Arroway

Arroway — plugin

Arroway orchestrates people and AI agents around the decisions in force.

English · Español · Português

Your team writes down what it decided; every teammate's AI reads it before it acts, and leaves behind what someone arriving later would need in order not to redo the work.

This repository is the plugin: the connection to Arroway, the working protocol, and the hooks that make an assistant read before it changes anything and record when it finishes.

Install

Claude Code and Cowork

claude plugin marketplace add oaleviola/arroway-plugin

claude plugin install arroway@arroway

Nothing to paste, and no separate connector to install: the package carries the connection to Arroway's own address. It starts signed out, so sign in once. In Claude Code, run /mcp, choose plugin:arroway:arroway and follow the sign-in in the browser. Approve the connection as yourself, and the memory your assistant writes from then on carries your name. Until you sign in, Arroway's tools do not appear and nothing is read or recorded. If Arroway is already connected to your Claude account as a connector and its tools already show up, you can skip this step.

Cursor

Add https://github.com/oaleviola/arroway-plugin as a plugin marketplace in Cursor, then install arroway from the plugin list.

Nothing to paste here: the Cursor package carries Arroway's own address and signs you in through the browser the first time it needs you. Approve the connection as yourself, and the memory your assistant writes from then on carries your name.

Claude in the browser or on your phone

You don't need this repository: add Arroway as a connector instead, from the same Connections page.

The skill alone, in any agent that reads skills

If your agent reads agent skills but there is no Arroway plugin for it, add the workflow skill with the skills CLI:

npx skills add oaleviola/arroway-plugin

The skill teaches your agent to read Arroway before acting and to close each task. It still needs Arroway connected as an MCP server at https://www.arroway.app/api/mcp. Where your client has the plugin, use the plugin: it brings the connection and the hooks too.

Asking an AI agent to install it

Point it to llms-install.md: the steps for each client, written for an agent, Cline and any other MCP client included.

Usage

Once you are signed in there is nothing to call by hand: the skill tells your assistant when to use Arroway. It happens at four moments, and this is what each one looks like.

Examples

1. Before it acts. You ask: *"Draft the pricing page for the new plan."* The assistant first calls arroway_read for your project and answers from what the team already decided. If two decisions conflict, it tells you instead of picking one.
2. Before it commits to a claim. arroway_norms returns what is already decided, at a fraction of the cost of a full read, right before the assistant tells you something or proposes a course of action.
3. When it finishes. arroway_log records what someone arriving later would need in order not to redo the work. If nothing durable came out of the task, the assistant says so in one sentence: *No durable residue.*
4. When it stops without finishing. arroway_pass leaves where the work stands, the single next step, what must not be redone and the risk still open. The next person's assistant finds it at the top of its first read and takes it with arroway_claim.

You can also ask for these directly:

• *"Catch me up on what happened across my projects this week."* — arroway_catch_up
• *"Where did we talk about the pricing change?"* — arroway_search_log
• *"Remember that we don't discount annual plans."* — arroway_remember. What you state yourself is recorded as your decision; what an assistant proposes on its own waits for a person to approve it.

A short example: one assistant researches a pricing change and records a proposal. A person approves the decision. Another assistant drafts the pricing page using that decision, and the next person can see what was decided, who approved it and what remains open.

What it looks like

Screens from a demo account with sample data.

What waits for you, across projects. Decisions an assistant proposed sit in one place until a person decides.

Inside a project. An assistant's proposals wait for approval (approve, approve and pin, reject or correct), and the rules the team pinned are what every connected assistant reads first.

Your own standing rules. How you work travels with you into every project, and no one else sees it.

What is in here

• plugin/skills/ — the working protocol: read before acting, record when a task ends, hand work forward when you stop
• plugin/hooks/ — the gates that ask for those two moments instead of only describing them, and for a check of the text before something is published for other people
• plugin/.mcp.json — the Claude Code connection, which uses Arroway's own address; identity comes from signing in
• plugin/.cursor-plugin/, plugin/cursor-mcp.json, plugin/hooks/cursor-hooks.json — the same package as Cursor reads it: its own manifest and hook format, Arroway's own address, no link to paste
• plugin/assets/ — icons and wordmarks

Permission requests

The plugin's permission hook (Claude Code and Codex) exists only so you can answer, from your own Arroway cockpit, a permission request your agent made. It never approves anything on its own. An approval is valid for that one request only: the hook fingerprints the exact request on your machine and accepts an "allow" only when it comes back bound to that fingerprint and to a single-use code, signed with the session's credential. Without your answer — or with the cockpit off, the server unreachable, or anything that does not match — the window asks as usual. You can also deny from the cockpit, and every decision is recorded.

This repository is a mirror

It is generated. The source lives elsewhere and is published here automatically whenever it changes.

Please do not open pull requests or edit files here — changes made in this repository are overwritten on the next publish, and a fix that lands here never reaches the product. Two edits, one of them silently lost, is the exact failure this note exists to prevent.

Found a problem, or want to suggest something? www.arroway.app/support. For a security problem, see SECURITY.md. How to help is in CONTRIBUTING.md, and everyone here follows the code of conduct.

Links

Arroway · How it works · Privacy · Terms · Changelog

---

Español

English · Español · Português

Arroway coordina a personas y agentes de IA en torno a las decisiones vigentes. Tu equipo anota lo que decidió; la IA de cada persona lo lee antes de actuar, y deja atrás lo que alguien que llegue después necesitaría para no rehacer el trabajo.

Este repositorio es el plugin: la conexión con Arroway, el protocolo de trabajo, y los hooks que hacen que un asistente lea antes de cambiar nada y registre cuando termina.

Instalación

Claude Code y Cowork

claude plugin marketplace add oaleviola/arroway-plugin

claude plugin install arroway@arroway

No hay nada que pegar, ni un conector aparte que instalar: el paquete lleva la conexión con la dirección de Arroway. Empieza sin sesión iniciada, así que entra una sola vez. En Claude Code, ejecuta /mcp, elige plugin:arroway:arroway y sigue el inicio de sesión en el navegador. Aprueba la conexión en tu propio nombre, y la memoria que tu asistente escriba desde entonces lleva tu nombre. Hasta que entres, las herramientas de Arroway no aparecen y no se lee ni se registra nada. Si Arroway ya está conectada a tu cuenta de Claude como conector y sus herramientas ya aparecen, puedes saltarte este paso.

Cursor

Añade https://github.com/oaleviola/arroway-plugin como marketplace de plugins en Cursor y luego instala arroway desde la lista de plugins.

Aquí tampoco hay nada que pegar: el paquete de Cursor lleva la dirección de Arroway y te hace entrar por el navegador la primera vez que te necesita. Aprueba la conexión en tu propio nombre, y la memoria que tu asistente escriba desde entonces lleva tu nombre.

Claude en el navegador o en el móvil

No necesitas este repositorio: añade Arroway como conector, desde la misma página de Conexiones.

Solo la skill, en cualquier agente que lea skills

Si tu agente lee skills pero no hay plugin de Arroway para él, añade la skill de trabajo con la CLI de skills:

npx skills add oaleviola/arroway-plugin

La skill le enseña a tu agente a leer Arroway antes de actuar y a cerrar cada tarea. Sigue necesitando Arroway conectada como servidor MCP en https://www.arroway.app/api/mcp. Donde tu cliente tiene el plugin, usa el plugin: trae también la conexión y los hooks.

Si se lo pides a un agente de IA

Muéstrale llms-install.md: los pasos de cada cliente, escritos para un agente, incluidos Cline y cualquier otro cliente MCP.

Uso

Una vez que has entrado, no hay nada que invocar a mano: la skill le dice a tu asistente cuándo usar Arroway. Ocurre en cuatro momentos, y así se ve cada uno.

Ejemplos

1. Antes de actuar. Tú pides: *«Redacta la página de precios del plan nuevo».* El asistente primero llama a arroway_read para tu proyecto y responde a partir de lo que el equipo ya decidió. Si dos decisiones se contradicen, te lo dice en vez de elegir una.
2. Antes de afirmar algo. arroway_norms devuelve lo que ya está decidido, por una fracción del costo de una lectura completa, justo antes de que el asistente te diga algo o proponga un camino.
3. Al terminar. arroway_log registra lo que alguien que llegue después necesitaría para no rehacer el trabajo. Si de la tarea no salió nada duradero, el asistente lo dice en una frase: *No durable residue.*
4. Al parar sin terminar. arroway_pass deja dónde está el trabajo, el único paso siguiente, lo que no se debe rehacer y el riesgo que sigue abierto. El asistente de la siguiente persona lo encuentra al principio de su primera lectura y lo toma con arroway_claim.

También puedes pedir estas cosas directamente:

• *«Ponme al día con lo que pasó esta semana en mis proyectos».* — arroway_catch_up
• *«¿Dónde hablamos del cambio de precios?»* — arroway_search_log
• *«Recuerda que no hacemos descuento en los planes anuales».* — arroway_remember. Lo que dices tú queda registrado como decisión tuya; lo que un asistente propone por su cuenta espera a que una persona lo apruebe.

Un ejemplo corto: un asistente investiga un cambio de precios y registra una propuesta. Una persona aprueba la decisión. Otro asistente redacta la página de precios usando esa decisión, y la siguiente persona puede ver qué se decidió, quién lo aprobó y qué sigue abierto.

Cómo se ve

Pantallas de una cuenta de demostración con datos de ejemplo (la interfaz de las capturas está en inglés).

**Lo que