Personal tool · AI assistant

Every project at a glance. And a voice that answers.

A brain at the center, nerve fibers to every project: Mission Control shows what’s open, where AI agents such as Claude Code or Codex are working right now and what today holds. I ask by voice and hear the answer. Agents review tasks in the background or carry them out in a visible session; what they’re allowed to do depends on the job. Calendar and important emails are built in, and on the go the same assistant answers in a Telegram bot. Built for my own work, as an example of how an AI tool with its own design can fit the services you already use.

Scope: Concept, design and development
Stack: Node.js, three.js, Whisper, ElevenLabs, Claude Code
Use: Private, server on my own Mac

A working day, asked and answered.

Every picture comes from the demo mode: invented projects, appointments, emails and reports, so no real data is shown. The interface and the flows are the same as in the real Mission Control.

Mission Control with the Shop-Relaunch project card open and the answer to “What’s on today?” as a caption above the “Frag Mission Control” button — Questions by voiceOne click, one question. Speech recognition (Whisper) transcribes it locally on the Mac, Claude answers sentence by sentence, an ElevenLabs voice reads it out, and the camera flies to the project in question.

Claude’s report on the task “Add alt text to product images” with summary, findings, suggestion and next step — Reports from agentsClaude Code or Codex review a task in the background and file their report on the project’s card: the summary on top, then findings, suggestion and next step.

Overview with the brain at the center, the projects around it and the calendar for the coming days opened — Calendar and important emailsTop right: the next few days from the Mac’s calendar and the emails that need something from me. The assistant knows both, and today’s plan works around the appointments.

Deep zoom into the neural tissue: glass cell bodies and glowing fibers in the projects’ colors — An endless zoomScrolling leads through the glass shell into the neural tissue and into a cell whose nucleus holds the brain again. The jump back to the start stays invisible.

On the go, right in Telegram.

Questions by voice messageThe voice message is understood just as on the Mac; the answer comes back as a voice message in the same voice, with the text underneath.
Checks in on its ownEvery morning the day’s appointments and most important items arrive, even when the Mac is off.
The same assistant as on the MacCheck the status, take notes, tick things off, plan the day or have an agent review something; its report comes back to the chat as a summary. Answers arrive while the Mac is running.

Telegram chat with the Mission Control bot: an update for today, a question as a voice message and the spoken answer with its text underneath

Every agent gets exactly the access the job needs.

Permissions per jobWhat an agent may do follows from the kind of job. Further levels, such as changes on a separate working copy, can be added.
Reviews in the backgroundFor reviews, Claude Code and Codex get read-only access, enforced by the tools themselves, not just asked for in the prompt. Tested on a throwaway folder: neither could create a file.
Changes in a visible sessionWhen something needs to change, Mission Control starts a Claude Code or Codex session in the project folder. It states its plan first and asks before it changes anything.
Limits in the backgroundAt most three reviews at once, each with a 15-minute time limit.
List changes can be undoneBefore the assistant checks off, adds or edits anything, it saves a backup; “undo” restores the previous state, on the Mac and in the Telegram bot.
Reachable from this Mac onlyThe server only accepts requests from the machine itself. The Telegram bot fetches messages on its own, opens no port to the outside and answers a single chat.

Under the surface.

2,300cell bodies in the neural tissue
3.6 msrender time per frame at 2880 × 1800 pixels
0.35 sspeech recognition per question, locally with Whisper
0npm dependencies, no build step

Custom shadersThe brain is dark glass; the veins in its folds glow in each project’s color and are computed per pixel, so they stay sharp at every zoom level.
Answers sentence by sentenceClaude writes the answer as a stream. Every finished sentence goes straight to the voice while the rest is still being written.
Reads what’s already thereGit, Markdown notes and the Claude Code and Codex sessions, read-only. A new project is one entry in a file.
Calendar without a passwordA small Swift program reads appointments straight from the macOS Calendar app through EventKit.
Fast or thoroughQuestions about the current state go to a fast model; requests and changes to the list go to a more thorough one.
Routines in the cloudThe morning briefing and the email triage run as Claude routines on a schedule; clear appointment confirmations go straight into the calendar.

Node.js with no build step, three.js with custom shaders, whisper.cpp, ElevenLabs with a macOS voice as fallback, Claude Code and Codex, Swift with EventKit, Telegram Bot API, Claude routines. The server runs on my own Mac with no rented hosting; voice, AI models and routines come from cloud services.

Want an AI tool built around your workflow, with permissions you decide?

Tell me about your project