I run several coding agents at once. Claude Code for most of the work, OpenAI's Codex for reviews and tests, Google's Antigravity for research. Each one on its own account, each with its own role.

For a while that meant a terminal per agent. It worked until it didn't: I'd come back in the morning to an agent that had hit its usage limit during the night and simply stopped, half a task on disk and nobody to pick it up. Or my phone buzzing every minute because another agent was "Waiting." on a limit.

So I built Seat Console: a self-hosted web page (and a Telegram bot) that drives all of them.

Seat Console tour

What it does

One page for every agent. Claude Code, Codex and Antigravity seats sit in one sidebar, each with its own login and a usage meter that shows the real 5-hour and weekly windows and when they reset. Several accounts per agent work, each in its own seat.

Multi-sided projects. You write one brief. A planner seat turns it into a plan with tasks tagged by discipline (design, backend, video, copy, deploy...). You approve it with one tap, and every discipline runs in parallel on the seat fixed for it. Tasks that need another task's output wait for it, and at the end the owner seat integrates everything and checks it.

Limits are respected, not dodged. When a seat hits its usage limit, its tasks wait for that seat's reset and then continue where they stopped. Work is never moved to another account to get around a limit. That rule is in the code, with a test that proves no failover happens.

Less quota burned on cache. Long agent conversations re-read their whole history on every turn. Task briefs ask agents to hand heavy reading to subagents and keep a short progress file, and when a task's conversation gets long, the console continues the task in a fresh, short conversation on the same seat.

The small things that add up:

  • a conversation stopped by the usage limit resumes by itself after the reset
  • free limit resets show on each seat's card, one tap uses one
  • images both ways, including images the agents generate
  • search across every conversation, archived ones too
  • each task in a project can run on the model that fits it, not always the biggest
  • a settings page that shows a new install what it can turn on
  • 15 interface languages, right-to-left included

How it works

Plain Node, no framework, no build step. Each agent runs as its own CLI process (claude, codex, agy) under a small host process, so conversations survive a console restart. The page talks to the server over SSE. Codex and Antigravity are driven through thin adapters that translate their JSON event streams into the same shape the console uses for Claude.

Try it

git clone https://github.com/beriki770-ship-it/seat-console.git cd seat-console 

Then follow the one-minute quick start in the README. It's free and Apache-2.0. The usage-reset endpoints it reads are the ones the official CLIs use; they're undocumented and may change, so reading them is opt-in.

If you run more than one coding agent at a time, I'd like to hear what you'd want it to do next.