Answers

How to install and use Arroway in the main AI ecosystems

You pick the tool you already work in, connect Arroway to it, and start working. The installation page asks one question — where do you use AI? — and shows one route, not a menu of every possibility. It does not ask you to sign in until the moment a sign-in is actually needed.

Last updated November 14, 2026

One route at a time, for the tool you use

Claude, ChatGPT or Codex, and other tools that speak the same connection standard each get their own path, written for the screens you actually see in that tool. Commands appear when you reach the step that needs them, not all at once at the top. There is no generic “for IDEs” instruction, because the interfaces are not the same and a generic one would be wrong for most of them.

The connection is required; the plugin is offered only where it works

Connecting to the server is what makes the memory readable and writable — that part is the same for everyone. Some tools also support a small plugin, which adds the reminder to read at the start and write at the end. It is offered on those tools and not mentioned on the others; it is never presented as something you must install.

Personal account or team workspace: both routes, no interrogation

You are not asked what kind of account you have. A personal account gets a private address to paste into your tool — that address is your identity, so treat it like a password. A Claude Team or Enterprise workspace gets the sign-in address of this site, and each person authorises individually. Either way, what you write is attributed to you, not to a workspace.

When not to connect yet — and what never goes in

If you work alone, in one tool, and a single instruction file still fits on a page and still gets read, you are not yet in the situation this solves: keep the file and revisit when a second tool or a second person arrives. Do not connect an account you do not control, and do not paste the private address into a shared document, a commit or a chat — it is your identity, so anyone holding it writes as you. And whatever you record, passwords, tokens and API keys do not go into the memory, even if the assistant offers them as ‘context’: that is the one line the product refuses to cross, whoever asks.

The mistake that makes the first week disappointing

People try to load everything at once and approve it in bulk. The project fills with forty entries nobody read, the assistant's first reads are crowded, and the habit of reviewing dies in a week. Start with the five things your team repeats most often, approve them one by one, and let the rest arrive as the work produces it. A small project that is true beats a large one that is mostly old.

What it looks like in practice

A hypothetical four-person product studio: two people work in Claude, one in Codex, one in ChatGPT, and the rules that govern the work are scattered over about twenty documents. Nobody needs to migrate anything on day one. The first person connects, asks the assistant to read the project, and gets back an empty project plus its stated purpose. The point of the first hour is not to load everything — it is to get five or six things the studio would be annoyed to explain again, recorded as proposals a person approves. The other three connect over the following days and find those five or six already waiting.

The first hour after you connect

  1. Pick the tool you already work in on the installation page and follow the one route it shows. Do not look for the steps here: this page deliberately keeps no copy of them.
  2. Open a new conversation and ask the assistant to read the project before it starts anything. An empty project is a normal first answer, not a failure.
  3. Say in one plain sentence what the project is for — for example, “the studio's client work for the retail accounts”. The statement is what keeps a later decision from being filed in the wrong project, and a person can correct it at any time.
  4. Point the assistant at the three to five documents that already settle arguments in your team: the naming convention, the review rule, the thing you will never ship on a Friday. Do not point it at all twenty. Volume is not the goal.
  5. The assistant proposes what to record. Read each proposal, keep the ones that are true and still current, edit the ones that are almost right, and drop the rest. What you keep is a decision; what you did not read stays a proposal.
  6. Next day, open a different tool and ask it to read the same project. The decisions from step 5 should already be there, along with a note of what the read delivered and what it left out. If they are not, that is a defect to report, not something to repeat yourself to cover.

Questions people ask about this

Where do I actually do it?
On the installation page of this site, which is the only place the steps are written. This page explains what installing means and what you get; it deliberately does not keep a second copy of the instructions, because two copies drift.
Do I need to know what MCP is?
No. It is the connection standard underneath, and the routes are written for the interface you see in your tool, not for the protocol.
Is there anything to keep updated on my machine?
No. What gets installed does one small job — remind the assistant to read and write — and it is frozen at install time on purpose. Everything that could change lives on the server, so an improvement reaches everyone the moment it ships.
Install ArrowaySee how it works →

Where this is verifiable

The installation page and the distribution decision recorded in the repository. Availability in a given directory or marketplace is stated only after external confirmation.

https://www.arroway.app/en/answers/install-across-ai-tools

Continue exploring

See all answers