GitHub pages
November 2026
mAGI is a Fedora/Hyprland setup where the main interface is one LLM agent in tmux. Looking back at the repos, most of what I built over the last four years points at it. This post goes through them in order.
At Hear Us I built Joy, a conversational bot for wellbeing check-ins. It ran on Slack, then on WhatsApp: a state machine on top of Dialogflow, with Firestore for events and conversation memory. This was before ChatGPT. It was the first thing I built that people talked to instead of clicking through.
The same year I started keeping my dotfiles on GitHub. Neovim and the terminal. Everything below lives there.
At Mavex AI the product was Mavy, an executive assistant. The user talked to one thing; behind it a router sent each request to one of fourteen specialists: calendar, email, tasks, search and so on. Function calling was barely out, so the router was GPT-4 forced to return one of two JSON shapes: answer directly, or hand off. I wrote about it here. The search service behind it picked its retrieval cutoff by clustering similarity scores instead of using a fixed threshold (post).
One assistant in front, specialists behind it. mAGI has the same shape.
I pulled the agent loop out into AgentZero, a minimal tool-calling agent. Its README has a wishlist. Two items:
Then PlanWeaver, a plan-and-execute agent, and ChainFactory: a small DSL for chaining LLM calls, with parallel steps and typed outputs, published on PyPI as chainfactory-py. The spec drafts are on this site.
daily-notes.nvim is a daily notes plugin for Neovim: one file per day, year/month folders, headers carried over from the previous day. doctr was a daemon that read git diffs and kept documentation in sync with an LLM.
Neither is an agent. Both are about not losing track of things, which is most of what mAGI’s ledger and notes do now.
Most of the agent work is from this year.
By bridge, the pieces were all there: one agent in front, workers in tmux windows, voice, a browser agents can drive. They were spread across repos and each needed setting up.
mAGI puts them in one install. You log in and the orchestrator is already running in tmux. You speak or type to it. It starts other agents, operates the desktop through hyprctl, and asks you only for decisions.
Some parts map straight back:
The pattern I see: I keep building one thing to talk to, with other things doing the work behind it, and something that remembers what’s still open. It moved from a chat app, to a product backend, to my own desktop.
| When | Project | What it was |
|---|---|---|
| 2022 | Joy (Hear Us) | Wellbeing chatbot on Slack and WhatsApp, with conversation memory |
| 2022 | dotfiles | Neovim and terminal config |
| 2023 | Mavy (Mavex AI) | One assistant, fourteen specialists behind a router |
| Jan 2024 | AgentZero | Minimal tool-calling agent |
| Feb 2024 | PlanWeaver | Plan-and-execute agent |
| Jul 2024 | ChainFactory | DSL for LLM pipelines, on PyPI |
| Jun 2025 | daily-notes.nvim | Daily notes in Neovim |
| Aug 2025 | doctr | Docs daemon driven by git diffs |
| Jan 2026 | multiclaude | Parallel Claude Code agents in tmux |
| Mar 2026 | FlowPilot | Browser automation with human handoff over CDP |
| May 2026 | trie | Index of meaning and intent for a codebase |
| May 2026 | Soren | Agent swarm in tmux with mailbox and journal |
| Nov 2026 | bridge | Talk to one Claude Code session; it runs the rest |
| Nov 2026 | mAGI | A Fedora/Hyprland setup built around LLM agents |
mAGI is on GitHub at pankajgarkoti/magi; a star helps other people find it.