← Steve Govoni

Go7 Workhorse: All My AI Subscriptions, Working Together

Role: Founder, builder, and lead, Go7 Studio Dates: Started a few months ago; ongoing Links: https://go7workhorse.com | https://github.com/go7studio/Go7-Workhorse (MIT, Windows and macOS)

Workhorse connects my AI apps. Codex can call Claude or Grok workers, and they can call it back. Every AI subscription I pay for works together on the same job instead of sitting in its own silo.

The problem

I pay for more AI than I can spend. Grok, Claude Code, Codex, Cursor, plus my own local models. Each one was an island. Its own window, its own login, its own meter.

None of them could hand work to another. If Codex was the right driver but Claude was better at one piece, I was the glue. I copied context between tools, guessed which model fit the task, and lost track of what was left on each plan.

I wanted them to work as one team: any agent able to call the others, each used where it is strongest.

What I built

Go7 Workhorse is a native desktop app that turns separate AI subscriptions into one working crew.

It runs on Windows and macOS. I use it every day on my Mac. It is open source under MIT, and it is a Go7 Studio product that I built and lead.

Architecture

Multi-agent orchestration. Electron, React 19, and TypeScript. Agents connect over the Agent Client Protocol. The orchestrator ranks worker spawns by domain benchmark and cost, so each task goes to the best-fit, cheapest capable model. Missions have cost caps. A judge scores the mission report. Usage is metered per vendor. Workers run in isolated git worktrees with sandbox profiles. Schedules, goals, plans, and subagents are journaled and survive a restart.

Local inference on DGX Spark. The local side runs on an NVIDIA DGX Spark (Grace Blackwell, 128 GB unified memory). I built a Python GPU job broker with epoch-fenced leases, unified-memory admission control, a SQLite job journal, SHA-256 artifact checks, and staged canary rollout with automatic rollback. A model router drains in-flight work and hot-swaps weights behind an OpenAI-compatible gateway. Multiple local models run there, including a 27B-class Qwen on SGLang. Because the gateway speaks the OpenAI format, a local model joins the desk like any other worker.

I build Workhorse by directing a supervised fleet of AI coding agents under test-enforced repository and release rules.

Go7 Studio

Go7 Studio also ships games. Live titles: PATHOGENEER, Terraceweave, and Cargo Pop in Godot; Rampart and Outclimb in Unity; Empire Tycoon in Flutter. Gate Wars TD is in development in Godot, and Roblox titles are in development.

Results

Takeaway

The value isn't any one model. It's getting the subscriptions I already pay for to call each other, with cost caps, judge scoring, and each vendor's login kept separate.