NEW

SurrealDB 4 is here, introducing SurrealDB Earth and SurrealDS!

Learn more

GTA 6 easter eggs: a live tracker built on SurrealDB

Community
Tutorial

Nov 21, 20267 min read

Martin Schaer

Martin Schaer

Show all postsGTA 6 easter eggs: a live tracker built on SurrealDB

Our newsletter

Get tutorials, AI agent recipes, webinars, and early product updates in your inbox every two weeks

GTA 6 players have already found a literal Easter egg: a painted egg in a bird's nest on a tower in Leonida, and throwing it triggers a cutscene (IBTimes, 19 November 2026). The other finds are scattered across Reddit, X and Discord, so I built a live tracker on one SurrealDB database where you can add yours.

Open the egg tracker

Press "Report a find", click the map where you found it, then describe it and say which older game or show it nods to. If someone has already reported something close by, the tracker shows it first, so you can confirm theirs instead of adding a duplicate. You can attach a screenshot too: your browser shrinks it before it is sent. A moderator checks every new find, and until then only you can see it.

The tracker is a small tour of what one database can do:

An egg is a place, a picture and a reference to something else. Most databases make you pick one of those. SurrealDB stores all three on the same record:

schema.surql (abridged)
DEFINE TABLE egg SCHEMAFULL;
DEFINE FIELD name ON egg TYPE string;
DEFINE FIELD at   ON egg TYPE geometry<point>;          -- where on the map
DEFINE FIELD image ON egg TYPE option<bytes>;           -- the screenshot, shrunk in the browser
DEFINE FIELD shot ON egg TYPE option<array<float>>;     -- its embedding, written by a worker
DEFINE INDEX egg_shot ON egg FIELDS shot HNSW DIMENSION 512 DIST COSINE;

-- egg ->references-> work   (GTA IV, GTA V, Dexter, ...)
DEFINE TABLE references TYPE RELATION IN egg OUT work;

Positions use the grid from the open community map of Leonida: 0 to 2000 on each axis. The starting data is a handful of launch-week finds reported by GameSpot and IBTimes, with approximate positions.

Every query below was run on SurrealDB v3.3.5 and v4.0.1 on 20 November 2026.

You've got a screenshot of a mural, a sign or a strange statue, and no idea where it was. Turn it into an embedding with CLIP, then ask for the nearest screenshots, but only inside the part of the map you're searching:

LET $vice_city = { type: 'Polygon', coordinates: [[[1150, 650], [1700, 650], [1700, 1150], [1150, 1150], [1150, 650]]] };

SELECT name, hint, at, vector::distance::knn() AS dist
FROM egg
WHERE shot <|3,40|> $shot AND at INSIDE $vice_city
ORDER BY dist;

The polygon filter sits in the same WHERE as the vector search, so it's applied during the index walk rather than after it. Filtering after the walk can leave you with nothing (I measured it here).

CLIP puts images and text in the same space, so $shot can also come from a sentence. "A pink flamingo mural" finds the screenshot of a pink flamingo mural.

Your browser shrinks the screenshot before it leaves your computer. In my test, a 1.9 MB photo of 4000×3000 pixels became a 67 KB WebP, and the re-encoding drops photo details such as GPS. The image goes to SurrealDB as bytes. A small Python worker on my computer, embed.py, finds the images that have no embedding yet, runs CLIP and writes the 512 numbers to shot.

"Search by screenshot" takes the same route through a short-lived search_request record, so a search waits for the worker's next run.

The same script embeds by hand, which is how I made this post's demo data:

uv run embed.py --image screenshot.png
uv run embed.py --text "a pink flamingo mural"

You're standing at the foot of the radio tower. What else has been found within 100 map units?

LET $here = [1280, 950];
SELECT name, math::round(vector::distance::euclidean(at.coordinates, $here)) AS away
FROM egg
WHERE vector::distance::euclidean(at.coordinates, $here) < 100
ORDER BY away;
Result
[
  { away: 14, name: 'The literal Easter egg' },
  { away: 54, name: "Dexter's apartment" },
  { away: 81, name: 'Wade on the Final Act podcast' }
]

The map grid is flat, so this uses straight-line distance. geo::distance would assume latitude and longitude and give you metres on a sphere. The geospatial docs cover both.

Easter eggs are mostly callbacks: to GTA V's Mount Chiliad mystery, to GTA Online's jetpack, to a TV serial killer's flat in Miami. A callback is an edge, so the question is a graph query. Run it here:

Turn the arrow around and you get every nod to one game. The RELATE statement creates the edges:

SELECT VALUE name FROM work:gta_v<-references<-egg;
Result
['Wade on the Final Act podcast', 'The bunker under the mountain']

The tracker talks straight to SurrealDB from your browser. Each visitor signs up anonymously: the browser makes up a random secret and keeps it. The rules live in the schema, so they apply whichever client connects:

DEFINE TABLE egg SCHEMAFULL
  PERMISSIONS
    FOR select WHERE status = 'approved' OR found_by = $auth.id
    FOR create WHERE $access = 'hunter'
    FOR update, delete NONE;

DEFINE FIELD status ON egg TYPE string
  VALUE IF $access = 'hunter' THEN 'pending' ELSE $value ?? 'pending' END;

-- The screenshot: 200 KB at most, and only its finder can read it back
DEFINE FIELD image ON egg TYPE option<bytes>
  ASSERT $value = NONE OR bytes::len($value) <= 204800
  PERMISSIONS FOR select WHERE found_by = $auth.id;

-- One "I found it too" per hunter per egg
DEFINE INDEX confirmed_pair ON confirmed FIELDS in, out UNIQUE;

A new find is visible only to the person who reported it until I approve it. Sending status: 'approved' from the browser console gets you a pending egg. The real schema also checks that an upload is a JPEG or WebP, and visitors cannot write the embedding, so a find always matches the picture a moderator approved. The counter at the top of the tracker updates from a LIVE SELECT, and every refresh goes through the same permissions. For the thinking behind this, read even a jailbroken LLM can't exceed its database permissions.

A Rockstar studio lead said in August 2026 that GTA 6 takes NPC interactions "much further" than Red Dead Redemption 2. A hint like the reported one about a bunker under a mountain is two strangers in a bar, said once, in passing. If your captions go into SurrealDB Agent Memory as you play, an agent can find that line days later:

from surrealdb.memory import Memory

with Memory(context="leonida", endpoint=ENDPOINT, api_key=KEY) as memory:
    memory.remember(
        "NPC in a bar: there's a bunker under the mountain, full of jetpacks.",
        labels=["source=captions", "place=545,1465"],
    )

Days later:

You: Someone mentioned a bunker. Where was I?

Agent: On Tuesday evening, two NPCs in a bar said there's a bunker under the mountain, full of jetpacks. You were at 545, 1465 on the map. Want me to add it to the tracker as a lead?

(The conversation above is illustrative.)

Agent Memory pulls entities and relations out of the text, such as the bunker, the mountain and the jetpacks. So "where did I hear about jetpacks?" finds the same line.

create-surrealdb-poc scaffolds a SurrealDB proof of concept in TypeScript or Python, with the SurrealDB agent skills already installed. It asks for a project name and a language, and flags skip the questions:

npm create @martinschaer/surrealdb-poc
# or: bun create @martinschaer/surrealdb-poc
# without the questions:
npm create @martinschaer/surrealdb-poc my-poc -- --lang py --install

The new folder has:

  • The skills from surrealdb/agent-skills: surrealql, surrealql-functions, surrealql-performance, surrealdb-cli and surrealdb-docs, plus surrealdb-js or surrealdb-python for the language you picked. They live in .agents/skills/, are linked into .claude/skills/ for Claude Code, and skills-lock.json records where they came from.

  • A starter that connects, signs in and runs a query. TypeScript needs Node 22.18+ or Bun, and Python uses uv.

  • A Justfile and a .env.example for a throwaway local server.

  • A README.md that also covers SurrealDB Studio and deploying to SurrealDB Earth.

Four commands later you have a result:

cp .env.example .env
just install
just db     # an in-memory database, in one terminal
just run    # in another: prints "hello from SurrealDB"

Then open the folder in an agent that loads skills, such as Claude Code, and ask it to interview you and write a spec for your POC. The skills tell it how SurrealQL really works, and the in-memory database lets it test its own schema and queries instead of guessing. When the POC needs to leave your machine, one command creates an Earth instance (after surreal install cloud and surreal cloud login):

surreal cloud instance create my-poc --region earth

Yes. Open the tracker, press "Report a find", click the map and describe it. A moderator checks every find before other players see it.

No. Your browser shrinks the image and sends it, and a worker turns it into an embedding. You only run embed.py yourself if you want to reproduce the queries in this post.

One SurrealDB instance, queried straight from the browser, with a small Python worker for the embeddings. The queries in this post work on SurrealDB v3.3.5 and v4.0.1.

Run npm create @martinschaer/surrealdb-poc and pick TypeScript or Python. It sets up the SurrealDB agent skills and a starter that connects to a local database.

No. It is an unofficial fan project built to show what SurrealDB can do.

Found something Rockstar hid? Add it to the tracker, and share this post with the friend who still hasn't found the egg.

Add your find

Here are the schema, seed data, queries and the worker files.

GTA VI and Grand Theft Auto are trademarks of Take-Two Interactive. This post isn't affiliated with or endorsed by Rockstar Games.

Our newsletter

Get tutorials, AI agent recipes, webinars, and early product updates in your inbox every two weeks