A Full Gamified Treasure-Hunt Game live in 2 Days.

This is the record of Tokyo Mystery, nomarz’s own product. Since launch, AI runs its operations and social media too. Every number is measured from the git history and the AI session logs.

Open tokyomystery.com

  • 36 hours

    From first instruction to live

  • 125

    AI agents started

  • 430 K lines

    Of code and tests

  • 1,600 dollars

    AI usage at API list price

From hour 0 to hour 64

  1. Oct 4, 0:26 0 h

    The spec is handed over

  2. Oct 4, 22:19 22 h

    First commit

  3. Oct 5, 12:34 36 h

    First live release

  4. Oct 6, 16:14 64 h

    Last build

  5. Oct 8

    Own domain

A treasure hunt you solve by walking the streets of Tokyo

  • A 3D map of central Tokyo seen at an angle, with buildings coloured by height.
    Tokyo in 3D
  • Today’s map with the Edo coastline and landmarks laid over it, and a panel for choosing the era.
    Edo laid over today’s city
  • The scroll of a map called “Lights of Shinjuku”, with a treasure chest and its difficulty, steps and finishers.
    A treasure scroll for each map
  • A list of daily quests, some finished and some in progress.
    Daily quests

The screens are from the test environment.

People decided. AI built it in parallel.

  1. Decide

    A person writes one spec.

  2. Divide

    AI lays the base and sets each area.

  3. Build

    Agents build and test, each in its own area.

  • Lead
  • Build and research 62
  • Knowledge keeper 53
  • Shared work 5
  • Advice (design and review) 3
  • Back end 2

AI agent

About 430,000 lines of code and about 4,000 tests

  • Go 178,065 lines
  • TypeScript 127,107 lines
  • Svelte 101,968 lines
  • Other 26,148 lines
  • 1,212 Go tests
  • 2,750 Front-end tests (approx.)
  • 16 End-to-end test files
  • 43 Database migrations
  • 17 API documents

Every agent worked from the same memory

  • 542 recalls

    Memory

    Short facts and decisions, each with its evidence.

  • 221 searches

    Documents

    The right passage from specs and procedures.

  1. Read

    Look up memory and documents before starting.

  2. Build

    Check what was found against the code.

  3. Report

    Attach what was learned to the report.

  4. Write back

    The knowledge keeper updates the record.

Since launch, a bot runs it

Post cards flying out of a phone into the night sky over Tokyo.

Grok AI agent

Social media

  1. Receives the assets

  2. Asks a person for the rules

    Person

  3. Makes and publishes posts

Admin console

  1. Drafts a map

  2. Publish check

  3. Test play

  4. Owner approves

    Person

  5. Live

Assets handed to the bot

  • 162 screens
  • 89 artworks
  • 21 logos
  • 5 videos

Can

  • Draft maps, quests and themes
  • Support players
  • Edit prizes and partner shops

Asks a person first

  • Publishing and scheduling
  • Grants and codes
  • Revealing riddle answers
  • Deleting

Cannot

  • Orders, refunds, ledger
  • Admins and settings
  • Audit log

Results so far

［What the bot has actually done: number of posts, map drafts created, and so on］

The AI usage came to about $1,600

About 1,500–1,700 dollars

  • Re-reading context already loaded 1,031 USD
  • Writing context 369 USD
  • What the AI wrote 86 USD
  • Extra if written context is kept longer +221 USD

This is the three-day build at API list prices. What you actually pay depends on your plan. Server and domain costs and people’s time are not included.

Where it stands

  • Live

    Live at tokyomystery.com (checked October 12, 2026).

  • Payments and maps

    As of October 8, payments ran in test mode and the one map was a test map.

  • The bot’s guides

    Written from the documents and code; as of October 8 they had not been checked by clicking through the screens.

  • Analytics

    As of October 8, the analytics store was not yet on the server.

How we measured
Time
The first instruction is the first record in the AI session logs (Oct 4, 0:26). The first release is the first release run (Oct 5, 12:34). The end is the last commit of Oct 6 (16:14).
Lines
Lines in the Go, TypeScript, Svelte, SQL, CSS and JavaScript files tracked by git, tests included, Markdown excluded. About 430,000.
Tests
Go: the number of functions whose name starts with Test (1,212). Front end: the number of it and test calls across 256 test files (about 2,750), an estimate.
Agents
From the session logs: 125 agent starts and 77 follow-up instructions. Knowledge base: 542 memory recalls and 221 document searches.
Cost
The token counts of 20,379 responses in the logs, multiplied by API list prices (per million tokens: Opus 5.5 $4 input and $20 output, Sonnet 5.5 $2 and $10, cache reads $0.20). Cache writes are given as a range, at 1.25× and 2× the input price.
Assets
The assets handed to the bot are the file counts of the marketing asset folder (162 screens, 89 artworks, 21 logos, 5 videos).

Zero to one, then run it all with agents

  1. Decide

    A person sets what to build and where approval is needed.

  2. Build

    Agents build in parallel.

  3. Launch

    A person approves the last call.

  4. Operate

    A bot runs it every day, within its permissions.

  5. Learn

    What is learned is kept for next time.

Tell us how your work flows. We’ll show you where AI saves the most time.

Free AI audit (30 min)