Vox Manifestor

Vox Manifestor is AR Labs' Android app that a person operates almost entirely by speaking. An AI character, the "genie", helps them choose up to five wishes, talk each one through, and describe both their present situation and the outcome they want. It was our first hand-built conversational agent: the place where speech recognition, speech synthesis, a language model and a conversation state machine were first made to work together on a phone. This page tells the app's build history. From 2025 the same app grew a daily voice check-in and is presented as Coherence, whose page describes that conversation design.

What it does

The main screen holds five wish slots. Spoken commands drive it: "manifest" starts a new wish, "select" picks one, "define" opens the concept screen for the selected wish, and "affirm" starts an affirmation session. Wishes can be added, edited, swapped and deleted by voice or by touch.

The concept screen is where the conversation happens. For the selected wish the genie explains the idea of a present state and a desired state, asks questions, and fills in up to five items on each side from the person's answers. The affirmation tool generates personal affirmations from a wish and reads them aloud. A status panel shows what the speech recogniser heard and what the agent is doing, and conversation history is kept and grouped by day.

sequenceDiagram
  participant U as Person
  participant S as Speech-to-text
  participant A as Conversation agent
  participant B as Language model
  participant T as Text-to-speech
  U->>S: speaks
  S->>A: transcript
  A->>A: command or conversation?
  A->>B: prompt with current state
  B-->>A: reply and next actions
  A->>T: genie's words
  T-->>U: speaks back

How it was built

The app is written in Kotlin with Jetpack Compose and stores its data in a Room database. Google Cloud Speech-to-Text and Text-to-Speech handle the voice, and Google's Gemini model is the "brain" behind the genie. It works with a Bluetooth microphone, so a person can use it through earphones.

Development ran in short versions from October 2024, and the changelog records the path:

• A reflex agent first. The earliest versions recognised fixed commands and moved through hard-coded state pathways. By December 2024 the agent could create, change and delete wishes by voice reliably.
• Then a brain. At the end of 2024 a language model was added behind the concept screen, with a question-and-answer loop that fills in present and desired states.
• Then planning. In February 2025 the agent began keeping a rolling list of its next three intended actions. Planning a few steps ahead let it do one thing per turn instead of trying to cover everything in a single reply.
• Then the core loop. From April 2025 the app gained a daily check-in conversation run through chained prompts, and later the extraction of recurring themes from what the person said. That line of work is described on the Coherence page, and the essay CORE-LOOP: Managing Complex Two-Way Voice Conversation via Agent / LLM, written at the time, sets out the conversation-loop design.

Most of the engineering effort went into making a spoken conversation behave. Each of these was a real problem with a specific fix:

• The app hearing itself. The microphone could pick up the genie's own voice and treat it as input. The fix was explicit coordination between speaking and listening, and later dedicated echo prevention.
• Commands inside sentences. A word like "exit" in the middle of a sentence should not end the conversation, so command parsing takes context into account instead of acting on keywords.
• Homophones. "To" and "two" sound the same, and wish slots are numbered, so an interpretation layer resolves common homophones.
• Slow turn-taking. Waiting for a long silence before processing speech felt sluggish, so the screen gained a "send now" control and an interrupt button.
• Duplicate agents. Leaving and re-entering a screen could start a second conversation alongside the first. Conversation scope is now managed explicitly, with separate commands for pausing and for ending.

The idea that carried forward from this period is that the language model's job is to extract structure from what a person says, and the app's job is to hold that structure, so that the conversation's state lives in code rather than in the model's memory.

What exists today

The app reached version 2.22 in November 2025, with a Material 3 dark theme and a terminal-style display of the agent's activity. The public documentation site describes the genie-era features and the plans of the time, such as brainstorming wishes, milestones and deeper goal hierarchies. The app's source code is private.

In 2026 work began on a redesign around a daily ritual. A person earns energy through intentional actions, turns it into lasting crystals, loads those into a goal, and then shapes the goal into a symbol, a sigil, through a guided AI conversation. The redesign was built in vertical slices, each a thin piece of working behaviour from screen to database, through our coding pipeline, the route every code change takes: spec, test-first build, second-model review and a person's merge. Development has been paused since July 2026.