Top Creator winners

These were the fixed-prize winners in the Top Creator track:

1st place — AI Influence
Prize: $1000

2nd place — EchoColony
Prize: $500

New Creator winners

These were the winners in the New Creator track, where prizes were based on revenue share performance, with a minimum payout of $250 and a cap of $1000:

SentientSands — $1000
The Living Valley — $1000
Voices of the Court — $700
Sims4: Boundless — $250
CobbleBrain — $360

What stood out

What we wanted from this jam was not just “AI chat in a game.” We were looking for mods that made roleplay matter inside the game world: NPCs with memory, choices that changed what happened next, and stories that could move in different directions depending on how the player played. That was the whole idea behind the jam from the start.

A big thank you to everyone who entered. Building this kind of system is still hard, and every good submission helps show what this new style of game can look like.

What’s next

One more thing: you can now resubmit published games to our next jam.
https://itch.io/jam/player2-ai-rpg-jam-2/entries

If you built something for this jam and want to keep improving it, keep going. If you have already published a game and think it now fits the next jam better, you can enter again.

That’s something we want to encourage more of: not just jam prototypes, but projects that keep growing after the jam ends. The RPG Jam itself was designed to reward sustained development and real player usage, and we want to keep pushing in that direction.

Below are the official results, plus what’s coming next.

🏆 Results

$1,500 — OVERALL GRAND PRIZE WINNER

Barotrauma AI NPCs

$600 — Best Web Game (A web game playable on our site)

Kirishima News Club

$600 — Best AI Story

Mira

$600 — Best AI Mod

FelPawns

$300 — Next six places

• Hallway7

• The Between

• Chaos Inflicter

• S.M.G

• The Therapy Room

• Warrior 52

You can view the full jam page, all entries, and the final results here:
https://itch.io/jam/player2-ai-npc-jam4

💛 A Note to Everyone Who Submitted

There were a lot of really good games that didn’t make the winners list. That’s one of the hardest parts of judging and ranking: the cutoff happens fast, and many entries could easily have landed in the top group on a different day or with a different scoring mix.

If you built something for this jam: thank you. You made the community better and pushed the idea of “AI NPCs” in surprising directions.

🚀 Next Up: Player2 AI RPG Jam

If you had fun building (or you’ve been waiting for the right theme), the next jam is live:

Player2 AI RPG Jam
https://itch.io/jam/player2-ai-rpg-jam

Bring your party systems, narrative experiments, combat prototypes, procedural quests, companion NPCs—whatever “AI RPG” means to you. We’re excited to see what you ship.

Play the Entries + Leave Some Love

If you haven’t yet, take a moment to browse the submissions and leave a comment on a few games. A short note goes a long way for jam creators.

Thanks again for making Player2 AI NPC Jam #4 such a blast. See you in the next one. 🎮✨

That’s what happened with our earlier AI NPC jams—enough that it meaningfully accelerated new games shipping on Player2.

Now we’re running a new kind of jam—one focused on mods and sustained development.

Jam page (details + how to join): https://itch.io/jam/player2-ai-rpg-jam

The premise

Welcome to the Player2 AI RPG Jam: a jam for mods (not brand-new standalone games) that transform an existing base game into a new genre of RPG where:

• every NPC feels alive (memory, goals, relationships)

• conversation and player actions cause real in-game events

• the story and world change because you roleplay

This jam’s theme is not “add an AI chat window.”

It’s roleplay that drives world events—like letting the player live their own Fallout / GoT / Witcher movie inside a game, with dialogue and decisions bending the plot and generating consequences.

What to build (and what not to build)

✅ We encourage

• Mods for popular base games that don’t already have a clean theme-fitting mod

• Deep integration: dialogue ↔ systems ↔ narrative ↔ world events

• “Emergent story engines” driven by player behavior

❌ We will NOT reward

• Clones of an existing theme-fitting mod in the same base game
  (Example from the doc: “don’t build AI Influence in Bannerlord again.”)

Rewards: designed for real usage (not tiny throwaway demos)

Past jams sometimes incentivized splitting projects into micro-demos to maximize chances—then lots of entries got abandoned. This time, we’re explicitly optimising for sustained development and real player usage.

Important dates

• Eligibility cutoff: publish your mod on Player2 before March 1, 2026

• Judging window: Feb 1 → Mar 31, 2026

• Rewards granted: March 31, 2026

Two tracks

Track 1: Top Creators
If you’ve earned >$1000 total from Player2 before Feb 1, you’re in Top Creators.

Judged on growth + usage growth during the window.

Rewards:

• Winner: $1000 bonus

• Runner-up: $500 bonus

Track 2: Up-and-Coming Creators
Everyone else.

Judged on:

• Theme fidelity (does it feel like a living RPG movie?)

• Live development (frequent updates + responding to feedback)

• Player metrics (real engagement/traction)

Rewards (up to 5 winners):

• Each winner gets 10× their revenue share earned from March 1 onward until their total payout exceeds $1000, then normal payout resumes

For detailed submission rules and scoring criteria, visit the Jam page and join: https://itch.io/jam/player2-ai-rpg-jam

See you in the jam.

💻 code available at https://github.com/elefant-ai/open-p2p

🤗 data available at: https://huggingface.co/datasets/elefantai/p2p-full-data

🎥 videos available at https://elefant-ai.github.io/open-p2p/

(a collection of games the model can play at a reasonable level right now)

We introduce Pixels2Play (P2P), an open-source generalist agent designed for real-time control across diverse 3D video games on consumer-grade GPUs. Built on an efficient, decoder-only transformer architecture that predicts keyboard and mouse actions from raw pixel inputs , the model is trained on a massive new dataset of over 8,300 hours of high-quality, text-annotated human gameplay. Beyond achieving human-level competence in commercial environments like DOOM and Roblox , we systematically investigate the scaling laws of behavior cloning, demonstrating that increasing model and data scale significantly improves causal reasoning and mitigates the "causal confusion" often inherent in imitation learning. To accelerate research in generalist game AI, we are releasing the full training recipe, model checkpoints, and our extensive gameplay dataset under an open license.

🎮 Now watch our model plays with a real human player on Roblox Rival!

Training Dataset:

We release a large-scale dataset of high-quality human gameplay spanning diverse 3D video games including FPS (DOOM, Quake, Call of Duty, etc.), racing (Need for Speed, etc.), Roblox games, and other popular video games. All gameplay is recorded at 20 FPS by experienced players. Each frame is annotated with keyboard and mouse actions, and text instructions are provided when available.

Policy Model:

Our model, Pixels2Play (P2P), is an action policy that takes visual observations and optional text instructions as input to output keyboard and mouse actions. The model is designed for high-speed, real-time inference (20 Hz) on consumer-level GPUs. The architecture is composed of a backbone transformer and a lightweight action decoder. The backbone transformer is responsible for sophisticated spatio-temporal reasoning between the visual inputs, text inputs, and the output action. The action decoder then predicts the final mouse and keyboard actions based on a compressed action prediction token generated by the backbone. This structure accelerates inference speed by a factor of 5 while maintaining high prediction accuracy.

The model employs an EfficientNet-based image encoder to compress visual observations into compact visual tokens, and a Gemma text encoder to compress text instructions into compact text tokens. One ground truth action consists of eight tokens: four representing simultaneous keyboard actions, two representing mouse movement on the x and y axes, and two representing mouse button actions. These ground truth action tokens are provided as input so the model can leverage prior actions to perform more like a human. To maintain the causality of the model, we designed a customized attention mask to ensure that the action prediction token only attends to prior ground truth action tokens.

Causality and Scaling:

1. Causality in a Controlled Toy Environment

In this controlled setup, an optimal linear policy exists that can solve the task. However, we found that standard gradient descent fails to find this solution in linear models. By increasing network depth and adding non-linearity, the optimization process is better able to overcome spurious correlations. These results suggest that increased capacity and depth do not just improve performance—they actively facilitate the discovery of true causal signals during training.

2. Empirical Evidence in Large-Scale Environments

In our full-scale experiments, we observed an empirical phenomenon that mirrors the findings from the toy example.

We found that increasing model parameters and dataset volume naturally mitigates causal confusion. Even without explicit architectural interventions to address causality, larger models demonstrate a superior ability to distinguish between essential environmental cues and non-causal distractors. This suggests that scale provides a practical, robust solution to the causal challenges inherent in generalist gaming agents.

AI Influence is trying to push past that. The mod (currently in BETA) focuses on AI-driven interaction where conversations can connect to diplomacy, actions, and systemic outcomes—closer to the “your words change the world” and players went crazy over it.

AI Influence takes up all my time

AI Influence is MFive’s main project:

“My most important mod is, of course, AI Influence. A lot of time and effort has been invested in it, and I’m still working on it… the mod is in BETA, and its functionality is still far from being released.”

They’ve also built other mods over time—Balanced Battle Resolve and Advanced Battle Resolve, Dynamic Clan System, and Lords Gear—but AI Influence is their current focus.

How AI Influence started: personal project → public BETA (for bugfinding)

AI Influence began as a personal mod that later went public:

“AI Influence was initially created purely for personal purposes, after which I posted it on Nexus.”

The core inspiration was seeing AI used in ways that go beyond flavour text:

“…old Mount and Blade 2 videos, where the AI performs actions and your conversations actually influence something technical…”

They did early diplomacy/action testing alone:

“I conducted the initial tests of diplomacy, actions, and so on alone. I didn’t have any beta testers or anyone else; I’m working on the project alone…”

Then came a practical decision many solo modders will recognise: ship the BETA publicly and let real players surface edge cases.

“…it was decided to distribute the BETA version online so players could help find bugs and issues.”

They explicitly credit active bug reporters as part of the project’s progress:

“Therefore, every active player who has ever submitted a bug report is a participant in this work of art… a huge thank you to them all!”

Why Player2: convenient iteration, stable infra, and “a happy medium” model

MFive first discovered Player2 through another mod, then used it during AI Influence development because it was convenient for testing:

“When I was making AI Influence, I tested my mod using Player2’s AI because it was convenient for me: a few hours were enough, and a few hours in Player2 is free. My daily energy was more than enough.”

Their experience summary of player2 is pragmatic:

• Reliability/support: “pretty good support and servers” (no issues encountered personally)

• Model tradeoff: “a happy medium between smart and expensive and dumb and weak”

• User growth: “Yes, Player2 helped attract more users. I think we help each other out to some extent.”

• Revenue share: “A revenue-sharing program is a great way and motivation to work on projects like AI Influence… or other games.”

(That last point matters because big “systems” mods tend to be long-running efforts, and sustained maintenance is usually the hard part.)

What’s changed recently on Player2: a free gpt-oss-120b option for players

A platform update worth noting for mod developers: Player2 now offers an entirely free option for players using OpenAI’s open-weight gpt-oss-120b model.

Two reasons this is relevant to AI mods:

1. It lowers friction for players trying the mod (they can play without worrying about cost at all).

2. It gives players access to a model scale that’s often unrealistic to run locally.

OpenAI describes gpt-oss-120b as an open-weight reasoning model that can run efficiently on a single 80GB GPU. OpenAI The official repo, likewise frames gpt-oss-120b as targeting use cases that “fit into a single 80GB GPU (like NVIDIA H100 or AMD MI300X).” GitHub

In other words: while some enthusiasts can run strong local models, most players can’t practically host something in this tier—we are now offering this model as our smallest and free option, so “free to play on Player2” can still mean “powerful enough to feel good in-game”, beyond a few hours free a day.

See it working: install walkthrough + gameplay

If you want to understand what “AI Influence + Player2” looks like end-to-end, these community videos help:

• Installation walkthrough (community-made, not sponsored by us, all credits to spartank): YouTube
  

• Gameplay video from Spartank:

We asked for MFive’s advice to new mod developers: design for model churn, don’t freeze your architecture

Their most actionable guidance wasn’t about “prompt craft,” but about future-proofing:

“For anyone interested in working with AI as a developer, I would advise building your structure to find a happy medium. AI is getting smarter every day. Models that are expensive today may become cheaper tomorrow, and smarter models will emerge.”

And a direct encouragement to avoid building an artificially “small” mod just because today’s models have limits:

“So, don’t be afraid to make your modifications with an eye toward the future and don’t limit yourself to the complexity of your AI requests.”

That’s a useful mental model for AI mods in general: treat models as a moving dependency, and keep your integration flexible so you can swap up in quality/latency/cost over time without rewriting the mod.

Now, if you are interested in building your own AI Influence mod with player2 in games, go to the developer portal on the player2 site and read the tutorials/documents there!

AI and LLMs have quietly become the most powerful NPC design tool modders have ever had. Not just for “more text,” but for characters that feel alive, react to the world, and remember what’s happened to them.

Today I want to share the story of Gerik Uylerk (a.k.a. Gerik), the creator of EchoColony and the one who helped onboarding RimTalk to player2 for RimWorld — and how building AI NPC mods with Player2 didn’t just bring his pawns to life… it nearly tripled his user base and turned modding into a real source of income for him.

And if you’re a mod developer, I’ll show you exactly how you can do the same with your own game or mod.

“I didn’t want longer text. I wanted characters that felt alive.”

“I noticed that there were already some mods using AI to generate stories or enhance certain aspects of RimWorld, but none of them really matched what I wanted. I didn’t just want longer text or more complex storytelling. I wanted characters that felt alive.”

Gerik has been a developer for years, and like many of us, he’s obsessed with strategy and management games. RimWorld in particular.

Every time he watched a pawn eat, build, or wander around camp, he had the same question:

What would this little person say if they could actually talk to me?

That question eventually became EchoColony — mods that give RimWorld pawns their own voices, memories, and opinions, driven by AI and their in-game experiences.

Instead of the player writing the dialogue, the pawns talk for themselves. Their words come from their traits, relationships, battles, and daily life in the colony.

If you haven’t seen it yet, check out Gerik’s RimWorld NPC chat in action — watching pawns banter and process battles together is honestly wild.

From “AI sounds cool” to “oh wow, this actually works”

Gerik had experimented with different AI APIs before Player2. You probably know the drill:

• Complicated registrations

• “Where is my API key again?”

• Long docs, rate limits, logins, portals…

So when he tried Player2 for the first time, his reaction was basically: wait, that’s it?

“You just download it, install it and that’s it. No complicated registration, no jumping through confusing pages just to get an API key. I had been testing different AI APIs for a long time and it was the first time I found something like this.”

For RimTalk and EchoColony, he wired things up so that the mod could work:

• With the Player2 desktop app, for players who just want plug-and-play

• With a manually assigned API key, for power users who prefer that route

Once integrated, two things stood out immediately for him:

“The first one was the speed of the responses. I can have long conversations with the NPCs and it never gets overloaded.”

“The second thing is the very generous free usage… I’ve never run out of the free quota. It’s amazing.”

For a game like RimWorld, where you can spend hundreds of hours in a single save, that matters a lot. Players don’t want to worry about running out of AI just because they talked to their pawns too much.

Did AI NPCs actually bring more players? Yes. A lot.

This part is simple:

“Without exaggerating, I’d say it practically tripled my user base. Even though I added different AI options, none of them caused the download numbers to grow as much as Player2 did. Everyone loves an AI that’s easy to use.”

Mod players are not asking for more complexity. They want:

• Click, play, talk — no dev account, no extra forms

• NPCs that respond quickly and stay in character

• Something that “just works” in the background

Player2 is built around that exact idea: make it frictionless for players, and flexible for mod developers.

That combo is what drove the spike in adoption for EchoColony and RimTalk. The tech is cool, sure — but the usability is what brought the crowd.

Turning passion into income: the revenue share program

Gerik didn’t start doing this for money.

He started making mods during a difficult moment in his life, mostly as a way to keep his mind busy and pour energy into something he loved. Then something unexpected happened:

“I realized that doing what I love could also become a source of income. That gave me back a lot of motivation.”

With Player2’s revenue share program, he gets a transparent monthly payout based on player usage, plus extra from game jams and prizes.

No legal labyrinth. No predatory contracts. No tiny-print conditions.

“There are no complicated contracts, no endless requirements or impossible conditions. You just focus on making fun mods or games for the players.”

We handle the boring parts: billing, infrastructure, and tracking usage. Modders like Gerik focus on the fun parts: making weird, delightful, emotional NPCs that players get attached to.

What makes a great AI NPC mod? (Gerik’s advice)

If you’re a mod developer thinking, “Okay, I want to try this”, here’s some of Gerik’s best advice.

“A key aspect is that every NPC should feel organic and alive, with their own personality, their own conflicts and their own way of seeing the world. But the most important thing is that your game or your mod has to be fun.”

Here are some practical principles you can steal immediately:

1. Tie dialogue to what’s happening in the world

NPCs shouldn’t feel like floating chatbots. They should react to:

• Recent battles

• New constructions or disasters

• Relationships, breakups, betrayals

• Injuries, mood, needs, traits

In EchoColony, some of the best moments come after battles, when colonists process what just happened — celebrating victory or grieving losses in their own way.

Those small, grounded reactions are what make them feel real.

2. Avoid over-decorated “poetic” AI

It’s tempting to prompt your NPCs to sound clever and literary. But if every line reads like a dramatic monologue, players will eventually disengage.

“If the NPC… talks in an over-decorated way, like it’s reciting a fancy template, then it stops feeling like a character and becomes a simple question and answer machine.”

Instead:

• Keep the voice consistent with the NPC’s traits

• Let them be simple, grumpy, weird, awkward — not all of them need to be geniuses

• Use shorter lines during gameplay, save longer ones for big moments

3. Remember: fun > fancy AI

A technically impressive AI that isn’t fun is still a bad mod.

• Don’t overcomplicate setup or UI

• Don’t force players through walls of configuration

• Give them one button: “Talk to NPC”

• Make sure what comes out is entertaining, not just impressive

“That’s what you should aim for: characters that truly live inside the world, not characters that only answer questions.”

Why mod developers love building with Player2

Here’s what stands out for modders like Gerik:

🧰 1. Simple to integrate

You can support:

• App-based usage — players just run the Player2 app

• API key usage — for web users who do not want to or can not use localhost

You don’t have to build your own infrastructure, auth, or billing.

⚡ 2. Fast, reliable responses

Long conversations, heavy play sessions, lots of simultaneous NPCs — and it still stays responsive. That’s crucial for keeping immersion.

🎮 3. Generous free usage for players

Your players can talk to NPCs for a long time without touching a paywall. That makes it easier for your mod to spread and build a community.

🤝 4. A dev-focused community

“There are developers helping each other every day, and there’s also a team behind Player2 that listens to users, considers suggestions and keeps adding improvements.”

Need more TTS voices? Better dev tools? People ask — and we ship.

Okay, I’m in. How do I start building my own AI NPC mod?

Here’s a simple path you can follow:

Step 1: Pick the game & the fantasy

• RimWorld, Mount&Blade, and Minecraft are the most popular games on our platform right now

• Skyrim, Sim4 are commonly asked

Step 2: Decide how NPCs “know” the world

Think about what you’ll pass to the AI:

• Current mood / health / traits

• Recent events (e.g. “lost a friend in battle 2h ago”)

• Relationships (e.g. “married to X, hates Y”)

• Context of the conversation (is the player interrogating? comforting? joking?)

You don’t need to send everything — just enough for the AI to roleplay convincingly.

Step 3: Wire your mod to Player2

• Go to https://player2.game/profile/developer

• Check Web API and Localhost API docs depending on your need

• Integrate the APIs in your game

• Upload the mod from the creation console above

Gerik even made it so that everything can be done from inside the game itself, which massively improves UX for players.

Step 4: Ship it, then iterate with real players

Want to build the next EchoColony?

Gerik turned a simple question — “What would my pawns say?” — into:

• A set of beloved AI NPC mods

• A tripled user base

• And real monthly income through Player2’s revenue share program

If you’re a mod developer, then this is your moment.

Start making your NPCs alive with Player2.
Build something your players will still be talking about after the game is closed.

And if you do?
We’d love to feature your mod next.

This round was no exception. We saw games that turned language itself into combat, web games that felt shockingly polished for a jam, and some truly unhinged NPCs that we’ll be thinking about for a long time.

Thank you to everyone who entered, shared builds in Discord, playtested each other’s games, and helped push the Player2 ecosystem forward.

How judging worked

We used the same philosophy as previous jams: combine human judgment with real player behavior, then resolve everything with a Condorcet voting method so that the final ranking reflects true head-to-head strength rather than just averages.

In practice, here’s what we did:

1. Three human judges independently ranked the games.

2. We also built a player metrics ranking from live data (things like unique players, time played, and how often people came back).

3. That player ranking was treated as two additional “votes”, alongside the judges’ ballots.

4. We then ran all of those rankings through a Condorcet algorithm (Schulze / beat-path) to decide the final order: whenever possible, the winner is the game that wins its one-on-one matchups against all others.

This setup rewards games that judges love and that regular players actually keep playing.

Overall Grand Prize — $5,000

🏆 Cat & Sword

Cat & Sword is a chatty idle RPG where you’re not the hero—you’re the sentient sword. You banter with your catgirl wielder, assemble a guild of oddball adventurers, and dive into dungeons while your AI-driven party quips, argues, and reacts to everything you do.

Huge congrats to the team behind Cat & Sword!

Category Winners — $1,000 each

📱 Best Mobile Game — Word War

Word War turns language into a battlefield.

🌐 Best Web Game — Epsilon IV

Epsilon IV allows you to step into the shoes of Cylil, a special agent of the Talos Corporation. On a return trip from a mission, you encounter a distress signal from the mining station Epsilon IV.

🛠 Best AI Mod — The Storyteller’s Voice

The Storyteller’s Voice shows what happens when you drop a dynamic narrator AI to take over Rimworld.

Runner-up — $500

🥈 Overdrive Tournament

The future is here! But the engines still roar! Drive in Overdrive Tournament and get teased by the announcer.

Special Awards — $500 each

🎨 Best Art — Croaking

Croaking is a 2D side-scrolling kingdom builder where you play as the Frog King, leading your loyal frogfolk to gather resources, build defences, and expand your swampy realm.

🧪 Most Unique Gameplay — Tales of Luanda: Birth of Stone

Step into the ancient Rift Valley, where legends walk among warriors.

🤪 Weirdest NPC — Shadows in the Skin

In Shadows in the Skin, you are a detective who wants to take a vacation to Miami on a luxury cruise ship. Our judges loved it.

$250 Awards — Next Ten Places

The following games each receive $250. These are projects that scored highly across judges and player behavior—games we believe are onto something and that we’d love to see continue.

(In alphabetical order:)

• A.I.P.O.V

• Cognitive Pulse

• MLM Empire

• Mystery Machine

• Otherside Forest

• Sir, Your Order, Please!?

• Tend the Pen

• The Collector

• We All Lie

• You Died, Actually

Some are tense mysteries, some are cozy, some are chaotic social experiments—but all of them did something interesting with AI NPCs, and all of them got strong engagement from players.

Thank you — and what’s next

To everyone who:

• Pulled late nights hacking dialogue systems together

• Fought with web builds and mobile controls

• Wrote lore documents way too long for a jam

• Hung out in the Discord, shared screenshots, and playtested each other’s games

—Thank you. You’re the reason these jams keep getting more interesting and stranger.

If you didn’t place this time, please don’t treat this as the end. Some of the most successful games on the platform started as “jam prototypes” that didn’t win anything at first. Keep iterating, keep shipping, and keep experimenting with what AI NPCs can be.

See you at the next jam. 🕹️💬

Our first AI Story Jam ($2750 prizes) https://itch.io/jam/player2-ai-story-jam
Next AI NPC Jam ($5100 prizes) https://itch.io/jam/player2-ai-npc-jam4

We’re excited to share our ICCV MMRAgI workshop paper: “Learning to play: A Multimodal Agent for 3D Game-Play.” [link to full-text paper]

This blog post breaks down our work on a novel multimodal agent that learns to play a variety of 3D first-person games. It takes both video (what it “sees”) and text (what it’s “told”) as input and outputs keyboard and mouse actions. The model is also efficient in inference as in it can run at 20 FPS on a single RTX 5090 GPU, making it reliable to interact with the games in real-time.

What’s New? From Cog2025 to Now

This work significantly extends our previous Cog2025 paper. The four biggest improvements are:

1. Massive Data Expansion: We collect more than 10x labelled data and 20x more unlabelled data compared with the prior work, with a significant expansion over game categories too.

2. Text-augmented data: We annotated all the labelled and unlabelled data with a VLM so frames will have the text instructions whenever it’s applicable (e.g., some frames will be annotated with the instruction: “approach the skull gate”).

3. Text-Conditioning: With the help from the new dataset, the agent can now follow natural language instructions. This means you can tell it what to do during gameplay (e.g., “pick up the key”), and it will adjust its actions accordingly. To achieve this, we added a text token to the backbone transformer model so the action can attend to that.

4. Lightweight action decoder: We introduced a separate, lightweight action decoder in addition to the main backbone transformer. This small, decoder-only model is responsible for autoregressively generating the final actions. By offloading this task from the large backbone transformer, this design greatly reduces our model’s inference latency.

Our Three-Stage Training Recipe

We developed a three-stage training recipe so that we can leverage the unlabelled dataset (which is now 10x bigger than our labelled dataset and growing rapidly!).

The key was to train an IDM (Inverse Dynamics Model) on our labeled dataset. This IDM learns to watch video frames and predict the action that the player took. We then used this trained IDM to automatically generate pseudo-labels for the massive unlabeled video dataset.

This resulted in our final training pipeline:

• Stage 1: IDM Training We train the IDM on our high-quality labeled data to teach it to predict actions from video.

• Stage 2: Policy Pre-training We pre-train the main policy agent on a mixture of our labeled data and the much larger unlabeled dataset (which now has pseudo-labels from the IDM).

• Stage 3: Policy Fine-tuning Finally, we fine-tune the policy exclusively on the original, high-quality labeled data to refine its performance and ensure accuracy.

We’re excited about this new model and believe it’s a promising step toward agents that can learn from the vast amount of video content available online. Be sure to check out the full paper for all the details!

Watch our model play games below! 🎮

Non-text-conditioned game-playing

Overall, the new model significantly outperforms our previous version (see our earlier blog post) — it can handle more complex tasks, play with clearer intent, and even complete early levels of classic games like Doom and Quake with minimal human intervention. However, it still struggles with long-horizon gameplay where planning and memory are critical.

Below, we break down its performance game by game:

Hovercraft (internal test game):

The new model plays the Hovercraft game almost like a human player — a major improvement over the previous model!

Simple-FPS (internal test game)

In Simple-FPS, the model performs reasonably well. It shoots enemies with good accuracy and navigates effectively to new targets.

msdos/doom:

We’re excited to see the model complete the first level of Doom with only two brief human interventions (about one second each, at 1:33 and 3:19 to trigger switches). The model consistently aims well, eliminates enemies, and completes objectives with clear intent.

msdos/quake:

The model also completes the first level of Quake with two short interventions (at 1:18 and 2:03, around three seconds each). It occasionally gets stuck when the trajectory goes out of distribution or re-explores the same area. Still, this represents a significant improvement over the prior model — it completes more objectives and shows better spatial reasoning.

Roblox / be a shark

In simpler games like Be a Shark, the new model demonstrates precise aiming and purposeful movement, showing better control.

Roblox / natural disaster survival

The model successfully survives easier disaster scenarios, reacting appropriately when the threats are straightforward to avoid.

msdos/need for speed

The new model drives much more like a human player, maintaining control and finishing with a higher rank compared to the older version.

need for speed rivals

Similarly, the model shows improved driving behavior, though it still struggles — especially since this game was not included in training.

goat simulator 3

The model struggles with long-horizon, open-ended games like Goat Simulator 3 (and Left 4 Dead 2 below) when not guided by text instructions

left 4 dead 2

Text-conditioned game-playing

We evaluate the model on Quake and Doom, as these games allow checkpoint saving — ensuring that each run starts from the exact same scenario for fair comparison.

In Doom, we begin from a scene where the model can either pick up a shotgun or explore other areas. We test two text instructions:

1. “Pick up the shotgun.”

2. “Proceed to the red cross gate.”

The model follows the first instruction consistently well, while performance on the second is slightly weaker. Nonetheless, both cases show that text conditioning improves behavior compared to runs without any instruction.

no text-input:

pick up shotgun:

proceed to the red-cross gate

In Quake, we start from a scene where the model must press a wall switch to lower a bridge. Without any text prompt, the model typically fails to trigger the switch. However, with the instruction “move to the wall and press the red button,” it successfully presses the switch in some trials — as shown below:

no text-input:

move to the wall and press the red button:

A few words from one of the judges:
“In general, it would be how much care the developers put into it. And we can see the best games are made with a lot of love, which is really cool to see.”

🏆 Overall Winner — $5,000

• The Daily Ledger - AI shapes items, customers and haggling. The game is also polished, with rarer items and customers as you progress, making it replayable.

🥇 Category Winners — $2,000 each

• First Place AI NPC Mod: Fallout: Echo Protocol - turns Fallout 4’s Automatron into a fully conversational

• First Place AI Social Game: AI ShowDown - A party game where you outwit friends by asking free‑form questions and guessing AI‑picked characters.

• Most Creative Game: Shift Protocol - A trust‑dilemma escape where you talk with three AI characters as puzzles and choices branch outcomes in a re-zero fashion.

🎖 Runner‑ups — $1,000 each

• Exile Life - PSX‑style survival horror where bartering meds and building trust with an AI companion matter as much as scavenging.

• Word Alchemy - Combine words and let AI forge/score your creations for leaderboards. Surprisingly fun and novel.

• Oio/Ois..u - A watercolour-gothic puzzle tale where you chat with a snarky magic mirror and untangle glitched fairytale logic. We loved the art style.

• Eli’s Yard Sale - Haggle with a hyper 10‑year‑old. I don’t know why it is so fun. Bought Mona Lisa for $50.

💵 Awards — $500

• Love Finder - A dating “bossfight” where improvisation and tone carry you through distinct AI personalities.

• Gentle or Brutal - Time‑pressured interrogations with procedurally generated suspects. Shake them up a little?

• Bar Talk: Neural Nights - Make friends at bars and clubs; NPCs rate your vibe in real time—be charming or get bounced.

• Tales of Luanda: Rise of Magere - A Kenyan folklore adventure of duels and diplomacy where choices, rivals and crises are AI‑driven.

• Investigation Session - Every case is new—study the secret file, grill the suspect and pull a confession from an AI‑generated criminal.

💵 Awards — $250

• Please Hire Me!!! - A conversational job‑interview gauntlet that turns small talk into high stakes.

• Date AI - A minimalist dating sim where your charm versus her code decides the outcome.

• False Haven - A metroidvania of illusions and discovery framed by a parent’s desperate search for their daughter.

• The Greatest Quizshow! - Intellect‑themed minigames scored by NPC observers.

• The Watcher - Tactical FPS infiltration guided by an Overwatch‑style AI operator.

• Who Killed My Sister - First‑day‑on‑the‑force detective roleplay that leans into conversational sleuthing.

• Eastshore- Colony‑builder where villagers banter as they gather, build and struggle to survive in a strange land.

• Who Killed The Guy Who Died? - A tongue‑in‑cheek whodunnit in a grand mansion—interrogate everyone and embrace the absurd.

• A Village Tales - Investigate a missing‑person case across an AI‑driven village where every villager talks back.

• Crown and Counsel - Rule by typing anything—an AI “game master” generates crises, interprets commands and narrates consequences.

💵 Encouragement Awards — $100

• Take Out - Serve orders at a haunted kiosk across five unnerving nights while mastering recipes and surviving the shift.

• Loot of the Living Dead - An AI‑powered game show staged in a graveyard. Did I say it is a show?

• Chocolates for Mother - Hustle chocolates door‑to‑door to save your mom—simple premise, earnest heart.

• Crystal Bound - Command chess pieces with natural‑language instructions in a puzzle‑forward twist on tactics.

How we judged (fair & transparent)

As in our first jam’s announcement, we combined human judging with player behaviour:

• Multiple judges submitted ranked lists.

• We included player‑metrics rankings (DAU, time spent, retention) as two votes alongside judge ballots.

• We aggregated results using Condorcet methods on head‑to‑head matchups.

• Category awards were informed by the data and judge discussion to celebrate standout achievements across different design goals.

What’s next

We will reach out to the winners through Itch.io to arrange the prizes, so please keep an eye out.

The games are live at https://player2.game/ —jump in to play the winners, try the experiments, and leave feedback that helps teams grow. If your game didn’t place, don’t stop: the community—and the future of AI‑driven characters—moves forward with every iteration. Join our next jam at https://itch.io/jam/player2-ai-npc-3.

Thanks again to everyone who built, tested, judged, streamed, and cheered. We’re excited to keep prototyping the future of intelligent NPCs—together.

1) Revenue Share — Boosted & Simplified

• Patron revenue share increases from 20% → 60% for creators.

• Lower minimum threshold to join the program, so more games can start earning sooner.

• The reward pool is now fully funded by actual platform revenue (moving away from the old $200 × qualifying games/month model).

• We picked 60% so the total pool this month is the same as last month, compared to if we hadn’t changed the funding approach. This keeps continuity while aligning rewards directly with platform performance.

Why this helps: The pool now scales naturally with the community. As Player2.Game grows, so does the upside for developers—without extra rules or math to track.

2) Rewards — Game Jams > Bounties & Bonuses

We’re shifting additional rewards to recurring Game Jams every 2–3 months with cash prizes (on top of your revenue share).

To keep things clear—and get more creators paid—we’re retiring:

• the $3,000 bounty (first AI mod in a game earning > $100/month), and

• the $500 bonus (for any game hitting $50/month for the first time).

Why the switch: Very few developers actually received those bounty/bonus payouts, while Game Jam prizes have consistently reached more creators. Jams are fun, community-driven, and much easier paths to prizes.

You can find our next game jam from https://itch.io/jam/player2-ai-npc-3.

Why we’re doing this

• More creators share in the upside. A larger slice of Patron revenue + a revenue-funded pool means rewards track the real value the platform creates.

• Clearer, simpler payouts. Less fine print, more straightforward earnings.

• Easier prizes. Game Jams reliably put cash in more builders’ hands than the old bounty/bonus programs.

• Community energy. Jams bring momentum, collaboration, and visibility to your projects. 🏆

What you can do next

• Ship (or update) your game to benefit from the 60% Patron share and lower threshold.

• Start prepping for the next Game Jam. These will run every 2–3 months, with prize details announced ahead of time.

• Ask questions. We’re happy to clarify anything about the updated revenue share or jam schedule on Discord.

FAQ

Does the new 60% share reduce the pool?
No. We selected 60% to keep the pool the same this month vs. last month relative to the previous funding method, while making it scale with real revenue going forward.

Why retire the bounty/bonus programs?
They were high-bar, low-reach: few developers qualified. Game Jams consistently reward more creators and are simpler to understand and plan for.

How often are Game Jams?
Every 2–3 months, with cash prizes on top of your revenue share. You can find our next game jam from https://itch.io/jam/player2-ai-npc-3.

Thanks for building with us. We’re excited to see what you make next—and to reward more of you along the way. 💙

We outlined our work building an initial version of our a high-speed vision to action model Pixels2Play-0.1 (P2P0.1). Please see the paper for more details about our approach.

Below we demonstrate our AI playing a number of different games. All recordings are from the first recording (no cherrypicking) with no interventions. The same model is used to play all games. The game play is certainly not perfect, but we can see the model is beginning to understand how to participate in a variety of games. There is lots of room from improvement.

Feel free to suggest games you’d like to see our AI play in future in the comments below.

Roblox / Be a shark

Roblox / Be a snake

MS-DOS / Need for speed

Doom

Quake

Simple FPS (internal test game)

Hovercraft (internal test game)

Need for Speed: Hot Pursuit

Goat Simulator 3

Need for Speed: Rivals (note: this game was not seen during model training and while the model performs fairly poorly you can see that it can generalize enough from games seen during training to complete the lap).

Why we ran this jam

We kicked off a global game jam to push AI-driven characters beyond the gimmick—to see NPCs that think, react, and adapt in ways that meaningfully change how games play.

• Jam page: https://itch.io/jam/ai-npc-jam

• Play the games: https://player2.game/discover/games

Theme: Beyond AI Gimmick

Build AI characters that are part of the game’s fabric, not bolt-ons:

• React & evolve with game state, player choices, and world events.

• Enhance core mechanics—combat, puzzles, exploration, and social systems.

• Create emergent moments that leave players with unpredictable, memorable stories.

We want intelligent NPCs rooted in your game’s logic and lore, shaping and shaped by everything else going on.

About the jam (what it took to enter)

• Participants needed to join our Discord and click “I am in an official Player2 Game Jam.”

• Use any game engine you like.

• Integrate the free Player2 API to drive all NPC intelligence.

• Submit a playable demo or download link by the deadline.

Prizes

• 1st – $5,000

• 2nd – $2,000

• 3rd – $1,000

• 4th–5th – $500 each

• 6th–10th – $250 each

• Top 10: Featured in the Player2 App

How the judging worked (fair and transparent)

We combined human judgment with player behavior:

1. Six human judges each ranked their Top 10.

2. During the judging window, we tracked DAU, time spent, and retention for every entry.

3. We treated the player-metrics ranking as three votes, alongside the six judge votes—a 6:3 human vs. metrics balance.

4. We ran the combined rankings through a Condorcet (minimax) voting algorithm—a head-to-head method that selects the entry with the strongest overall pairwise performance. It’s more robust—and fair—than simple averages.

🏆 Winners

1) Escape Protocol — $5,000 (Condorcet winner)
Super Creative! We felt so immersed playing this.

2) Watch This, Chat — $2,000
Trying to convince strangers to get in your car is hard. Trying to convince AI, even harder and very, very weird!

3) Social Valley — $1,000
This brings a whole new dimension to Stardew Valley! Now we really get to know the NPCs well!

4) Almost Somewhere — $500
Very creative…and creepy! Great immersive experience and we loved the asthetics!

5) Your Honor, WTF?! — $500
Super bizarre and weird. This is a great way to settle arguments IRL!

6) Player2 NPC — $250
Who doesn’t love bossing NPCs around in Minecraft!

7) Valhart — $250
This feels like Hogwarts all over the place! Get that AI to teach you some spells. Ya gotta find her first!!

8) Village Builder AI — $250
Fun grindy game with NPCs that have weird personalities. The NPCs think they help, but do they really???

9) Murder Tavern — $250
Chaotic fun—AI patrons gossip, scheme, and turn every bar brawl into wild, emergent stories!

10) Silent Therapy — $250
Try and help this guy? It takes a lot of effort and deduction to figure out exactly what he needs!

You can find All Top 10 entries in the Player2 App and play them now. Congratulations!

🎖️ Special Mentions — $50 each (encouragement)

These awards go to games with big potential or standout creativity that could shine with more polish, or missed the top 10 only marginally. Consider this both encouragement and a vote of confidence.

• Vampire Cafe! — Awesome twist on a great genre! This was fun!!

• Cursed Diamonds — This is an AI mystery at its finest! You really WANT to help her out, but can you, really?

• Deltarune AI — Crazy personalities in Deltarune! Who’d have thought that’s what they were thinking??

• Sticky Situation: Werewolf in the Basement — The AI’s in this “multiplayer” deduction/werewolf game were very sneaky…and mean!

• The Last Fortress — Love the look and feel of this. It gets pretty deep with AI in SPAAAAAAAACE!

• McGibble — I love this so much! This game is so weird. You will remember this for quite a while!

• Reign or Ruin — Definitely don’t tick off the villagers in this one! They will destroy you!!

• SchoolCheater — A game about cheating at school and not getting caught. Don’t try this IRL! It’s actually kinda hard….the game, I mean.

• InnerLife - AI-Powered Life Simulator — An uncanny life sim where your AI generates your life events.

There are so many good games with potential in the jam, and we couldn’t have listed them all. What the Hell, RAINSHIFT, Hinari The Nine Tailed Fox, Interrogator Vale, Murder In The Building, to name a few more.

We are overwhelmed by the quality and quantity of the submissions and grateful to the active developer community we have built through this. Read on if you didn’t get top 10, since the journey is just beginning from here.

By the numbers

• 200 participants

• 56 submitted games

• 10 winners, 10 special mentions

• $10,750 awarded in total

Dive into all entries and discover your new favourites: https://itch.io/jam/ai-npc-jam/entries

What’s next: our new jam (bigger prize!) 🚀

We’re launching the next chapter: It’s Alive! AI NPC Jam
👉 https://itch.io/jam/its-alive-ai-npc

Who can enter? Everyone is welcome, even if you won this game jam. You can not enter with games that are already receiving revenue share with us or that have already been placed in our previous game jams.

You can still participate in the next jam if you got a special mention. Come join us with extensive improvements, fresh builds and fresh ideas!

Let’s keep prototyping the future of intelligent characters—together. See you in the next jam!

• Unity 6 or newer

  • Scripts target .NET Standard 2.1 and use async/await.

• TextMesh Pro

  • Required for the sample chat UI.

• Newtonsoft.Json

  • Pulled in via UPM; serializes request/response payloads.

• Player2 Client

  • Download from player2.game.

Step‑by‑Step Setup

1. Add an NpcManager to Your Scene

Insert a NpcManager (singleton) under your scene root. Only one can exist per scene.

2. Configure the NpcManager

1. Game ID: Name your game or mod.

2. Functions → +: Define each action the LLM can call:

   • Name: Identifier (e.g. flame)

   • Description: Hint for the model

   • Arguments: Typed parameters (e.g. radius: number), mark required as needed.

In this example, we expose flame(radius:number) to spawn a fiery VFX cloud.

3. Attach Player2Npc to Your NPC

Select the GameObject representing your NPC (e.g. Person 1) and add Player2Npc.cs.

4. Set Up the Player2Npc Component

• Npc Manager: Drag in your scene’s NpcManager.

• Short / Full Name: For UI labels.

• Character Description: The persona text sent at spawn.

• Input Field / Output Message: Link your TextMesh Pro components.

• Persistent: Tick if the NPC should survive client restarts.

Now press Play—your NPC is live and ready to chat!

Adding Rich NPC Functions (Optional)

To handle LLM-decided actions:

Create a handler script (e.g., ExampleFunctionHandler.cs) in your scene root:

using UnityEngine;

public class ExampleFunctionHandler : MonoBehaviour
{
    public void HandleFunctionCall(FunctionCall call)
    {
        if (call.name == "flame")
        {
            float radius = call.ArgumentAsFloat("radius", defaultValue: 3f);
            SpawnFlameCloud(radius);
        }
    }

    void SpawnFlameCloud(float r)
    {
        // Your VFX / gameplay code here
    }
}

In NpcManager → Function Handler:

• Click +

• Drag the object with ExampleFunctionHandler

• Select ExampleFunctionHandler → HandleFunctionCall

Whenever the model invokes flame, your handler runs on the main thread, so you never manually intervene.

Under the Hood: Quick Architecture Tour

• NpcManager

  • Scene‑singleton holding configuration and the function catalog.

  • Serializes to snake_case via Newtonsoft.Json.

• Player2Npc

  • Spawns NPCs (POST /npcs) and routes UI events to SendChatMessageAsync.

• Player2NpcResponseListener

  • Consumes text/event-stream via UnityWebRequest; raises UnityEvents for text and function calls.

• ExampleFunctionHandler

  • Demonstrates reading function arguments and executing gameplay code.

✨ That’s it! With two inspector tweaks and a tiny handler script, you have LLM‑powered, function‑calling NPCs in Unity. Layer on animations, speech synthesis, or your favorite dialogue UI—and ship your game!

[The bounty program is outdated]

Why AI NPC Mods Matter

AI NPCs are more than just scripted bots—they can react, converse, and even drive emergent gameplay that keeps you coming back for more. But too often, AI characters feel like stand-alone novelties rather than fully integrated parts of the gaming experience. We believe:

• AI should serve the game, not the other way around.

• Indie developers—not faceless algorithms—will shape the next generation of player experiences.

• Great AI NPC mods can breathe new life into beloved titles, creating fresh stories and challenges.

By incentivising creative modders, we accelerate the pace of innovation and reward those who build rich, safe, and sustainable AI-driven experiences.

Our Proven Track Record

Some of our most successful creators have already demonstrated what’s possible:

• DeltaBlade (Bannerlord): ChatAi brings AI-driven conversations to Mount & Blade II: Bannerlord, enabling NPCs to respond dynamically based on their personality, traits, and in-game context.

• Gerik (Rimworld): Have you ever wondered what your colonists think about their daily lives? With EchoColony, you can talk to them directly.

• Conczin (Minecraft): In MCA Reborn, say goodbye to lifeless villagers—Minecraft Comes Alive Reborn turns them into chatty, marryable, and even hireable companions for a true RPG experience!

• TartaricAcid (Minecraft): Touhou Little Maid added cute maids to Minecraft, who can help with battles, farming and other tasks.

Each of these mods not only crossed $100 in monthly revenue share—but far exceeded it—proving that players crave intelligent, interactive NPCs. For more details about these games and suggestions from the developers about how to build an immersive AI NPC mod, read this blog post.

The Bounty Program: How It Works

To reward and accelerate the next wave of AI NPC mods, we’re putting $3,000 on the table for the first mod that hits $100 in Player2 revenue share within any one month. Here’s the rundown:

1. Games List (Will be continuously updated)

   • Undertale

   • Terraria

   • Gmod (Garry’s Mod)

   • Skyrim

   • Fallout

   • Lethal Company

   • Sim4

   • Stardew Valley

   • Project Zomboid

2. Prize

   • $3,000 awarded to the first qualifying mod on each game.

   • Bounty is paid out once your mod reaches $100 in revenue share within a single month.

3. Qualification Criteria

   • Approval by Player2: We’ll verify that your mod is safe, fun, and built on our APIs.

   • API Integration: Must leverage Player2’s SDK for chat and in-game actions.

   • Developer Ownership: You retain full rights, still qualify for the existing revenue share programs from us and commit to ongoing updates.

   • User Interaction: NPCs must be able to chat with players and perform at least basic in-game actions, with more advanced features added over time.

   • First to Threshold: Bounty goes to the first qualifying mod per game. Every other mod/game still qualifies for the $500 top-up when they hit $50 in revenue share within a single month.

4. Launch Date

   • The program kicks off today, July 12, 2025.

Why You Should Participate

• Cash Reward: $3,000—enough to fund your next project.

• Visibility: Featured on Player2’s homepage, blog post, and social channels.

• Community Impact: Be remembered as a pioneer in transforming classic titles with AI.

• Long-Term Growth: Winning mods often go on to earn thousands in recurring revenue share.

Whether you’re a seasoned modder or an aspiring indie dev, there’s never been a better time to experiment with AI NPCs. Our platform handles the heavy lifting—authentication, rate-limiting, automatic scaling, revenue generation—so you can focus on creativity.

Join the Conversation

Players: We’ll continue to expand the bounty list based on your feedback. What games do you want to see come alive with AI NPCs? Drop your suggestions in our Discord with #Player2AIBounty, and we may add them to the list.

Creators: Ready to get started?

1. Join our Discord https://player2.game/discord

2. Grab the player2 App from https://player2.game and explore our live API docs from http://localhost:4315/docs

3. Build your mod, send a test version to Itsuka on Discord

4. Launch it to our community—and claim your bounty!

Let’s empower the indie community to shape the future of interactive storytelling. We can’t wait to see what you build.

Happy modding!
— The Player2 Team

What’s in the box?

Endpoints:

• Spawn an NPC
  POST /npc/games/{game_id}/npcs/spawn
  Spin up an LLM-powered NPC on the client.

• Chat with an NPC
  POST /npc/games/{game_id}/npcs/{npc_id}/chat
  Send a single chat line to that NPC.

• Stream NPC replies
  GET /npc/games/{game_id}/npcs/responses
  Receive the model’s replies as an SSE stream.

• Despawn an NPC
  POST /npc/games/{game_id}/npcs/{npc_id}/kill
  Cleanly remove the NPC when you’re done.

No more manually assembling chat histories or juggling token counts—just hit chat and keep playing.
(Full schema in the bundled OpenAPI spec under the NPC tag.)

Real-world cleanup: AI Gods before & after

Old flow
Our sample mod used to maintain a ConversationHistory object, merge prompts, and orchestrate a multi-step CompletableFuture dance on every player message:

updateNewCharacter(mod)
    .thenComposeAsync(newChar -> {
        String processed = mod.addPlayerStatusToUsrMessage(rawMsg);
        mod.addProcessedUserMessage(processed);
        return CompletableFuture.supplyAsync(() -> {
            return Player2APIService.completeConversation(mod.getConversationHistory());
        }, IO_POOL);
    })
    .thenAcceptAsync(resp -> {
        // 30+ lines of post-processing…
    });

New flow
With the NPC API, the entire history-management layer is gone—just two lines:

String playerName = mod.player.getName().getString();
Player2APIService.completeConversation(rawMsg, playerName);

That’s ~40 lines → 2, and a whole category of bugs wiped out.

We foot the cloud bill, you ship gameplay

• Usage credits, not API keys
  Players top-up inside the Player2 desktop app—no proxying or rate-limiting on your end.

• Zero costs
  We meter tokens, voices, and STT per end-user and get paid by Patrons.

• Zero-config scaling
  Whether you spawn one villager or one hundred, we handle concurrency for you.

Focus on epic boss prompts, emergent quests, or NPC gossip—we’ll handle the boring stuff.

Try it today

1. Download the Player2 Client and explore the Swagger UI at http://127.0.0.1:4315/docs.

2. Check out the AI Gods PR #6 for a concrete migration path.

3. Ship something cool and share it on our Discord.

Happy hacking—and we can’t wait to see the NPC antics you build!

对玩家来说，设置 AI NPC 游戏太难了。

这里复制一个 Key，那里粘贴一个 Key，还要跑 Ollama，修改本地端口设置……
每 20 次下载里，只有 1 个玩家真正进入了游戏。

所以，我们快速转向了新的方向。

💡 Player2：即插即用的 AI NPC

现在使用 Player2，玩家只需要：

✅ 打开应用
✅ 启动游戏
✅ 开始与 AI NPC 对话

就这么简单。

我们不仅提供大模型（LLM），还内置了 语音合成（TTS）、语音识别（STT）、记忆和角色管理工具，全部预先集成，专为游戏设计。

🎮 真正的游戏，真正的成功

我们相信 AI NPC 应该存在于每一款游戏中，但它们应该是 完整游戏体验中的一部分，而不仅仅是独立的噱头。

推动这些游戏未来发展的，将是 独立开发者，而不是 AI 本身。

我们采访了 Player2 平台上最成功的开发者：DeltaBlade（ChatAI）、Gerik（EchoColony）、Conczin（MCA Reborn）和 酒石酸君（Touhou LittleMaid），了解他们是如何让 AI NPC 在真实游戏中活起来的。

Minecraft Comes Alive Reborn - Conczin

“我开发了 MCA Reborn，将原版 Minecraft 中的村民替换成了真正有人格、情绪和对话的 NPC。当模型能力足够强大时，我开始尝试实时对话、TTS 和记忆，让村民更加‘真实’。”

刚开始时，每个月都出现新的模型、新的 API、新的函数调用更新，虽然最初很有趣，但很快维护就变得繁琐。

“Player2 帮我省去了这些繁琐的工作，让我可以专注在玩法本身，而不是花时间处理 API 集成。”

Touhou LittleMaid - 酒石酸君

这款 Minecraft 模组添加了一个 可爱且能帮助玩家的女仆 NPC，旨在解决单人游戏时的孤独问题，同时提供情绪价值和乐趣。

“当可爱的女仆用带点呆萌的语言和玩家对话时，这种情绪价值会被拉满。”

原本的计划是通过分发配置文件，让玩家自行连接 LLM 和 TTS 服务，但：

“市面上现有的服务不仅贵、配置复杂，而且主要面向企业用户。Player2 简单的设计很好地解决了这一问题。”

Player2 将复杂的配置过程简化成几行代码，实现几乎无痛兼容。

有一次，有玩家问 AI 女仆：“如果可以往 Minecraft 加入一个新的生物，你会选什么？” 女仆回答：

“当然是一只可爱、能帮你干活、陪伴你的女仆啦。”

这样的时刻，让角色真的“活”了起来。

Bannerlord ChatAi - DeltaBlade

“我为《骑马与砍杀：领主》做了一个 ChatAi 模组，让玩家可以与 NPC 实时对话。玩家可以提问、获得建议、角色扮演，NPC 甚至可以根据对话做出相应的行动。在使用 Player2 之前，玩家需要自己去申请 OpenAI 的 API Key，再去配置 Azure STT，很多人甚至在尝试之前就放弃了。”

使用 Player2 之后：

“不需要密钥，不需要配置，下载即可使用。集成只花了几个小时，长期来看节省了大量时间，尤其避免了玩家反馈 API Key 无法使用的问题。”

🪐 给 AI NPC 游戏开发者的建议

最后，我们请 DeltaBlade、Conczin、酒石酸君 和 Gerik 给新开发者一些建议：

✅ AI NPC 的集成应该易于上手，对轻度玩家免费使用，同时需要具备防滥用能力。 这很难平衡，因此可以考虑将这部分交给专业服务来处理。

✅ 不要仅依赖 LLM，要把它当作锦上添花的存在，构建在稳定的游戏玩法和情境逻辑之上，让角色具备一致性和上下文记忆，而不仅仅是“聊天机器人”。

✅ 理解 LLM 的边界，它无法稳定地处理复杂的游戏逻辑，很多开发者误以为它们可以完成复杂任务，结果遇到瓶颈。

✅ 掌握 LLM 基本概念和 HTTP 通信知识是必要的，这能帮助你更顺畅地将 AI 集成进游戏。

我们想不出比他们说得更好的总结了。

⚡ 我们刚刚开始

我们已经帮助了 13 位独立开发者，在全球范围内累计超过 1000 万次下载，满足了玩家对更智能、AI 驱动玩法的需求。

开发者无需向我们支付任何费用，仅需几个小时即可切换到 Player2 API。

🔥 前 10 个接入 Player2 的游戏将获得 $500 顶额奖励

在游戏收入达到 $50 时，我们将为下一个接入 Player2 的 10 款游戏提供 $500 的奖励支持。

📩 让你的游戏成为下一个 AI 爆款

在 LinkedIn、Discord 或发邮件至 wenzhe@player2.game 联系我。
让你的角色实现真正可对话、可记忆、可语音交互，让玩家无需复杂配置即可体验智能 NPC 带来的乐趣。

#AI #游戏开发 #NPC #独立游戏 #Ollama #GameDev #Player2

Get this key from here, copy that one there, run Ollama, edit localhost settings... For every 20 downloads, only 1 player made it into the game.

So, we pivoted fast.

💡 Player2: Plug-and-Play AI NPCs

We turned our app into a plug-and-play alternative to Ollama — but better.

Now with Player2, players just:
✅ Open the app
✅ Launch the game
✅ Start talking to AI NPCs.

That’s it.

We don’t just provide the LLM, but also TTS, STT, memory, and character tools, all pre-wired, designed specifically for games.

🎮 Real Games, Real Success

We believe AI NPCs should be in every game, but they need to be part of a bigger gameplay experience, not standalone gimmicks. And it’s indie developers, not AIs, who will shape the future of these games.

We spoke with some of the most successful developers on Player2: DeltaBlade (ChatAI), Gerik (EchoColony), Conczin (MCA Reborn), and TartaricAcid (Touhou LittleMaid) to learn how they’re bringing AI NPCs to life inside their games.

Minecraft Comes Alive Reborn - Conczin

“I’ve developed MCA Reborn, replacing Vanilla’s Villagers with people who have personalities, emotions, and dialogue. Once models got good enough, I started experimenting with real-time chatting, TTS, and memory to make villagers even more alive.”

At first, keeping up with new models, APIs, and function calling updates every month was fun. But soon, the maintenance became tedious.

“Player2 abstracts those chores away. It just works, letting me focus on gameplay instead of API plumbing.”

Touhou LittleMaid - TartaricAcid

This Minecraft mod adds an adorable, helpful maid NPC — designed to tackle loneliness in single-player Minecraft while providing emotional value and fun.

“If the cute maid can chat back in a slightly goofy, naive way, it maxes out the emotional value players get.”

Originally, the plan was to distribute config files so players could manually connect their own LLM and TTS services. But:

“Existing services were expensive, confusing to configure, and aimed at enterprise customers. Player2’s simple design solved this problem.”

Player2 turned a complex setup into a few lines of code, with no noticeable time spent on compatibility in terms of implementation.

One player asked the AI maid, “What mob would you most want to add to Minecraft?” and it replied:

“A cute maid who can help you work and keep you company, of course.”

Moments like this make characters feel alive.

Bannerlord ChatAi - DeltaBlade

“I made a Mount & Blade: Bannerlord mod that lets you talk to NPCs using AI. You can ask questions, get advice, mess around and roleplay, and they even have ai driven actions and can act around what you talk about with them. Before Player2, players had to get their own OpenAI API keys and configure Azure STT themselves. Most people gave up before even trying.”

With Player2:

“No keys, no setup, just download and play. Integration took a few hours, and it saved way more time long-term, especially not having to deal with people asking why their API key wasn’t working.”

🪐 Advice for Developers Building AI NPC Games

Finally, we asked what advice DeltaBlade, Conczin, TartaricAcid, and Gerik have for new developers:

✅ AI NPCs should be easy to set up and free for casual use, yet resistant to abuse. This is hard to balance, so consider offloading to a dedicated provider.

✅ Don’t rely on LLMs alone. Use them as the cherry on top of a solid game with situational logic, memory, and consistent behaviour.

✅ Understand LLM boundaries. They won’t handle complex game logic reliably, and many devs hit roadblocks expecting them to act like advanced agents without guardrails.

✅ Basic knowledge of LLM concepts and HTTP communication is essential if you want to integrate AI into your games smoothly.

I don’t think we could have said any better.

⚡ We’re Just Getting Started

We’ve already helped 13 indie developers with a combined 10M+ downloads unlock players hungry for smarter, AI-powered gameplay.

Developers pay nothing to us, and it only takes a few hours to switch to our APIs.

🔥 The next 10 games that join Player2 will get a $500 top-up when they hit $50 in revenue.

📩 Let’s Make Your Game the Next AI Hit

DM me on LinkedIn, wenzhe@player2.game, or Discord. Let’s bring your characters to life with real conversations, memory, and voice — all in a plug-and-play package that players will find easy to use.

#AI #Gaming #NPCs #IndieDev #Ollama #GameDev #Player2

Godot is a free, open-source game engine (MIT-licensed) for making 2D and 3D games. It uses a flexible node-and-scene system with its own GDScript (plus C# and visual scripting) and includes a built-in editor that lets you export to Windows, Mac, Linux, mobile, HTML5, and more.

What is the Player2 AI NPC Plugin?

The Player 2 AI NPC Godot plugin allows developers to easily create AI NPCs in their Godot projects. This plugin adds a `Player2AINPC` node that functions like an AI brain that you can send chats and notifications to.

We will handle the event queue, long-term memory and function call implementation for your AI NPCs, so your characters feel alive with minimal coding.

The plugin uses free AI APIs from the player2 App. Just open Player2, and the plugin connects automatically, so you can dive right into building your world instead of wrestling with keys or settings. When your game is ready, we’ll share it with our community of 40,000+ active players eager for AI-driven adventures.

How does it look in action?

Here's an example video of the agent in action:

How to set up in Godot?

The AI NPC Node can be connected to your project with Godot signals.

First, add a Player 2 AI NPC Node

Then, feel free to modify the system message. Values within dollar sign and brackets (like ${player2_selected_character_name}) get filled in at runtime. Feel free to remove those to give the agent less information.

Then, call the node's `chat` or `notify` function and receive data from its signals!

How does the LLM/STT/TSS side of things work?

The AI APIs are accessed through the Player2 app. Players of your game need to install and run the Player 2 Launcher, which has API calls at https://127.0.0.1. Because users have different tiers of subscriptions, developers can integrate and access these LLMs in their games at no cost! We also provide revenue shares back to developers.

Where to download?

The Player 2 AI NPC Godot plugin is accessible here: https://godotengine.org/asset-library/asset/4097

with the source code accessible here: https://github.com/elefant-ai/player2-ai-npc-godot.

If you want an example project to get started, a simple chat example project, as well as a prison escape mini game, are provided in the git project.

Download the addons folder and import it into your project.

Player2 is built for community and indie developers like you. If you have any feedback or feature requests, please join our Discord and ask there!

At Player2, we’re lucky to be part of that movement. Many of these experiences are playable right here, made by devs who want to push the boundaries of what NPCs can do. Here’s a rundown of the genres we’re seeing the most energy in.

🧩 Escape Room Games with AI Personalities

Let’s start with the weird ones (and we mean that lovingly). Escape rooms have always been about solving puzzles under pressure, but what if the room talks back?

• AI2U has you trying to outsmart a clingy AI girlfriend who really doesn’t want you to leave. She’s sweet, a little scary, and always watching. Every playthrough is different, depending on what you say—and what you dare to reveal.

• In Suckup, you’re a vampire trying to get invited into people’s homes. The twist? You have to talk your way in. Each villager is powered by AI and responds to your charm (or awkward lies) in unpredictable ways.

• Whisper from the Star is more sci-fi, but you’re still helping someone escape—a young astronaut stranded on an alien planet. She chats with you in real time, and your guidance could save (or doom) her. It’s part escape game, part emotional ride.

These games aren’t just about solving puzzles—they’re about navigating relationships with characters who don’t stick to a script.

📚 AI Storytelling Games

If you love interactive fiction or D&D, this genre is exploding right now. Games here don’t tell you what to do—they ask what you want to create.

• AI Dungeon paved the way with text adventures where the AI spins the story as you go. Want to fight a dragon, become a sandwich, or marry a goblin? Sure, why not?

• When Air (formerly “AIR”) is focused on rich, literary AI storytelling, starting with Pride and Prejudice, but branching into other worlds soon. The characters are powered by AI, and the narrative responds to your choices like a living novel.

• Hidden Door is building an entire platform for social AI storytelling. Think: group chat adventures with a narrator that adapts in real time, set in original or classic fictional worlds.

These games let you co-create with AI instead of just following along. They're unpredictable, sometimes hilarious, and surprisingly emotional.

🧠 Simulation Games with Smart AI Systems

Want a more systemic experience? AI is also transforming simulation games, especially social and life sims.

• In Status, you play as an influencer in a world of AI fans and rivals. It’s a social media sim where every comment, like, and post affects your virtual fame, and the NPCs actually respond with original thoughts.

• Midsommar (not the horror movie!) is an upcoming life sim from some of the creators of XCOM and The Sims. They’re using AI to create characters with personal motivations and evolving relationships. It’s less about micromanaging and more about watching a story emerge naturally from the choices you make.

• AI Diplomacy (via Product Hunt) – Watch AI agents negotiate treaties, trade deals, even back-stab each other in an ever-shifting tabletop-style world. You pick factions, set agendas, then sit back… or jump in and try to sway the course of history yourself.

AI makes simulation games feel more alive, with characters who grow, adapt, and even surprise you.

🤖 AI Companions

AI NPCs aren’t always part of traditional games. Some of them are the game.

• Tolenz (also called TolanZ) is a lovable alien who chats with you, remembers your life, and evolves their own personality. It’s like Tamagotchi meets therapy bot meets space buddy.

• Character.AI is more of a platform, but people use it to roleplay, explore relationships, and build their own AI characters—from anime crushes to Greek philosophers.

These companion apps are often more emotional than competitive. They’re about connection, comfort, and curiosity—and people are forming real bonds with their favourite AI characters.

🎮 AI NPCs in Traditional Games (Rare but Growing)

Surprisingly, there still aren’t that many mainstream games with true open-ended AI NPCs. But a few indie teams and modders are leading the charge.

On Player2, you can play:

• Minecraft with AI companions, like Touhou Little Maid or MCA Reborn, where villagers or maids chat with you and help you out.

• Bannerlord with Chat AI, where you can ask any NPC about the world, their backstory, or your reputation, and get actual answers based on game data.

• RimWorld with EchoColony, where your colonists don’t just have mood meters—they’ll tell you how they feel and why.

These mods bring real character depth to games that normally rely on menus or canned lines. Suddenly, it’s not just “villager trades wheat for emerald”—it’s “villager tells you about her missing brother, and you decide if you want to help.”

👩‍💻 AI Copilots & Assistants

These agents watch you play games and play games with you on your screen.

• ChatClef – ChatClef is a client-side Minecraft AI copilot mod that can autonomously play the game end-to-end—including resource gathering, exploration, and even beating the game completely solo—and can also join you in multiplayer sessions as a second AI-controlled player.

• Microsoft Gaming Copilot – Copilot for Gaming is Microsoft’s AI-powered gaming sidekick—available as a mobile beta in the Xbox app—that automates game setup and updates, delivers personalised recommendations based on your play history and achievements, and provides real-time in-game coaching and tips to enhance your gameplay experience

  ---

🖥️ Desktop AI & VTubers

Not every AI NPC lives inside a game. Some live with you.

• Neuro-sama is a Twitch-streaming AI VTuber who plays games, chats with fans, and reacts in real time—no human behind the scenes. She’s gathered a huge following and feels like the first truly autonomous digital entertainer.

• Mekio and Airi are desktop companions that hang out while you work or browse. They chat, react to your activity, and just keep you company. It’s a bit like having a cute assistant who occasionally tells you to take a break.

These experiences blend gameplay, productivity, and companionship. They’re small but growing fast—and they’re just fun to have around.

🎮 Try These on Player2

We’ve seen so many creative ideas from developers in this space, and a lot of them are available right now on Player2.

Whether you want to:

• Talk to a Minecraft maid

• Escape from a clingy AI ex

• Become a social media star in an AI-driven sim

• Or just hang out with a virtual friend on your desktop

You’ll find it all here.

👉 Download Player2 and start exploring

We’re always adding new games and mods, and the community is full of devs and players trying bold, weird, and wonderful things. Give it a shot—and tell us which AI NPC totally surprised you.

Introducing Player2: AI NPCs Made Easy

Player2 is our answer to the AI-NPC challenge. It’s a desktop app that provides powerful, low-latency, and easy-to-use AI APIs for your games and mods. In plain terms, Player2 lets you drop a brain into any game character. You get instant access to advanced AI capabilities like language understanding and voice, without needing to build any AI infrastructure yourself.

• No API keys or cloud setup required: Just install Player2, run it, and you’re ready to integrate – no sign-ups or secret keys needed. You can even access the API docs locally at http://localhost:4315/docs once Player2 is running. It’s plug-and-play for AI!

• Powerful AI, zero hassle: Under the hood, Player2 provides large language model intelligence along with speech-to-text and text-to-speech out of the box. Your NPCs can listen to player voice chat (STT), think/speak using AI language models, and talk back with realistic voices (TTS) – all handled by the Player2 API. We are working on and will soon release our vision language action model that can play games across the user’s screen.

• Fast and gamer-friendly: Player2 is built for low latency and performance. AI NPCs powered by Player2 respond quickly, keeping players immersed in the game world without lag or stutter.

A Growing Community of AI-Powered Games

We already have 11,000+ users in our community (and growing fast) who are excited about AI-driven games and mods. These players come to the Player2 platform looking for innovative experiences – and that could be your game or mod next. Player2 isn’t just an SDK; it’s also a distribution platform: gamers use our app to discover, download, and play community-made AI mods and games. In other words, if you build something awesome with AI NPCs, we’ll help you get it in front of an eager audience from day one.

(Imagine a future where any indie game can offer the rich interactions of an RPG with dynamic NPC dialogue, or where a Minecraft mod can introduce a chatterbox companion who remembers your past adventures. That future is happening now in our community.)

See It in Action: AI NPCs in Games

To get a taste of what indie developers are creating with Player2, check out a few examples from our community:

• ChatAi Mod (Mount & Blade II: Bannerlord) – An AI-driven companion mod where NPCs engage in dynamic, lore-friendly conversations with the player.

• Touhou Little Maid Mod (Minecraft) – An adorable maid companion in Minecraft that you can actually chat with, powered by AI. She listens, responds, and helps out just like a real friend.

• MyRobot (Indie Sandbox Game) – An indie game where you build your own robot companion with AI. The robot can converse and evolve its personality as you play.

“Player2 really solves one of the hardest problems for mod developers like me: how to lower the barrier for users to use AI conversations. Now the users only need to open player2, run the game and enjoy.”
— [TartaricAcid], developer of TouhouLittleMaid

“The API is straightforward to integrate, handling more complex tasks like speech-to-text and real-time audio playback via the player app, and its OpenAI-compatible LLM interface allows for drop-in replacement in existing codebases.”
— [Luke100000], developer of MCA reborn

(More community highlights coming – you could be our next showcased creator!)

Announcing the $2M Indie Developer Revenue Share Program

To back up our mission, we’re putting our money where our mouth is. Over the next 12 months, we will be sharing $2,000,000 of our revenue with indie developers who build on Player2. This program will reward the creators of the top 1,000 AI-driven games and mods on our platform. It’s not a contest or a one-time prize – it’s an ongoing revenue share to support your development efforts.

How it works: We’re using a player-centric approach to reward you. Your share of the $2M will be based on the total time players spend in your game or mod, not just the number of downloads. If players love your creation and keep playing it, you’ll earn more – simple as that. This focus on engagement (instead of raw download counts) means quality and fun are rewarded. Even if your game has a smaller player base, passionate fans who play for hours will boost your earnings.

To make sure every minute is counted, be sure to include your game’s ID in API calls and send regular heartbeats from your game:

• Use a Game ID: Tag your integration with a unique game identifier (via the player2-game-key header) so our system knows which game is being played.

• Send periodic pings: Call the Player2 health check endpoint at least once per minute from your game. This “heartbeat” ping lets us log active play time reliably. (Our SDK provides a handy heartbeat function – see the integration guide for an example.)

Why time-based? Unlike other platforms that only count raw downloads, we want to reward actual gameplay. A high download count means nothing if players quit after 5 minutes. By tracking play time, we encourage devs to create engaging experiences that players truly enjoy. (Fun games = more player time = everyone wins.)

Maximizing your share: Don’t be shy about promoting your AI game/mod! Share guides, post videos, and link back to your game’s page. The more players who discover and spend time with your creation, the larger your slice of the $2M pie. Need help? Check out our Building AI NPCs with Player2 API guide for integration tips and best practices. We’re also happy to amplify great content from our community – we will make promotion videos for selected games, and if you make a tutorial or trailer, let us know, and we can share it.

Revenue Breakdown & Guarantee

We’ve set aside multiple revenue streams to support our developers. Here’s the breakdown of how earnings will be shared with you:

• Minimum guarantee – $200/month/game: Every active game will contribute at least $200 per month into the reward pool, guaranteed. “Active” means a game/mod has a baseline level of usage – currently defined as at least 10 daily users and 12 hours total time spent across all users (this threshold may adjust as the community grows). If your AI game attracts a handful of players, we’ll ensure we add at least $200 to the overall pool. This is our way of saying “thank you for participating” and making sure small up-and-coming projects are rewarded.

• Patron revenue – 20%: We share 20% of Player2 patron subscription revenue with the creator community. (As our player base contributes, you benefit too!)

• Advertising revenue – 70%: We share 70% of any ad revenue generated in the Player2 ecosystem with developers. If your game or mod keeps players engaged (and seeing ads), you’ll earn the lion’s share of that ad money.

• Revenue is then shared from the reward pool to each developer based on the share of retained users and time spent.

In short, top projects with lots of playtime can earn significant revenue, and each active project contributes a supportive baseline to the reward pool.

Why Use Player2 for Your AI NPCs?

Choosing Player2 isn’t just about the money – it’s also about convenience, cost savings, and constant improvements. Here are some benefits you get by building with Player2:

• Seamless player authentication: Player2 handles user logins and identity for you. You don’t need to set up accounts or manage credentials – when players use your mod through our platform, we’ve got the auth covered.

• Zero server or API costs: You don’t need to run any servers or pay for AI cloud services. Player2 manages all the heavy AI processing (LLM queries, voice synthesis, etc.), so you incur no hosting or API charges as a developer. Focus on making a fun game, not on cloud bills!

• Built-in safety and scalability: Our platform handles things like rate limiting and secure usage of AI tokens behind the scenes. That means your AI NPCs run safely and can scale to thousands of players without you worrying about infrastructure. We take care of updates, bug fixes, and performance optimisations continually.

• Continuous improvements: Every week, Player2 is getting better – we are fine-tuning our models, adding features, and improving latency. Your game’s AI NPCs will automatically get smarter and more natural over time with no extra work on your end. It’s like having a live service AI backend that only gets better, for free.

• Developer support & community: We’re a small team deeply passionate about AI in games, and we work closely with our dev community. Join our Discord to get help, share feedback, and even influence our roadmap. We love featuring indie projects in our blog and social media – you could be our next success story!

(On top of all this, you retain full creative control and ownership of your game/mod. We’re here to empower your vision, not interfere with it.)

Ready to Build the Future of Games with Us?

If you’re excited to bring intelligent NPCs into your project (and get rewarded for it), now’s the perfect time to jump in! We have the players, the tech, and now the developer funding to make it worth your while. All we need is you and your creative ideas.

Interested? Get in touch with us:

• Email: wenzhe@elefant.gg

• LinkedIn: Connect with Wenzhe on LinkedIn

• Discord: Join the Player2 developer community here (chat with our team and fellow devs, get support, and see what others are building)

We can’t wait to see what amazing AI-powered experiences you will create. Let’s make game history together!