Run three agents. Answer only the one that asks

Put Claude Code, Codex and OpenCode on projects of their own and let them work side by side on one Compute, each on its own subscription or key. One list shows every tab and flags the one waiting for you. Pro runs five terminals at once for $75 a month, and a new project costs $0.

Get my Compute See the plans

From $39 a month · Docker and sudo included · Cancel from the dashboard any time

Terminals · every project5 tabs · 3 projects

Project · tab	Agent	Last line	State
web · t1	Codex	Which date library should the calendar use?	Needs you
api · t1	Claude Code	▸ go test ./... ok 412 tests	Working
api · t2	shell	dev@fleet:/work/api$	Idle
web · t2	Codex	▸ pnpm build compiled 214 modules	Working
docs · t1	OpenCode	✓ docs/api/orders.md written, 14 endpoints	Done

Two agents on one laptop fight over everything

They share one CPU, one set of ports, one Docker daemon and often one checkout. Branches collide, test databases clash and the fan never stops.

Then you click through every terminal to find the agent that stopped to ask a question twenty minutes ago.

A morning with three agents

09:00
Split the work by project
The API, the web app and the docs become three projects on your Compute, each with its own checkout in /work. Each one is a folder, and a new one costs $0.
09:05
Start them all
Every new tab is one launch: Claude Code in the API, Codex in the web app, OpenCode in the docs. Type the task and move on to the next tab.
09:40
One asks, you answer
The list marks the quiet tab Needs you and your phone buzzes. Answer in one line and go back to your own work.
12:30
Merge three branches
Each agent pushed its own branch from its own project. Nothing collided because no two agents shared a checkout.

Three agents that never collide

Separate checkouts, nothing colliding
Each agent works in its own project with its own checkout, settings and login, and sees nothing but that project and its own home folder.
The quiet tab finds you
Terminals mode lists every tab of every project as Working, Needs you, Done or Idle. Turn on push or Telegram and your phone hears about the quiet one.
The right agent for each repo
Each project gets the agent that suits it: Claude Code, Codex or OpenCode, each on its own subscription or API key.
Room on every subscription, in view
The limits board shows which account has headroom, so the next job starts where the window just refilled.

The size for this job

Pro · $75

Pro runs five terminals at once on 4 vCPU and 8 GB: three agents, a test run and a shell, with room for builds and a database.

Manifestdetachbox · node 1

Compute: fleet
Projects: /work/api · /work/web · /work/docs
Plan: Pro · 4 vCPU · 8 GB · 100 GB · $75/mo
Agent: Claude Code · Codex · OpenCode
Tabs: 5
Services: postgres:18
Runs: Around the clock, lid closed

Pro

Our pick for agent work

$75a month

5 agents plus the Docker services they test against

vCPU: 4
RAM: 8 GB
disk: 100 GB
Internet: 100 Mbit
Terminals at once: 5

Get my Compute

Questions about this

Does each agent need a Compute of its own?

No. Your one Compute runs every project side by side, each agent on its own subscription. Your size sets how many terminals run at once: 3 on Start, 5 on Pro, 8 on Max, 10 on Business, 15 on Business Plus, 20 on Business Max, 30 on Business Ultra. Need more at once? Change size on the running Compute, no restart.

How does the list know a tab needs me?

The agent CLIs report their own state: working, waiting on you or done. The tab waiting for you lights up in its place in the list. Switch on push, Telegram or email and your phone hears about it too.

Can one subscription serve several agents?

Yes. Connect the account once and launch as many agents on it as you like, in any project. They share that account’s limits, and the board shows what is left.

What this job uses

More jobs people run