The AI interprets.
Code counts.
Code computes ~25 deterministic aggregates. The AI only reads those numbers and returns a typed profile through one forced tool call.
Now in early access for game studios
The intelligence layer that makes
any game understand its players
Plug Locus into an existing game with two HTTP calls. It turns raw gameplay into a living AI profile of every player, and powers playstyle matchmaking, adaptive difficulty, wellbeing alerts and recommendations with zero AI calls on the real-time path.
- 0
- LLM calls on the hot path
- 1–7ms
- real-time reads
- 2
- endpoints to integrate
- 164ms
- to onboard a new genre
Player DNAplayer-00 · via llm
Aggressive Tactician
confidence 32% · 3 matches
Rolling win rate vs flow bandlast 40 matches
win ratebot skill45–55% band
Matchmaking · balanced3 ms · 0 LLM calls
- player-fErratic Gambler0.91
- player-aSharp Opener0.84
- player-dSharp Gambler0.78
Live pipeline SSE connected
Built for every kind of game
- Card & skill gaming
- Shooters
- MOBA & team
- Strategy & board
- Casual & puzzle
- Multi-title studios
Why don't games understand
their players yet?
Games collect mountains of behavioral data and reuse almost none of it. Every studio rebuilds player understanding from scratch, one feature at a time.
| Feature | Built on | Sees | Shared profile |
|---|---|---|---|
| Matchmaking | ELO model | wins / losses | ✕ |
| Difficulty | hand-tuned rules | level progress | ✕ |
| Churn | separate classifier | session counts | ✕ |
| Safety | rules engine | spend | ✕ |
| Next title | nothing yet | no data | ✕ |
✓With Locus, one Player DNA powers all of them
What makes Locus different?
Locus builds one profile per player, the Player DNA. An AI reads your gameplay once per session, off the request path. Every real-time feature then reads that profile with simple math.
The AI interprets.
One vocabulary for every player
Every player in every game is scored 0–100 on the same Core 7 traits, so profiles compare across players and titles.
Precompute, then read
Everything live is arithmetic over 7 numbers. No tokens, no network hop, and it keeps working if the model provider goes down.
01
Your game
manifest + events
02
Ingestion
202 · instant
03
Aggregator
deterministic math
async · 1 AI call / session
04
AI profiler
Core-7 + evidence
05
Player DNA
7-dim vector
real-time · 1–7 ms · 0 LLM calls
Matchmakingvector KNN
Adaptive difficultycontrol loop
Wellbeingauditable rules
Recommendationscosine ranking
Seven traits.
Every player, decoded.
Nobody fills in a survey. Every trait is a by-product of play, read from choices, timing, stakes and behavior after a loss. Pick an archetype to see its DNA.
Patient Grinder
Slow, safe, relentless. Wins by not making mistakes.
What can Locus do
for your game?
Post a manifest, stream events, and every feature below is ready to use. No model to train, host or babysit.
flow band 45–55%
Adaptive difficulty
Keeps each player in the flow band by tuning your own levers to their DNA.
6 auditable rules
Wellbeing & churn signals
Flags loss-chasing, tilt, fatigue and more, each with a suggested intervention.
00Aggressive Tacticianskill 72 · aggression 78
vs
11Patient Grinderskill 74 · aggression 30
Playstyle matchmaking
Lobbies that are fair and fun: close in skill, contrasting in style. Similar, balanced, rival and skill-only modes.
- blitz-ladder0.91
- study-mode0.74
- club-night0.52
Recommendations
Ranks your modes, events and tables against each player's DNA.
- riskyActionRatio 0.50→ risk 62
- baselineDecisionQuality 0.73→ skill 72
- postLossQualityDrop 7%→ composure 51
Evidence behind every score
Each trait names the numbers that produced it, and each flag carries the values that triggered it. Ready for support teams and regulators.
{ "gameId": "rift-clash", "genre": "moba", "actionVocab": […] }200 · onboarded
Onboard with one file
A JSON manifest. No code changes, no retraining.
Survives the model being down
The real-time path has no model dependency. If an AI call fails, profiling falls back to a deterministic profiler and keeps going.
matches3 / 5
coldStart · blended with prior
Cold start, solved
Usable matchmaking from a player's very first session.
dna.updated
Live ops console
Watch profiles form in real time over a streaming feed.
Two endpoints.
That's the whole integration.
No SDK lock-in, no training, no inference infrastructure. If your server can send JSON over HTTP, your game can be AI-enabled this week.
- 01Describe your game Post one JSON manifest: your actions, which are aggressive or risky, and the difficulty levers you can turn. Once, about 15 minutes.
- 02Stream your events Send decisions and outcomes as they happen. Fire-and-forget: it returns 202 immediately and never blocks your game loop.
- 03Read the intelligence Call matchmaking, difficulty, wellbeing or recommendations whenever you need them. Synchronous, 1–7 ms, zero inference.
Any engineUnity, Unreal, Godot, Flutter, web: plain HTTP from your server
Any genreproven on card, strategy, shooter and MOBA manifests
Strict validationmapping bugs are rejected at integration time
Debuggablesee the exact numbers behind any profile
How Locus makes
an impact
Retention isn't one feature. It's every touchpoint feeling right. Locus works the levers that decide whether a player comes back tomorrow.
Keep players in the flow zone
Players who always win get bored; players who always lose quit. Adaptive difficulty steers every player toward a 45–55% win rate, shaped to how fast they think and how they handle losses.
Matches people want to replay
On the same player pool, DNA-balanced lobbies beat skill-only lobbies on predicted engagement 11 times out of 12, with one tie and zero losses.
11 W1 T0 L
Catch players before they burn out
Loss-chasing, tilt and fatigue are flagged in real time with a suggested intervention: a break prompt, a lower-pressure mode or a fair lobby.
Personalization from day one
Recommend the modes, events and tables each player is built for. Cold-start players get sensible defaults from the population prior.
Defensible by design
Every intervention carries its evidence, such as “5 losses in a row while stakes rose +0.31”. That is an audit trail a reviewer or regulator can work with.
Ship AI features in days
One integration replaces four separate pipelines. Your next title reuses the same DNA vocabulary instead of starting cold.
Cost calculator
AI cost scales with sessions, not traffic.
A model on the request path means one inference per matchmaking tick, difficulty check and wellbeing poll. Locus makes one call per session, with a cached prompt prefix shared by every player of the game.
Model calls/day, AI on the request path4,000,000
Model calls/day with Locus100,000
40×fewer model calls
It recovers who players really are.
From behavior alone.
Simulated players were given hidden trait vectors that drove their play. Locus recovered them without ever seeing the hidden vectors.
0.959mean correlation with hidden traits (target 0.7)
93%archetype agreement with ground truth
0.93on a brand-new shooter genre, zero code changes
164 msto onboard that shooter and profile 12 players
Trait recovery
Pearson r · 30 players · 6 archetypes · 4 sessions eachMeasured with the repository's validation harness on a simulated population using the deterministic profiler. Figures show that the pipeline recovers latent player structure; they are not a claim about a specific live player base.
A real game. Real players.
Profiles forming live.
Nine Squares is a hot-seat game integrated using only the two public endpoints. Two people at one keyboard produce two independent Player DNA profiles.
Every score shows its work.
Black-box AI is a liability when real players, and sometimes real money, are involved. Locus attaches evidence to every trait and the triggering numbers to every flag.
- Player-safe archetypes and narratives you can show in-game
- Operator-facing evidence for support agents and auditors
- Profiles move smoothly; one odd session doesn't rewrite a player
- Confidence and sample size on every profile
{
"flags": ["loss_chasing", "fatigue"],
"severity": "high",
"detail": {
"loss_chasing": "5 losses in a row this session while
stakes rose (commitment trend +0.31) and
decision quality fell (-0.22)",
"fatigue": "session has run 104 minutes
without a break"
},
"nudge": "Offer a break prompt and a cool-down
before the next high-stakes table.",
"llmCalls": 0
}One layer. Any game.
Card & skill gaming
Rummy, poker and fantasy: spot loss-chasing and tilt early, route at-risk players to lower stakes, and recommend the right tables.
Shooters
Pair a flanker with an anchor, not two of the same. Tune enemy count and spawn pressure per player, gentler for those who tilt.
MOBA & team games
Balance teams by playstyle, not just MMR. One shared match id, independent profiles for every player in it.
Casual & puzzle
Keep players in the flow band with adaptive hints and timers. Suggest the next mode before they get bored.
Strategy & board
Grade every decision against your engine's evaluation and Locus separates low skill from high skill under tilt.
Multi-title studios
One DNA vocabulary across your portfolio. A new launch starts with shared player understanding instead of a cold start.
Start free.
Scale with sessions.
Real-time reads are never metered, because they cost nothing to serve.
Sandbox
Free
For prototyping and game jams
- Full API & live console
- Deterministic profiler
- Unlimited real-time reads
- Simulator & validation harness
- Community support
Studio
Usage-based
Billed per profiled session, not per request
- Everything in Sandbox
- AI profiler with prompt caching
- Matchmaking, difficulty, wellbeing, recs
- Multiple titles, one DNA vocabulary
- Integration support from our engineers
Enterprise
Custom
For publishers and regulated operators
- Private deployment on your Postgres
- Cross-title DNA programs
- Responsible-gaming rule tuning
- Dedicated success engineer
- Custom SLAs
Questions studios ask us
Will this add latency to my matchmaking queue?
No. The AI runs asynchronously, once per session, after ingestion has already returned. Matchmaking, difficulty, wellbeing and recommendations are vector reads plus arithmetic, 1–7 ms server-side, with llmCalls: 0 on every response.
Do I need to train a model or hire ML engineers?
No. You write a JSON manifest that describes your game in plain language and map your telemetry to a small event envelope. There's no per-title model to train or retrain.
What happens if the AI provider has an outage?
Your game doesn't notice. The real-time path never calls a model. Profiling falls back to a deterministic profiler and records which source produced each profile.
What data do I need to send?
At minimum: player decisions (what they chose, how long it took, how much they committed), match outcomes, and a session-end marker. Economy and social events are optional and power the wellbeing and sociability signals.
Which engines and languages are supported?
Anything that can make an HTTP request from your backend: Unity, Unreal, Godot, Flutter, web, or a custom engine. We also provide a dependency-free TypeScript reference client.
Which AI model powers the profiler?
Claude, through a single forced tool call that guarantees a typed profile, with the rubric and manifest prompt-cached across all players of a game. You can choose a larger model for accuracy or a smaller one for cost.
Can I explain a decision to a player or regulator?
Yes. Each trait stores the evidence behind it, each wellbeing flag carries the numbers that triggered it, and each matchmaking result includes a human-readable reason.
Locus
Ready to see Locus
in your game?
We are onboarding a limited number of studios. Tell us about your game and our team will scope your first integration with you.