A $20 AI Dev Setup Built on Pi, Herdr, and AXI Instead of MCP

A 23-year-old AI engineer at a big tech company walks through his personal AI development stack, and the whole thing runs on a $20 ChatGPT subscription. The core idea is that a customizable open-source agent harness plus agent-ergonomic CLI wrappers can replace both expensive API bills and MCP servers. He uses Pi as the harness, ChatGPT as the model provider, Herdr to manage multiple agents, and WezTerm as the terminal that holds it all together.

The Four-Part Stack

The setup has four layers. Pi is the agent harness, the open-source coding agent from Mario Zechner's pi-mono project. The model comes from a standard $20 ChatGPT subscription rather than pay-per-token API access; the status bar in his terminal shows a gpt-5.6-sol model with a 272k context window. Herdr sits above the agents as a multiplexer that manages several of them at once. Everything runs inside WezTerm. The pitch is that none of these pieces require enterprise tooling or a big API budget.

Why Pi: Everything Is Swappable

He picks Pi for customizability. On launch, Pi prints its loaded configuration in four blocks: Context (APPEND_SYSTEM.md and AGENTS.md files), Skills (axi, bro, chrome-devtools-axi, exa-axi, gh-axi, herdr, no-mistakes, unslop), Extensions (pi-subagents, chafa-image-preview, pi-transcribe, grok-build-ui, pi-cloak, save-md, pi-openai-compaction), and Themes (a custom grok-build theme). He calls out voice input specifically: the pi-transcribe extension gives him free native transcription, which he demos by speaking a prompt that lands in the input box as text. He has also themed the interface with thinking animations similar to Claude Code.

A Terminal-Centric Workflow

The rest of his life runs in terminal tabs next to Pi: NeoVim for editing, a TUI called nit that tracks daily calories and protein with 14-day charts, and a Spotify TUI for music. He prefers this because everything is keyboard-driven, multiplexed, and switchable between workspaces without touching a mouse. His NeoVim file tree shows dotfolders from a long list of coding agents he has tried (.claude, .codex, .cursor, .copilot, .continue, and more), which frames this stack as the one that stuck.

AXI: Replacing MCP with Token-Efficient CLIs

The most substantive segment covers AXI (Agent eXperience Interface), a framework by Kun Chen, whom the video credits as a principal engineer at Meta. AXI's position is that instead of connecting agents to MCP servers, you wrap existing capabilities in CLIs designed for agents. Instead of the plain GitHub CLI, the agent uses gh-axi. Instead of an Exa MCP server, it uses exa-axi. The wrappers return output in TOON (Token-Oriented Object Notation), a format that benchmarks at roughly 40 percent fewer tokens than equivalent JSON. He demos this live: a voice prompt asking Pi to search for Kun Chen's GitHub triggers the exa-axi skill and returns TOON-formatted results. His other skills follow the same pattern, including chrome-devtools-axi for browser control, unslop for cutting AI-sounding phrasing from responses, and a skill that teaches the agent how to build new AXI wrappers.

Skill Toggling as Context Hygiene

His favorite extension is Pi Skill Toggle, built by Dillon Mulroy, a Cloudflare engineer whose dotfiles include a home/.pi/agent setup. Running /toggle-skills opens a picker that marks each skill as agent-invocable or manual-only. In his config, chrome-devtools-axi, exa-axi, gh-axi, herdr, and unslop are agent-invocable, while bro and no-mistakes are manual-only. The reasoning is context management: if every skill is visible to the agent, you clog its context window. The agent should only see the skills you want it to use autonomously, and the rest stay available on demand.

Key Takeaways

• A $20 ChatGPT subscription can power a full agent workflow when paired with an open-source harness like Pi, avoiding per-token API costs.
• Pi exposes context files, skills, extensions, and themes as user-editable configuration, which is the main reason to pick it over closed harnesses.
• AXI replaces MCP servers with agent-designed CLI wrappers (gh-axi, exa-axi, chrome-devtools-axi) that return TOON output for large token savings.
• TOON formatting cuts roughly 40 percent of tokens versus equivalent JSON in published benchmarks.
• Not every skill should be agent-invocable. Toggling skills to manual-only keeps the agent's context lean and its autonomous choices predictable.
• Herdr adds agent-aware multiplexing (blocked, working, done, idle states) that tmux cannot provide.

Resources

• pi-mono -- Mario Zechner's AI agent toolkit: unified LLM API, agent loop, TUI, and the Pi coding agent CLI.
• Herdr -- Rust terminal multiplexer built for AI coding agents, with automatic agent state detection.
• AXI -- Kun Chen's design principles for agent-ergonomic CLIs, benchmarked against MCP and plain CLI.
• gh-axi -- GitHub CLI for agents built on AXI principles.
• chrome-devtools-axi -- Agent-ergonomic browser automation from the same AXI family.
• TOON -- Token-Oriented Object Notation, a compact JSON alternative with about 40 percent token savings in benchmarks.
• WezTerm -- Cross-platform terminal emulator and multiplexer written in Rust.
• Dillon Mulroy on GitHub -- Author of the Pi Skill Toggle extension and related Pi tooling.