I document the rapid evolution of tooling for my web app, nocodefunctions.com. The previous steps are documented in From NetBeans to Claude Code (Jan 2026).

In early January, I had left my coding environment in this configuration, which was great:

  • one single Debian server for dev and prod
  • Claude Code installed on it
  • I connect on SSH to it, from my local Windows machine
  • in the terminal (Putty), I “vibe code” the development of my web app
  • I can also SSH to my server directly from my Android phone using ConnectBot: effectively coding while on a metro ride.

This setup served me well, but I hit some real limits:

1. Coding from the terminal on the phone is OK but not fantastic

Sending instructions to Claude from a mobile with ConnectBot is not a great experience. The app connects to the server so this is a terminal interface, and on a mobile phone there are some keys (arrows, escapes, tabs…) that are not readily available, necessitating to install an app with a more advanced keyboard on the phone. In the end the experience is OK but not comfortable. Here is how it looks:

2. Claude: token rate limits every 5 hours and every week

Vibe coding increases productivity for sure, and it comes with token spending. I became regularly blocked by the 5-hour token reset period imposed by Claude. The weekly limit on token consumption was also looming large. I am on the $20 monthly plan with Claude, and didn’t want to jump to the $90 or $200 plan. The situation became very frustrating.

1. Adding Gemini CLI to the server!

Google’s most visible product to compete with Cursor or Claude is AntiGravity and so I had missed that Google has released a CLI version of Gemini (since summer 2025 as per this Github repo).

Anyway, if I am limited by Claude’s token limit, why not use the token allowance I have from my Gemini subscription?. Installing Gemini CLI on the server is straightforward.

At first, I switched between Claude and Gemini : launching one when the other’s limit was reached. If you wonder, Gemini 3.1 pro is good enough to work with on code. It is at the same level as Claude sonnet or opus. Gemini’s other models are insufficient.

But then I thought naturally…

2. Getting Claude and Gemini to coordinate their efforts!

It is natural to try the next step: getting Claude, Gemini, and possibly other forthcoming CLI agents to coordinate their efforts on tasks I submit to them.

To create this, I started a conversation with Gemini to devise a plan. My entire prompt was:

On my debian server I installed in 2026 both Claude Code and Gemini CLI with flags to bypass human confirmation . I created scripts and md files so that they should coordinate work when I formulate a task to accomplish . works great.
only issue is that when one is finished on a set of subtasks, it returns to the command line and wait for my instructions. there is no mechanism for it to detect that the other has written down new subtasks in the md file to be picked up now.
how to solve that? i would love that I first give them both a general task, and then they remain active on it. they would also send their remarks, questions, progress reports... in a centralized place where I could then answer (to both at the same time)

Gemini provided a simple plan, that I then asked my 2 agents (Claude and Gemini) on debian to implement. It works great now, mostly through a mix of shared .md files where both agents report their progress and pick up instructions to work on the next steps.

3. Using the web then Signal to improve my comfort as an orchestrator!

Since I left the comfort of using an IDE like NetBeans (2010-2025) then Cursor (November-December 2025), working in the terminal with Claude and Gemini is not so bad, but is not entirely comfortable either. Especially, as I explained above, when connecting on the terminal from my phone, things get really tiny and typing messages is not super easy.

First thing I did was asking the agents to create a web page interface: a page that I could double tap to edit it, so that I could monitor progress and add tasks directly in the text. They found a Python lib for that:

web interface for agents coordination

Not bad at all. Then I remembered that levelsio / Pieter Levels had mentioned his use of Telegram to pilot his agents: