Powerful AI Integration mod

Powerful AI Integration adds a broad AI layer to RimWorld: living pawn dialogue, text-to-speech, an AI storyteller, conversation and relationship memory, story arcs, story characters, negotiations, AI pawn control, prayers, faction context, and persona tools.

The point is not just to attach a chatbot to RimWorld. The mod tries to gather real game context, understand who is speaking, what is happening in the colony, what relationships already exist, what happened recently, and only then ask the model to write a line, judge a request, or choose an action.

Goal: the colony should feel alive, remember the past, and react to the present.

• Living pawn dialogue based on real in-game events.
  
• Colony chat, dialogue journal, lightweight overlay, ultra-light overlay, and Interaction Bubbles support.
  
• Built-in text-to-speech.
  
• AI storyteller: Powerful AI Director.
  
• Memory systems: conversations, relationships, rumors, conflicts, promises, story arcs.
  
• AI pawn control through natural-language commands.
  
• Negotiations with factions, traders, and prisoners.
  
• Pawn prayers to the storyteller.
  
• Story characters, factions, genres, narrative roles, canon rules, and personality types.
  
• Persona tools: generate, improve, randomize, assign, and evolve pawn personalities.

You can add this mod to an existing save.

If you want to remove it from an existing save, use the safe-removal tool first:

1. Load the save with the mod enabled.
   
2. Open the mod settings.
   
3. Find Safe removal.
   
4. Click Prepare current save for mod removal.
   
5. Save the game under a new name.
   
6. Quit RimWorld.
   
7. Only then disable or remove the mod.

Make a backup first. Removing the mod without cleanup can leave missing mod data in the save and may cause red logs, UI issues, flickering, or visual distortion.
Required:

• Harmony

Optional:

• Interaction Bubbles - shows generated lines as bubbles above pawns.
  
• A local model or cloud API - required for AI dialogue, AI command judgement, and AI-driven event selection.

If no model is configured, the mod tries to stay out of the way. Some local rules and fallback systems can still work without a model, but living AI dialogue, AI control, AI negotiations, and AI storyteller decisions require a working endpoint.
Some of my other mods may require Powerful AI Integration as a dependency. For example, an AI sculpture mod may use shared infrastructure, but it does not require you to configure dialogue or the storyteller.

If a companion mod is detected, the first setup flow can offer:

• Companion mod only - the main AI buttons and gameplay AI features from Powerful AI Integration are hidden.
  
• Full AI integration - dialogue, storyteller, TTS, AI control, and other systems are enabled.

This lets players subscribe to the dependency without getting extra UI if they only want the companion mod.

After installation, open the mod settings. The simple setup flow can help you:

• choose a model source;
  
• enter endpoint, model, and API key;
  
• test the connection;
  
• enable or disable TTS;
  
• choose dialogue frequency;
  
• choose where dialogue appears;
  
• choose a performance preset;
  
• set initial player prompts and general style.

If the model is missing or the test fails, the mod will warn you. That is intentional: it is better to see the problem immediately than wonder why pawns are silent.
Many services and local programs can behave like the OpenAI API. This means the mod sends requests in a familiar format:
In the mod settings, you usually enter the base address, without `/chat/completions` at the end. Examples:

• LM Studio: `http://127.0.0.1:1234/v1`
  
• Ollama: `http://127.0.0.1:11434/v1`
  
• OpenRouter: `https://openrouter.ai/api/v1`
  
• Player2: `http://127.0.0.1:4315/v1`

If you enter the full path like `/v1/chat/completions`, some providers may fail because the mod adds that path itself.
The model setup page usually needs these fields:

• Provider - connection type: local, OpenAI-compatible, Player2, etc.
  
• Endpoint - model server address.
  
• Model - model name.
  
• API key - provider key, if required.
  
• Timeout - how long the mod waits for a model response.
  
• Temperature - how freely the model writes.
  
• Max tokens - maximum response size.
  
• Preset - overall load level: Low, Medium, High, Maximum, or Custom.

For a first run, start with Medium. If the model is weak or local, start with Low. If the model is strong and fast, try High or Maximum.
LM Studio is one of the easiest ways to run a local model.

1. Install LM Studio.
   
2. Download a suitable chat/instruct model.
   
3. Open the Local Server tab.
   
4. Start the OpenAI-compatible server.
   
5. Check the address. Usually it is `http://127.0.0.1:1234/v1`.
   
6. In the mod settings, choose Local/OpenAI-compatible.
   
7. Endpoint: `http://127.0.0.1:1234/v1`.
   
8. Model: the loaded model name from LM Studio.
   
9. API key can stay empty if LM Studio does not require one.
   
10. Press the connection test button.

For GPUs around RTX 4060 8 GB, start with 7B/8B models in Q4 or Q5 quantization. If responses are slow, reduce the mod preset, dialogue frequency, and context size.
Ollama is convenient if you are comfortable with a small amount of command line use. Example:

Mod settings:

• Provider: Local/OpenAI-compatible
  
• Endpoint: `http://127.0.0.1:11434/v1`
  
• Model: the model name in Ollama
  
• API key: empty, unless you configured authorization

If the model starts returning too much reasoning or thinking text, enable thinking-output cleanup in the mod or configure the model/server so reasoning is not returned in the final answer.
You can use almost any OpenAI-compatible API:

• OpenAI-compatible providers.
  
• OpenRouter.
  
• Gemini-compatible OpenAI endpoint.
  
• DeepSeek-style endpoints.
  
• NVIDIA or other providers, if they expose an OpenAI-compatible chat endpoint.
  
• Player2.

Free and free-tier limits change often. The mod includes in-game hints for where to find keys and how to enter endpoints. If a provider does not work, leave a comment with:

• provider name;
  
• endpoint;
  
• model name;
  
• error text;
  
• log snippet, if available.

The mod supports saving multiple model presets. This helps when:

• one provider hits its limit;
  
• a local model is too slow for large scenes;
  
• you want a weaker model for normal lines and a stronger one for storyteller work;
  
• you are testing different endpoints;
  
• one model writes better dialogue while another understands commands better.

Create multiple presets and switch between them instead of retyping endpoint and model every time.
The mod supports built-in dialogue speech. The easiest starting option is EdgeTTS. EdgeTTS:

• can work without an API key;
  
• is available directly from the mod;
  
• is good for first setup;
  
• supports different voices;
  
• can automatically choose voice hints by pawn gender and settings.

Other supported options:

• Local HTTP TTS;
  
• OpenAI-compatible TTS;
  
• Fish Audio;
  
• SiliconFlow CosyVoice2;
  
• SiliconFlow IndexTTS2.

TTS settings can control:

• whether speech is enabled;
  
• automatic speech for new lines;
  
• volume;
  
• provider;
  
• default voices;
  
• per-pawn voices;
  
• text and audio synchronization;
  
• line replay button.

If synchronization is enabled, the mod tries not to reveal the next line while the previous one is still being spoken.

Dialogue is one of the core parts of the mod. Pawns can talk about:

• current work;
  
• food, sleep, recreation, and needs;
  
• wounds, sickness, and tending;
  
• combat and threats;
  
• guests, traders, prisoners, and enemies;
  
• family, children, weddings, and relationships;
  
• the past, rumors, conflicts, and promises;
  
• storyteller events;
  
• story characters and personal rules.

The mod distinguishes social status:

• Colonist speaks like a member of the shared home.
  
• Prisoner remembers they are confined.
  
• Slave speaks with coercion and control in mind.
  
• Guest does not sound like the owner of the colony.
  
• Trader thinks about route, goods, and profit.
  
• Enemy does not become friendly without a reason.
  
• Mechanoid can be disabled or sound like a soulless machine.

Dialogue can be shown in several ways:

• Dialogue journal - detailed history with context.
  
• Colony chat - a larger message-stream window.
  
• Lightweight overlay - a translucent window you can keep open.
  
• Ultra-light overlay - only name and text, almost no UI.
  
• Interaction Bubbles - bubbles above pawns, if the mod is installed.

Overlay settings include:

• size;
  
• transparency;
  
• font size;
  
• message count;
  
• conversation separators;
  
• settings button;
  
• pause chat reveal while the game is paused;
  
• auto-scroll down;
  
• clicking a pawn name to select that pawn.

If pawns talk too often:

• increase the minimum auto-dialogue interval;
  
• disable monologues or increase their separate cooldown;
  
• reduce ambient line frequency;
  
• enable auto-disable at high game speed;
  
• limit which pawn types are allowed to speak;
  
• reduce guest/trader/enemy weights if they fill the chat.

If pawns are silent:

• check whether the dialogue system is enabled;
  
• check whether the model test passed;
  
• check cooldowns;
  
• check whether dialogue is disabled at the current game speed;
  
• check whether that pawn type is allowed to speak;
  
• open the journal and debug logs.

The mod tries to distinguish active and passive threats. Active threat:

• raiders are attacking;
  
• an enemy is nearby and can reach the colony;
  
• mechanoids are active;
  
• manhunters are attacking;
  
• real combat is happening;
  
• there are wounds, bleeding, or downed pawns.

Passive threat:

• shamblers are far away and slowly approaching;
  
• insects are sitting in a cave and are not aggressive;
  
• a mech cluster is dormant;
  
• siege raiders have mortars but are not storming the base;
  
• anomaly objects exist on the map but are not attacking right now.

For passive threats, pawns may discuss the plan: fight, wait behind walls, close the gate, prepare traps, or choose the right moment.

For active threats, combat mode gets priority: danger, wounds, cover, positions, retreat, rescuing downed pawns, and post-battle reactions.

Powerful AI Director is a storyteller that uses local logic and AI where appropriate.

It can consider:

• current colony state;
  
• wealth, food, weapons, wounded pawns, prisoners, threats;
  
• story genre;
  
• scene roles;
  
• scene goal;
  
• event cooldowns;
  
• canon rules;
  
• story arcs;
  
• promises and payoff;
  
• factions;
  
• prayers;
  
• past events.

AI does not get permission to create any event from nowhere. The mod first collects legal game options, then the model can help choose a fitting one. If the answer is invalid, too late, or unavailable, fallback logic is used.
Roles are story forces: home, threat, mystery, relief, consequence, temptation, and so on.

Genre sets tone: survival, family saga, dark fantasy, horror, comfort, war drama, etc.

Canon rules are hard rules the AI should respect. For example:

• do not add new relatives without permission;
  
• do not kill story characters through text;
  
• do not break an established romance;
  
• do not turn a cozy story into endless catastrophe;
  
• do not accept new animals if that is an important rule for your story.

Canon rules are especially useful if you are roleplaying a specific story.
In the Characters tab, you can:

• mark a pawn as an important character;
  
• add a personality description;
  
• set their story role;
  
• describe relationships;
  
• tell the storyteller how to treat that character;
  
• create a character without a current pawn, so they can appear later;
  
• pull pawns from the loaded game;
  
• use stats, health, mood, traits, equipment, and social bonds.

If a pawn has a personal prompt, the mod treats it as important. But it should not break game facts: a prisoner remains a prisoner, an enemy remains an enemy, and the dead should not suddenly speak like the living.
Persona tools help configure pawn speech and character.

Available options:

• Generate from pawn - generate persona from the current pawn.
  
• Improve current - improve the current prompt.
  
• Make more dramatic - make it more dramatic.
  
• Make more grounded - make it more grounded.
  
• Make funnier - add humor.
  
• Make darker - make it darker.
  
• Make calmer - make it calmer.
  
• Roll Gen - local random archetype without AI.
  
• Persona type - choose a personality type from the list.
  
• Evolve persona - evolve personality based on memory, trauma, relationships, status, and events.

You can create configurable personality types: Stoic, Pessimist, Romantic, Cynic, Fanatic, Quiet, Joker, Aristocrat, and others. A personality type sets speech style and habits, but it does not have to replace the whole pawn.

AI Control lets you give the selected pawn a soft natural-language command.

Examples:

• "Go to the north door."
  
• "Eat something decent."
  
• "Go to sleep."
  
• "Pick up the best bow."
  
• "Help the wounded pawn."
  
• "Talk to the prisoner."
  
• "Attack the nearest enemy."
  
• "Carry the downed pawn somewhere safe."
  
• "Try doing the work that matters most right now."

The mod can search for objects, pawns, food, weapons, beds, places, AI markers, reachable cells near the target, and options from the right-click menu.

For safety, AI Control does not do some things:

• does not create new zones;
  
• does not place new blueprints;
  
• does not create bills;
  
• does not change global work priorities;
  
• does not arrest or capture without explicit supported logic;
  
• does not perform lovin;
  
• should not send a non-combat pawn to fight if they cannot fight.

AI Control is experimental. If a command is understood strangely, send the latest requests through the built-in report button.
You can place a special marker on the map, name it, and use it in commands.

Examples:

• "Go to the altar."
  
• "Move to the backup exit."
  
• "Stand at the western gate."
  
• "Return to the meditation garden."

This helps when a normal place description is too vague or the map has many similar objects.
Through the comms console, you can write a message to a faction as the player.

The model evaluates:

• how persuasive the text is;
  
• relationship with the faction;
  
• colony situation;
  
• a little luck;
  
• negotiation difficulty.

Possible actions:

• improve relations;
  
• request a trade caravan;
  
• request aid.

Negotiation difficulty is configurable. Medium is suitable for normal play, Hard is closer to the early mod behavior, and Easy makes player diplomacy text stronger.
If a trader is nearby, you can try to earn a small discount.

The result depends on:

• quality of the player's text;
  
• faction relationship;
  
• trade context;
  
• some randomness;
  
• negotiation difficulty.

The discount is usually limited and should not break the economy.
You can talk to a prisoner through AI text evaluation.

Possible effects:

• reduce resistance;
  
• recruit when resistance is low enough;
  
• change ideological certainty;
  
• convert;
  
• reduce will for future enslavement.

Cooldowns prevent spamming the same text repeatedly.
A prayer is not a direct command to the storyteller. It is an in-world request.

It can:

• be heard;
  
• be ignored;
  
• receive a gentle answer;
  
• receive a dangerous or ironic answer;
  
• enter the prayer journal;
  
• influence a future scene.

If a pawn prays too often, lore-friendly limits can be enabled: gods get annoyed, the request becomes weaker, the pawn starts looking like a whiner, and so on.

Prayers work best when an AI storyteller mode is enabled.

The mod stores several kinds of memory:

• Story journal - storyteller events and their meaning.
  
• Conversation memory - what pawns have already said.
  
• Relationship memory - warmth, trust, tension, resentment.
  
• Rumors - important overheard conversations that may return later.
  
• Continuing conflicts - arguments that do not vanish instantly.
  
• Story arcs - long-term arcs.
  
• Promises and payoff - setup for future events.

To avoid bloating saves and prompts forever, the mod stores limited records and cleans old data.
Rumors are not simply "everyone heard everything." The mod tries to respect privacy:

• public conversations may spread wider;
  
• private conversations may be overheard by chance;
  
• secret conversations should surface carefully;
  
• important overheard scenes may become rumors the next day.

Rumors help the colony remember drama without turning every dialogue into a recap of the past.
The mod supports family scenes:

• pregnancy;
  
• birth;
  
• first words;
  
• growth moments;
  
• learning;
  
• school scenes;
  
• children's games;
  
• children drawing on the floor;
  
• parent-child conversations;
  
• wedding ceremonies;
  
• wedding congratulations;
  
• jealousy or awkwardness when the context supports it.

If a wedding happened in the colony, the mod can remember it throughout the day and use it even in other scenes.
The mod adds reactions to many ordinary moments:

• research;
  
• book reading;
  
• recreation and games;
  
• billiards, chess, walks;
  
• cooking and food shortage;
  
• the first good meal after hunger;
  
• food poisoning;
  
• successful or dangerous hunting;
  
• medical care;
  
• mechanitors and mechanoids;
  
• genes, xenotypes, vampires, psycast levels, and royal titles.

Ordinary scenes matter because the colony should feel alive outside raids too.

The mod contains system prompts and scene-type prompts. They can be edited if advanced editing is enabled.

Recommended:

• do not remove output format rules;
  
• do not break rules against inventing game facts;
  
• do not force everyone to speak in one style;
  
• do not use a huge general prompt for weak models;
  
• use story characters and persona instead of one enormous global prompt.

If you heavily edit the system prompt and the model starts behaving strangely, reset it to default.
The settings window has search. You can search by English and localized names.

There is also an AI settings assistant. It can explain where a setting is and what it does. It requires a configured model.

The mod has a problem-report system through a Cloudflare receiver. It is enabled only with player consent.

You can send:

• an in-game error;
  
• recent model requests and responses;
  
• debug bundle;
  
• AI control issue;
  
• invalid model response format.

Before sending, the mod tries to filter errors that clearly belong to other mods, so players do not blame this mod for unrelated red logs.
If something breaks, the best report includes:

• RimWorld version;
  
• AI provider name;
  
• endpoint;
  
• model;
  
• local model or cloud API;
  
• whether TTS is enabled;
  
• AI selection mode;
  
• what happened in game;
  
• what you expected;
  
• Player.log or red log;
  
• bad model response example, if the issue is text quality.

If the issue is dialogue-specific, the "share recent requests/responses" button helps a lot.
If dialogue is too bland:

• check persona for important pawns;
  
• add story characters;
  
• configure genre;
  
• enable conversation memory;
  
• add personal speech habits;
  
• increase dramatic range for default scenes;
  
• use a stronger model or increase prompt budget.

If dialogue is too strange:

• lower temperature;
  
• reduce the global user prompt;
  
• reset the system prompt;
  
• disable overly heavy context blocks;
  
• check whether a pawn has an unsuitable persona;
  
• check whether pawn status is detected correctly.

• Start with preset Medium.
  
• Use Low for weak models.
  
• Increase auto-dialogue cooldown.
  
• Use a separate cooldown for monologues.
  
• Disable dialogue at high game speed.
  
• Limit guests, traders, and enemies with weights.
  
• Do not enable huge prompt budgets unless needed.
  
• For local models, choose a model your GPU can actually run.
  
• If the request queue grows, reduce dialogue frequency.

The mod tries to move heavy work into background processing or split context gathering into smaller parts, but RimWorld is still single-threaded in many places. If you have a huge map and many mods, careful settings matter.

Do I need internet?

For a local model, usually not after the model is downloaded. For a cloud API, yes, you need internet while playing.

Do I have to pay?

It depends on the provider. A local model is usually free after installation but needs PC resources. Cloud APIs are often paid by tokens or limited by free tiers.

Does the mod work without a model?

Some local logic and fallback systems work. But AI dialogue, AI control, AI negotiations, AI prayers, and AI event selection require a model.

Can I add it to an old save?

Yes. But remove it from a save only through the safe-removal tool.

Why are pawns silent?

Check whether AI requests are enabled, the model test passed, dialogue is enabled, cooldowns are not blocking it, and dialogue is not disabled at the current game speed.

Why are responses delayed?

The model may be slow, the local server may be busy, the request queue may be full, or the timeout may be too low. Increase timeout, reduce dialogue frequency, or choose a faster model.

Why is the model answering in the wrong language?

Check the dialogue language setting. The default can use interface language, but explicitly choosing a language is safer if the provider gets confused.

Why does the model write reasoning/thinking text?

Some reasoning models return thinking output. Enable thinking-output cleanup in the mod or configure the model/server so reasoning is not returned in the final answer.

For a first test:

• AI preset: Medium.
  
• Dialogue frequency: not too frequent.
  
• Auto-dialogue disabled at high game speed: enabled.
  
• TTS: start with EdgeTTS or disabled if you only test text.
  
• AI Director: you can keep the vanilla storyteller first and enable only dialogue.
  
• AI Control: enable after dialogue is working.
  
• Memory: enabled, but do not set huge limits.
  
• Prompt budget: default.

Once the model responds well and without lag, enable more features.
Powerful AI Integration is a mod about a living colony. It works best not when every pawn talks endlessly, but when rare lines land in the right situation: a prisoner fears the cell, a trader judges the walls, a child talks over the radio, a doctor calms a patient, an enemy threatens at the gate, and after the battle someone finally exhales.

If something happens too often, too rarely, too calmly, too dramatically, or just feels strange, please comment. The mod is actively evolving, and real examples from real colonies help make it better.

Patreon[patreon.com]