action boundaryLímite de acción

Which tools create, edit, merge, delete, deploy, purchase, or communicate externally? Allowlist only what the task needs.¿Qué herramientas crean, editan, fusionan, eliminan, despliegan, compran o se comunican afuera? Autoriza solo lo necesario.

CredentialsCredenciales

Prefer the provider's OAuth flow for remote servers and narrowly scoped host-managed secrets for local ones. Never place tokens in prompts, repositories, issues, or logs.Prefiere OAuth del proveedor para servidores remotos y secretos limitados administrados por el host para los locales. Nunca pongas tokens en prompts, repositorios, issues ni logs.

Untrusted contentContenido no confiable

Connected pages, issues, comments, files, and even tool descriptions can contain misleading instructions. Treat returned content as data, not authority.Páginas, issues, comentarios, archivos e incluso descripciones de herramientas pueden contener instrucciones engañosas. Trata el contenido devuelto como datos, no como autoridad.

Proof and exitPrueba y salida

Know where calls are logged, how results are verified, how access is revoked, and what must be undone if an action is wrong.Sabe dónde se registran llamadas, cómo se verifican resultados, cómo se revoca acceso y qué debe revertirse si una acción sale mal.

Use MCP for a boundaryUsa MCP para un límite

Choose it when an AI needs defined context or actions from another system—not as a substitute for clear workflow design.Elígelo cuando una IA necesite contexto o acciones definidas de otro sistema, no como sustituto de un flujo claro.

Host support variesEl soporte del host varía

One host may expose resources, approvals, extensions, and logs differently from another. Verify behavior in the surface you actually use.Cada host puede exponer recursos, aprobaciones, extensiones y logs de forma distinta. Verifica el comportamiento en la superficie que usas.

Connection is not acceptanceConexión no es aceptación

A successful tool call proves the connection worked. It does not prove the result is correct, authorized, complete, or ready to ship.Una llamada exitosa demuestra que la conexión funcionó. No demuestra que el resultado sea correcto, autorizado, completo o listo para publicar.

MCP, Skills, A2A, ACP: which connection do you need?MCP, Skills, A2A, ACP: ¿qué conexión necesitas?

MCP ↗

Agent ↔ tools and dataAgente ↔ herramientas y datos

Connect a capability; review each server and permission.Conecta capacidades; revisa cada servidor y permiso.

Agent Skills ↗

Reusable procedure → agentProcedimiento reutilizable → agente

Load instructions and resources on demand; not an access boundary.Carga instrucciones y recursos bajo demanda; no es un límite de acceso.

A2A ↗

Agent application ↔ agent applicationAplicación agéntica ↔ aplicación agéntica

Exchange tasks and status without exposing agent internals.Intercambia tareas y estado sin exponer el interior del agente.

ACP ↗

Editor / IDE ↔ coding agentEditor / IDE ↔ agente de código

Standardize the client connection; check supported protocol version.Normaliza la conexión del cliente; revisa la versión compatible.

These solve different boundaries, not competing levels of intelligence. None replaces authentication, authorization, or approval for consequential actions.Resuelven límites distintos, no niveles de inteligencia rivales. Ninguno sustituye autenticación, autorización ni aprobación para acciones importantes.

Collaboration, evidence, and deliveryColaboración, evidencia y entrega

GitHub + GitLab: where agent work becomes reviewableGitHub + GitLab: donde el trabajo del agente se vuelve revisable

GitHub Copilot’s app can export agent traces through enterprise-managed OpenTelemetry settings. Prompt and response content is excluded by default. Keep it that way unless an approved diagnostic need justifies content capture, and limit who can read the traces.La app GitHub Copilot puede exportar trazas mediante OpenTelemetry configurado por la empresa. Prompts y respuestas se excluyen por defecto. Conserva ese valor salvo una necesidad diagnóstica aprobada y limita quién puede leer las trazas. Official guide ↗Guía oficial ↗

Git records the history. GitHub and GitLab add the issue, review, automation, permission, and merge surfaces that turn an agent's file edits into a controlled team workflow.Git registra el historial. GitHub y GitLab agregan issues, revisión, automatización, permisos y merge para convertir ediciones de un agente en un flujo controlado de equipo.

GitHub Docs GitLab Docs

The universal agent-to-merge loopEl ciclo universal de agente a merge

1 · Issue

Define the outcomeDefine el resultado

Scope, acceptance checks, constraints, owner, and links.Alcance, criterios de aceptación, límites, responsable y enlaces.

2 · Branch

Isolate the changeAísla el cambio

Use a branch, fork, or worktree tied to the task.Usa una rama, fork o worktree ligado a la tarea.

3 · Commits

Record coherent stepsRegistra pasos coherentes

Keep the history understandable and the diff bounded.Mantén el historial comprensible y el diff limitado.

4 · PR / MR

Propose, do not declarePropón, no declares

Explain the change, evidence, uncertainty, and rollback.Explica el cambio, evidencia, incertidumbre y rollback.

5 · CI

Run configured proofEjecuta la prueba configurada

Builds, tests, lint, security checks, and previews.Builds, pruebas, lint, seguridad y previews.

6 · Review

Inspect the actual diffInspecciona el diff real

Resolve comments and verify behavior beyond the summary.Resuelve comentarios y verifica más allá del resumen.

7 · Merge

Make the acceptance decisionToma la decisión de aceptación

A person with authority accepts remaining risk and delivery state.Una persona con autoridad acepta el riesgo restante y el estado de entrega.

Git

Version historyHistorial de versiones

Local and distributed commits, branches, tags, diffs, and merges. Git is the history underneath either platform.Commits, ramas, tags, diffs y merges locales y distribuidos. Git es el historial debajo de ambas plataformas.

GitHub

Repository collaborationColaboración del repositorio

Issues, repositories, pull requests, reviews, Actions, status checks, Apps, protected branches, and rulesets.Issues, repositorios, pull requests, revisiones, Actions, status checks, Apps, ramas protegidas y rulesets.

GitLab

Project and delivery platformPlataforma de proyecto y entrega

A project can combine the repository with issues, merge requests, CI/CD pipelines, runners, permissions, security, and delivery settings.Un proyecto puede combinar el repositorio con issues, merge requests, pipelines CI/CD, runners, permisos, seguridad y entrega.

GitHubPull requests, Actions, rulesets, and MCPPull requests, Actions, rulesets y MCP

Plan in an issuePlanifica en un issue

Give the agent a linked issue with the goal, acceptance checks, repository, ownership, and what is explicitly out of scope.Da al agente un issue enlazado con objetivo, criterios, repositorio, propiedad y lo que queda fuera de alcance.

Review the pull requestRevisa el pull request

Read the purpose, linked issue, commits, Files changed diff, checks, comments, dependency or security findings, and the final review state.Lee propósito, issue enlazado, commits, diff de Files changed, checks, comentarios, hallazgos de dependencias o seguridad y estado final.

Use Actions as configured evidenceUsa Actions como evidencia configurada

Workflow jobs can build, test, lint, scan, or deploy. Passing status checks proves those configured jobs passed—not that every requirement is correct.Los jobs pueden compilar, probar, hacer lint, escanear o desplegar. Checks aprobados demuestran que esos jobs pasaron, no que todo requisito sea correcto.

Enforce the gateHaz cumplir la puerta

Protected branches and rulesets can require reviews, status checks, code owners, signed commits, or other policies before merge. A review request alone may not block merging.Ramas protegidas y rulesets pueden exigir revisiones, checks, code owners, commits firmados u otras políticas. Pedir revisión por sí solo quizá no bloquee el merge.

Grant repository access narrowlyLimita el acceso al repositorio

Prefer a GitHub App or fine-grained token with selected repositories and minimum permissions. A credential controls API reach; it is not a runtime sandbox.Prefiere una GitHub App o token granular con repositorios seleccionados y permisos mínimos. Una credencial limita alcance API; no es un sandbox de ejecución.

Treat GitHub MCP as an action surfaceTrata GitHub MCP como superficie de acción

GitHub provides an official MCP server in local and remote forms. Copilot Chat exposes a limited set; IDE and agent configurations can expose more. Begin read-only and allowlist the toolsets the task needs.GitHub ofrece un servidor MCP oficial local y remoto. Copilot Chat expone un conjunto limitado; IDEs y agentes pueden exponer más. Empieza en lectura y autoriza solo toolsets necesarios.

Agentic Workflows · generally availableAgentic Workflows · disponibilidad general

Define reasoning tasks in Markdown that compiles to Actions YAML. Start with issue triage or CI analysis, inspect generated permissions and safe outputs, and retain branch protections. Read-only defaults and sandboxing support review; the default Copilot engine requires a Copilot license.Define tareas de razonamiento en Markdown que se compila a YAML de Actions. Empieza con clasificación o análisis de CI, inspecciona permisos y salidas seguras y conserva protecciones de ramas. La lectura por defecto y el sandbox apoyan la revisión; el motor Copilot por defecto requiere licencia. Agentic Workflows ↗

Copilot Fusion + ProjectsCopilot Fusion + Projects

Fusion is a model-routing and orchestration system with Standard, High and Max effort; it replaces earlier auto/HydraFusion selections in supported clients. Projects and desktop computer use are public previews. Give a project explicit sources, tool access and review checks before parallel work starts.Fusion enruta y coordina modelos con esfuerzo Standard, High y Max; sustituye selecciones auto/HydraFusion anteriores en clientes compatibles. Projects y el uso del escritorio están en preview pública. Define fuentes, herramientas y comprobaciones antes de iniciar trabajo paralelo. Copilot Fusion ↗

Projects + computer use ↗Projects y uso del ordenador ↗

Pull request referenceReferencia de pull requests GitHub Actions GitHub rulesets GitHub MCP server

GitLabMerge requests, pipelines, protected branches, and MCPMerge requests, pipelines, ramas protegidas y MCP

Start with the projectEmpieza con el proyecto

A GitLab project is broader than its repository. It can carry issues, planning, code, merge requests, pipelines, permissions, and delivery configuration.Un proyecto de GitLab es más amplio que su repositorio. Puede contener issues, planificación, código, merge requests, pipelines, permisos y entrega.

Review the merge requestRevisa el merge request

Inspect commits, changes, discussions, reviewer state, approvals, conflicts, pipeline results, reports, and whether the issue's acceptance checks are met.Inspecciona commits, cambios, discusiones, revisores, aprobaciones, conflictos, pipelines, reportes y si se cumplen los criterios del issue.

Understand pipeline evidenceEntiende la evidencia del pipeline

GitLab CI/CD uses a .gitlab-ci.yml file to define jobs and stages run by runners. Merge request pipelines and merged-results pipelines answer different questions; know which one ran.GitLab CI/CD usa .gitlab-ci.yml para definir jobs y stages ejecutados por runners. Pipelines de merge request y merged results responden preguntas distintas; sabe cuál corrió.

Protect the destinationProtege el destino

Protected branches control direct pushes and who may merge. Approval rules, Code Owners, resolved discussions, and required pipelines can form separate merge checks.Ramas protegidas controlan pushes directos y quién puede fusionar. Reglas de aprobación, Code Owners, discusiones resueltas y pipelines requeridos forman checks separados.

Scope API and CI credentialsLimita credenciales de API y CI

For API or CI automation outside the official MCP OAuth flow, prefer a task-specific job, project, group, or fine-grained token over a broad personal token. Limit scope and lifetime, store securely, rotate, and revoke when the workflow ends.Para automatización de API o CI fuera del flujo OAuth oficial de MCP, prefiere un token de job, proyecto, grupo o granular para la tarea sobre uno personal amplio. Limita alcance y duración, almacena con seguridad, rota y revoca al terminar.

Connect in either MCP directionConecta en ambas direcciones MCP

GitLab's official MCP server uses an OAuth authorization flow and lets compatible external hosts reach project, issue, merge-request, and API operations. GitLab Duo can also act as an MCP client for approved outside servers. The server is generally available in GitLab 19.5. Select only needed toolsets, check group policy, and back off on rate-limit responses. Enabled Duo agents and flows can be started and tracked from compatible MCP clients.El servidor MCP oficial de GitLab usa un flujo de autorización OAuth y permite que hosts compatibles accedan a proyectos, issues, merge requests y operaciones API. GitLab Duo también puede ser cliente MCP de servidores aprobados. El servidor está disponible de forma general en GitLab 19.5. Selecciona solo toolsets necesarios, revisa políticas y respeta las pausas ante límites de solicitudes. Los clientes MCP compatibles pueden iniciar y seguir agentes y flujos Duo habilitados.

GitLab Duo CLI /goal

Define an outcome and acceptance checks for local goal-driven work instead of repeatedly saying “continue.” Enable Duo Agent Platform and check supported instance, CLI version, and tier. A completed goal still needs a reviewed merge request and passing pipeline.Define un resultado y criterios de aceptación para trabajo local orientado a objetivos en vez de repetir «continúa». Habilita Duo Agent Platform y verifica instancia, versión CLI y plan compatibles. Un objetivo terminado aún necesita un merge request revisado y un pipeline aprobado. Goal workflow ↗Flujo de objetivos ↗

Merge request reviewsRevisiones de merge requests GitLab CI/CD Protected branchesRamas protegidas GitLab MCP server

Practical bridgePuente prácticoUse MCP with Git platforms without giving away the merge gateUsa MCP con plataformas Git sin ceder la puerta de merge

ReadLee

Connect one repository or project read-only. Ask the agent to summarize the issue, relevant files, recent history, and open questions with links.Conecta un repositorio o proyecto en lectura. Pide resumen del issue, archivos relevantes, historial reciente y preguntas con enlaces.

PlanPlanifica

Require a bounded plan, owned files, expected checks, permission needs, failure modes, and the point where human approval is required.Exige plan limitado, archivos propios, checks esperados, permisos, fallos posibles y el punto de aprobación humana.

WriteEscribe

Enable only branch, commit, comment, or draft PR/MR tools that the job needs. Keep direct default-branch push, merge, deployment, and destructive tools blocked.Habilita solo herramientas de rama, commit, comentario o PR/MR draft necesarias. Bloquea push directo, merge, despliegue y acciones destructivas.

AcceptAcepta

Review the real diff and CI evidence, resolve discussion, name remaining risk, then let an authorized person decide whether to merge.Revisa el diff real y evidencia CI, resuelve la discusión, nombra riesgo restante y deja que una persona autorizada decida el merge.

Agent handoff contractContrato de entrega del agente

Require evidence that another person can review without replaying the whole session.Exige evidencia que otra persona pueda revisar sin repetir toda la sesión.

IntentIntención

Goal, linked issue, scope, owner, and acceptance checks.Objetivo, issue, alcance, responsable y criterios.

ChangeCambio

Repository or project, branch or worktree, commits, files changed, and PR/MR link.Repositorio o proyecto, rama o worktree, commits, archivos y enlace PR/MR.

ProofPrueba

Commands, tests, checks or pipelines, preview, reviewer comments, and observed results.Comandos, pruebas, checks o pipelines, preview, comentarios y resultados observados.

DecisionDecisión

Known uncertainty, remaining risk, rollback route, and the explicit next action for the human reviewer.Incertidumbre, riesgo restante, rollback y siguiente acción explícita del revisor humano.

Next learning stepSiguiente paso de aprendizaje

A merge finishes review. Ship & Operate teaches the deployment, verification, observability, and rollback loop that follows.El merge termina la revisión. Publicar y operar enseña el ciclo posterior de despliegue, verificación, observabilidad y rollback.

Continue to Ship & OperateContinuar a Publicar y operar

Agent operating rules

Instruction files: how agents learn your setup

Instruction files are markdown notes that agents load as persistent context: project rules, commands, style, review habits, tool boundaries, and the things you do not want to re-explain every session.

Codex AGENTS.md Claude memory

The simple version

Think of instruction files as onboarding docs for your agents. They do not replace your prompt, but they give every prompt the same ground rules before work starts.

1

Shared source of truth

Keep the common rules in `AGENTS.md` or a shared file under `docs/agents/`.

2

Tool-specific wrappers

Use `CLAUDE.md`, `GEMINI.md`, and `.cursor/rules` for the syntax and behavior each app expects.

3

Codex as maintainer

Ask Codex Agent to audit all instruction files, remove conflicts, and keep Claude Code, Gemini CLI, Cursor Agent, and Grok Build aligned.

How the files fit together

AGENTS.md
  Shared project rules for Codex and simple agent readers.

CLAUDE.md
  Claude Code project memory. Can import extra files with @path.

GEMINI.md
  Gemini CLI project instructions and memory notes.

.cursor/rules/*.mdc
  Cursor project rules: always, auto-attached, or agent-requested.

docs/agents/shared-agent-instructions.md
  Human-readable source that Codex can use to keep the others aligned.

Keep it short

Put durable rules here, not every idea. Long files crowd out the active task.

Make it testable

Say “run `npm test`” instead of “test thoroughly.” Concrete rules are easier to follow.

Refresh often

When the codebase changes, ask one agent to critique the instructions before another agent uses them.

Create the starter files

macOS

mkdir -p .cursor/rules docs/agents
touch AGENTS.md CLAUDE.md GEMINI.md docs/agents/shared-agent-instructions.md

Windows PowerShell 7

New-Item -ItemType Directory -Force .cursor\rules, docs\agents
New-Item -ItemType File -Force AGENTS.md, CLAUDE.md, GEMINI.md, docs\agents\shared-agent-instructions.md

A good Codex maintenance task

This is the kind of request that turns your setup into a repeatable system instead of a pile of half-remembered preferences.

Review AGENTS.md, CLAUDE.md, GEMINI.md, and .cursor/rules.

Create a short shared instruction source under docs/agents/.
Then update each agent-specific file so they agree on:
- project purpose
- setup commands
- test commands
- coding style
- file boundaries
- when to ask before taking action

Call out contradictions before editing anything.

Design the system around the model.Diseña el sistema que rodea al modelo.

The advanced map: context, access, actionEl mapa avanzado: contexto, acceso y acción

Where is the context?¿Dónde está el contexto?

Which AI can see it?¿Qué IA puede verlo?

Answer or action?¿Respuesta o acción?

Review the real thingsRevisa lo que realmente importa

From an advanced AI tool to a trustworthy AI systemDe una herramienta de IA avanzada a un sistema de IA confiable

User + taskUsuario + tarea

API + modelAPI + modelo

Output + toolsSalida + herramientas

Retrieve + groundRecuperar + fundamentar

Measure the jobMide el trabajo

Permissions + stopPermisos + detención

Workflow or agent?¿Flujo o agente?

API call, structured output, and tool contractLlamada API, salida estructurada y contrato de herramienta

Context and retrievalContexto y recuperación

Evaluation is a loop, not a launch checkboxLa evaluación es un ciclo, no una casilla de lanzamiento

Workflow vs agentFlujo de trabajo frente a agente

Safety and permissionsSeguridad y permisos

Human approval belongs at the boundaryLa aprobación humana pertenece al límite

Use the room that already has the context

Microsoft work context

Google and source packets

Local files, notes, and personal knowledge

Repos, branches, and project files

Know where the AI is working

Chat answer

Connected app

Local agent

Remote/work agent

Microsoft Copilot for web, work chat, and Microsoft 365 context

Microsoft Copilot

Copilot Chat

Microsoft 365 Copilot

Agents, Researcher, Analyst, and Office creation

Chat, Work, and Codex serve different jobs—and have different model choices.

ChatGPT app

Codex agent in ChatGPT desktop + CLI

Claude for careful work. Sonnet 5.5 for efficient execution. Opus 5.5 for proactive work. Fable 5.1 and Mythos 5.1 for the frontier.

Claude app

Claude Cowork

Claude Code

Gemini for Google context. Gemini Notebook for source packets. Antigravity for agent coding.

Gemini app + Workspace

Gemini Notebook

Antigravity CLI (consumer Gemini CLI transition complete)

Cursor IDE + Agent for learning by building

Why people pick it

Your progression path

Grok for current context. Grok Build for code. Grok Bot for demonstrated routines.

Grok app + xAI API

Grok Build

Grok Bot

CLI setup and prerequisites

macOS baseline

Windows baseline

Windows Terminal setup

A clean profile starter

MCP: the connection layerMCP: la capa de conexión

The AI applicationLa aplicación de IA

One server connectionUna conexión a servidor

Focused capabilityCapacidad enfocada

The real data or actionEl dato o acción real

What a server can offerQué puede ofrecer un servidor

A safe first connectionUna primera conexión segura

Local process boundaryLímite de proceso local

Remote service boundaryLímite de servicio remoto

MCP ↗

Agent Skills ↗

A2A ↗

ACP ↗

GitHub + GitLab: where agent work becomes reviewableGitHub + GitLab: donde el trabajo del agente se vuelve revisable

Define the outcomeDefine el resultado

Isolate the changeAísla el cambio

Record coherent stepsRegistra pasos coherentes

Propose, do not declarePropón, no declares

Run configured proofEjecuta la prueba configurada

Inspect the actual diffInspecciona el diff real

Make the acceptance decisionToma la decisión de aceptación

Version historyHistorial de versiones