My AI stack for running 100+ projects

Summary

The speaker, a CTO who has worked across more than 100 projects, describes a personal AI-agent setup for coordinating work across many projects. The system is designed to keep agents running continuously, provide remote access, support different model providers, retain shared memory, connect safely to work tools, and make it quick to start tasks.

Six Requirements and the Setup

1. Keep Agents Running

The speaker runs agents on a dedicated Linux PC rather than a laptop, so work can continue when the laptop is closed or offline. He repurposed an old gaming PC, installed Arch Linux, and connects to it remotely. He suggests reusing an old computer rather than automatically renting a VPS.

2. Access the Machines Remotely

He uses Tailscale to create a private network between his computers and phone, allowing them to connect even when they’re on different networks. Tailscale links can also be used to preview work produced by an agent.

3. Switch Between Models and Providers

T3 Code is the control interface he uses to manage agent environments such as Claude Code, Codex, and OpenCode. He can switch agents or use different ones within the same project.

T3 Code also offers a mobile client and a connectivity feature that the speaker describes as an alternative to setting up Tailscale separately when using a small number of devices.

4. Use Shared, Portable Memory

Instead of relying on each coding agent’s separate memory system, he uses a wiki based on Andrej Karpathy’s LLM Wiki approach: linked Markdown files that can also be viewed in Obsidian.

Agent instruction files tell agents to check the wiki for relevant context and add notes about important events, project starts, or decisions to an inbox. A periodic script incorporates those notes into the wiki.

5. Connect Agents to Tools with Appropriate Permissions

Agents can connect to services such as Jira, email, GitHub, GitLab, and task managers through APIs or MCP integrations.

He recommends using limited-scope API keys where possible. For older systems that only offer broad access, he describes creating a local MCP proxy that exposes only selected operations to agents.

For databases, he recommends restricted users with access only to the necessary tables, while avoiding sensitive data such as personally identifiable information.

6. Start Work Quickly Through a Central HQ

The speaker keeps an HQ project on each of his machines. It contains a list of projects and where they are installed.

He can tell HQ what needs doing; it locates the project and starts a thread on the machine where that project resides, provided the machine is online. This avoids manually finding projects and creating a new session in the right environment.

Starting Tasks from a Phone

An always-on Hermes Agent runs on the Linux PC and can access the MCP integrations, wiki, and project list. The speaker messages it through Element, using a self-hosted Matrix setup.

A phone shortcut records a voice instruction and sends it to the PC, where Whisper transcribes it. Hermes then creates a new T3 thread rather than immediately modifying the project.

This lets him capture tasks while away from his computer and review or continue the resulting threads later.

Overall Arrangement

The setup spans a MacBook Pro, Mac mini, Linux PC, and phone. T3 servers run across the computers, with clients on selected devices. HQ and the wiki are kept on each machine and synchronized with a command called fleet.

The speaker’s stated goal is to replace much of his to-do-list workflow with a stream of task threads that he can review and resolve.

Main Speaker and Sources

• Main speaker: Nimrod, presenting his own workflow and setup.
• Tools and projects discussed: Arch Linux, Tailscale, T3 Code, Claude Code, Codex, OpenCode, Andrej Karpathy’s LLM Wiki, Obsidian, Hermes Agent, Element/Matrix, Whisper, MCP integrations, and the speaker’s HQ project.