Skip to content
es
All work

AI HQ

A personal command center for AI work. It shows live Claude Code sessions, spend, and a local GPU that Claude can hand work to.

Role
Personal project · design and build
When
2026
Repo
Private repository
AI HQ dashboard in demo mode: the live constellation of two Claude Code sessions, their subagents and files, the local GPU with its runners, plus the sessions and project panels.
local: every service binds 127.0.0.1
100%
MCP tools exposed to Claude Code
10+
platforms: WSL2, macOS, Linux
3

The problem

I run several Claude Code sessions a day, each with its own subagents, and I also have an RTX 4090 sitting next to them. I had no single place to see what the agents were doing or what they cost, and no way to let them hand cheap work to the GPU instead of spending plan usage.

What I built

AI HQ is one dashboard, designed for an ultrawide screen, that shows all AI activity on a machine:

  • Live sessions and agents. A FastAPI server tails Claude Code’s JSONL logs into SQLite and streams changes over a WebSocket. The UI shows sessions, subagents and tool calls as they happen.
  • The constellation. A canvas graph in which sessions are hubs, subagents orbit them, and files light up when they’re touched. You can replay any day at up to 300× speed.
  • Spend and plan usage. Spend at API list prices, plus plan-window gauges that project whether the current pace stays within the limit.
  • A local GPU gateway. NVML telemetry, a job queue with a memory scheduler, and an OpenAI-compatible router in front of Ollama, ComfyUI and voice runners.

Hybrid mode

An MCP server gives every Claude Code project tools such as local_llm, transcribe, speak and generate_image. Claude hands simple drafting, extraction or media work to the local model, and the dashboard shows who sent what and what it would have cost in the cloud. Some of this portfolio’s Spanish copy was drafted that way, then reviewed.

Engineering choices

  • Local-only by design. Services refuse non-local hosts, keys stay in .env, and prompts never reach the dashboard.
  • Runs on every platform. Docker Compose with NVIDIA on WSL2 and Linux, a native virtualenv with LM Studio on Apple Silicon. A feature one platform lacks degrades cleanly elsewhere.
  • Verified like a product. pytest suites in containers, a typecheck, vitest, and Playwright screenshots at ultrawide, desktop and phone widths.