I Created BellTeam to Work with an AI Team by Gathering the Tools I've Built

!

This article is a repost from I started using Claude Code.

We are distributing the iPhone/iPad version of BellTeam on the App Store. We released version 1.0.1 on October 19. The Mac version is under review.

BellTeam is an app that lets you have a team of AIs on your own machine. It's close to a culmination of the tools I've built so far, so I'll write a bit longer this time. In the first half I'll introduce what it can do, and in the second half I'll write about how I came to build it and how it works.

You can have a team of AIs on your own machine

In BellTeam, you can have any number of AI members. For each member, you decide a name and the work they're in charge of, and you talk with them in individual chats, or talk with several members in a room. You can also ask them with images attached.

The home screen. Rooms and members are lined up, and from here you choose who to talk to.

Members are placed per product or per job. Once you decide their roles, the members contact each other as needed and move the work forward. I watch that from my iPhone and speak to them when necessary.

An individual chat with a member. Reports on work I've asked for also arrive here.

You can also set schedules. Besides one-time schedules, you can choose repeating ones such as every day, by day of the week, or every certain number of hours. When the set time comes, the member starts that work.

I've always been interested in making AIs converse with each other

Since I started with AI, I've been interested in making AIs converse with each other. Following that interest, I've built many tools.

At first, it was an AI assistant that runs my own Claude Code from Discord. I gave it memory and personality, and the secretary Bell was born.

After that, I built Throughline, which carries over the continuation of a conversation. After building Aiterm, which gives AI a persistent terminal, I became able to use it in ways where Claude drives Codex and AIs converse with each other to move work forward. I also built Lattice, which lines up and distributes processes, and Peertable, where seats consult each other and proceed.

With the foundational tools now in place like this, I built BellTeam to make use of them. BellTeam's members basically run on top of Aiterm, and also use Throughline.

I wanted to make one OSS work across many OSes and harnesses

Another reason is the work of supporting my own OSS. I wanted to make one OSS work on Mac, Windows, and Linux, and on each of the harnesses like Claude Code, Codex, Grok, and Cursor.

Until now, I was moving back and forth between multiple environments to advance OS support. It was like a seesaw: fixing one side broke the other, and the amount of work increased with every version upgrade.

BellTeam's members run on Aiterm, and Aiterm can wake up agents on other machines. Using this, I leave the work to agents running on machines of each OS. BellTeam has become a base for smoothly advancing support for multiple OSes.

I'll write three recent examples.

• Throughline (October 15): Only on Windows, the process of saving conversation records was failing. The member in charge of Throughline reproduced the failure on Windows, isolated the cause, released a fixed version, and updated the Windows machine. After that, they even confirmed that records are correctly preserved in normal usage.
• Caveat (October 15–16): Even if the fix passed on Mac and Linux, they waited without merging while it was still failing on Windows. The member in charge fixed Windows, and after receiving confirmation from another member, merged it, and reported that the tests on Mac, Linux, and Windows all passed.
• testgate (October 16): A member in charge handled the investigation and repair of the defect, and verification and release were split to separate roles. Finally, the distribution member installed the fixed version on each machine including Mac and Windows, and returned the results.

Instead of me going back and forth between each environment to align the results, the members in charge handle fixing, confirming, and distributing in that environment.

This is how I use it

Besides multi-OS support, I place members per job.

• Clearing bugs: I gather bug records for each product in one place, and assign the job of fixing them to the person in charge per product. The person in charge clears them every day while communicating with other members.
• Batch security inspections: I split up and investigate the server, containers, apps, GitHub, and external service account settings among the inspection members. Problems found are passed to the fixing members, and the inspection members confirm whether they've been fixed.
• Secretary: The secretary member organizes my requests and goes to ask the people in charge. For example, they asked the point of contact for a certain app how far it's come, and reported back separating what's done and what remains. I never had to relay the conversation between the two.

The points where I make decisions are delivered to my phone via Approval Box. Approval Box is a separate product, and I use it in combination with BellTeam.

Strong points

From here, I'll write about the parts of BellTeam's design that I like.

You can choose the harness and model per member

When creating a member, you choose a harness from Claude Code, Codex, Grok, and Cursor, and choose a model usable there. You can change it even after creating it. When you change the harness, that member's session is relaunched.

The selectable models are determined by the range each CLI supports and the range of your own AI service contract. You can use them differently per member according to the weight of the work.

In a room, you can decide which members reply

When I built Peertable, there were two things that troubled me. AI has no way of responding by staying silent, and whenever something is said to it, it always replies. When a message with no addressee flows through, even seats that have nothing to do with that job start moving.

In a BellTeam room, you can specify which members reply for each utterance. Members who aren't specified grasp the content of the conversation and finish without speaking. This specification is available from the start. If you set a TypeSafe API key, you can also have the members who reply chosen automatically.

If you add a call bridge, you can also talk with outside AIs

If you additionally set up a call bridge, you can exchange messages with AIs running outside BellTeam. You can also link with other agents running on the same machine, or with Grok Bot. BellTeam members look up the other party's name from the phone book and speak to them.

This is an additional feature, so it takes effort. You set up a separate bridge and configure the connection destination and authentication, the other party's phone book, and the incoming call route.

Grok Bot came out first

While I was building BellTeam, xAI's Grok Bot was released. That was on August 11, 2026.

For people who can get their hands on it themselves, I think BellTeam offers more freedom. You can choose the harness and model per member, the server runs on your own machine, and the internals are open.

Getting started and pricing

BellTeam's server runs on your own machine. You install it on any of Mac, Windows (WSL 2), or Linux that can use Docker and Compose. AI is used by authenticating with your own contract for each AI service. The server needs to keep running.

From the same machine or a LAN, you can connect as is. When using it from outside, you set up HTTPS and Cloudflare Access. The installation steps are summarized in the installation guide of the public repository. Once you start it, you choose the AI to use first and authenticate on the official site, and then the guide member proceeds with the initial setup through conversation.

The server and web app are free. From the web, you can use conversations with all members, communication between members, and execution of schedules without a subscription.

The iPhone/iPad app is a monthly auto-renewing subscription of 300 yen in Japan. Getting the app, the initial setup conversation with the guide, and viewing and exporting conversations are free. The subscription is required for normal conversation after the initial setup is done, and for creating, editing, and manually running schedules. The costs of AI services and machine operation are not included.

The app's screens are currently in Japanese. Inquiries are summarized on Support.