Stop opening Codex
and Claude Code.
코덱스와 클로드 코드에 들어가지 마세요!
One chat. A Main that plans the work and sends every task to the right machine and subscription.
Four machines, two agents, and you are the scheduler.
| Machine | Codex | Claude Code |
|---|---|---|
| Mac Studiohome · always on | 46% weekly | not signed in |
| MacBook Prohome | 13% weekly | 40% session |
| MacBooktravel hotspot | 17% weekly | 21% weekly |
| Cloud machineVercel Sandbox | device sign-in on the machine | API-key models via OpenCodex |
Which machine is awake? Which plan still has quota? Which tab was that task in?
Every new agent and every new cloud adds another tab to watch.
Quotas reported live by each machine's GOAT on 14 Oct 2026.
I ran out of compute. Then I found the pattern.
OpenCodex
Building a universal provider proxy for coding agents.
Home
Building it while raising a child at home.
Compute crunch
Ran out of compute on my own machines.
Many clouds
Cloud compute services kept appearing. Each one: another tab, another login.
The insight
One main orchestrator coordinating many workers is enough, and it stays maintainable.
“Give an instruction in one place, and the Main coordinates my work, allocating every task to the right machine and subscription.”
GOAT Anywhere · Fast Forward · Agents & Factories03 / 11One chat. The Main plans and allocates.
- 01Say it once. No picking a machine, a model or a tab.
- 02The Main plans. A small, fast model on AI Gateway splits the instruction into 1 to 4 tasks.
- 03The router allocates. Each task goes to a machine that is online, signed in to the right agent, with the most quota left.
- 04You stay in control. Every assignment shows where it went and why, with live status, steer and stop.
Main → router → machines × subscriptions.
You
Any browser. One chat box, live status, steer, stop.
Main · on Vercel
Planner: small, fast model on AI Gateway → 1 to 4 tasks.
- machine online
- agent signed in there
- lowest quota used
- schedules → always-on machines
- otherwise → cloud + AI Gateway
Each machine runs GOAT and its agents with your own subscriptions.
Three minutes, one instruction.
- 0:00Connect your agents once. Cloud machine launches; Codex device sign-in; Claude Code login inside the machine.
- 0:40One instruction in the Main chat.
- 1:00Chips appear: “→ Mac Studio · Codex” with a one-line reason each.
- 1:40Open GOAT on a machine through a one-time pairing link.
- 2:15Factory job on a schedule; Vercel Cron dispatches it.
- 2:40Steer or stop a running task from the browser.
It built itself today.
14 Oct 2026. Lane lengths are schematic; the order of events is real.
GOAT Anywhere · Fast Forward · Agents & Factories07 / 11Your tokens never leave your machine.
Codex
The official device-code sign-in runs on your machine. Your ChatGPT plan works there; the token stays there.
Claude Code
You sign in to Claude Code inside your own machine. No subscription proxying, ever.
OpenCodex
Routes API-key models through Vercel AI Gateway with Vercel OIDC. No keys in code.
Signed session
HttpOnly cookie behind an access code
Pairing links
single use, expire in 5 minutes
Isolation
one sandbox per user, never shared
Kill switch
every machine has a Stop button
The whole control plane is Vercel.
Next.js 16
Control plane, sign-in, Main chat, machine pages.
Vercel Sandbox
The cloud machine. GOAT healthy in ~15 s on Ubuntu with Node 24, from a 49 MB Linux runtime.
Vercel Cron
Dispatches due factory jobs every 5 minutes and refreshes active runs.
AI Gateway
The Main’s planner model, and every API-key model behind OpenCodex. OIDC auth.
Neon Postgres
Machines, jobs and runs.
Credits
Pro trial, $100 AI Gateway, $100 v0 and Neon, redeemed by a browser agent.
Plus GOAT (the multi-agent runtime on every machine) and OpenCodex (the model layer).
GOAT Anywhere · Fast Forward · Agents & Factories09 / 11How it meets the track.
- Deployed on VercelLive at goat-anywhere.vercel.app; cloud machines are Vercel Sandboxes.
- Runs async from schedulesFactory jobs: prompt + machine + schedule. Vercel Cron dispatches them with no one watching.
- Useful work without a prompt per stepOne instruction; the Main plans the tasks and routes them; agents on each machine work to completion.
- A person can monitor and steerLive status per task, a one-line routing reason, steer with a follow-up, stop, or open the machine’s GOAT.
Say it once.
Let the Main handle the rest.
- Rebalance running work when a plan nears its quota
- More runtimes: other cloud sandboxes and home servers
- More providers through OpenCodex
- Shared workspaces for small teams
Try it if you juggle more than one machine or subscription, and tell us where the Main routed wrong.
goat-anywhere.vercel.app · github.com/lidge-jun/opencodex