MimicAI

MimicAI is an advanced multimodal orchestration layer disguised as a single Discord bot. Instead of being limited to one personality, it acts as an engine for up to 50 unique, fully customisable profiles per user. You can build a tavern full of fictional characters, a team of assistants, or an entire OG crew, and have them all interact seamlessly in the exact same channel.

The Profile System At its core, the bot relies on a dynamic webhook technique. When a profile speaks, the system instantly generates a webhook to perfectly match that specific character's designated display name and avatar. For users who want a permanent presence, the system supports Child Bots. You can link a profile directly to a separate Discord bot application token, and the main MimicAI instance will orchestrate it remotely, placing actual bot accounts in your server's member list.

Deep Customisation Each profile is an isolated container. You define their backstory, personality traits, and physical appearance. Beyond the persona, you have absolute control over their mechanics. You can tweak specific generation parameters (temperature, top-p, presence penalties), enable realistic typing delays, or give them timezone awareness. You can even toggle web search capabilities (grounding) or allow them to scrape URLs sent in chat to read articles and documents.

Model Agnosticism You are not locked into one AI. Through Google Gemini and OpenRouter API integration, you can assign a different language model to every single profile. You can have a massive, high-tier reasoning model handling your complex narrator, while a cheaper, faster model pilots a quick-witted side character. If a primary model fails or hits a rate limit, the system automatically routes the prompt to your designated fallback model to keep the conversation moving without interruption.

You simply provide your own API keys and work from there. You are billed based on the API provider for what you generate. This bot does NOT require you to purchase a subscription to unlock features - it's all given to you for free.

Two Distinct Session Modes Interactions happen within isolated channel sessions, split into two modes:

Regular Sessions: You build a specific cast of characters (up to 10) for a channel. You can set a hidden Session Prompt to establish the scene behind the scenes. Users can control the flow entirely through reaction triggers on the bot's messages, allowing you to skip turns, mute messages from the AI's memory, force regenerations, or manually advance the round.

Freewill Mode: True automation. Admins can set channels to "Living" or "Lurking". Profiles are configured with specific personality percentages and wakewords. In this mode, characters will monitor the chat and randomly interject, argue with each other, or proactively start entirely new scenes without any user prompting based on probability rolls.

Narrative Command Tools The bot comes with a suite of utility commands to control the story:

Whisper: Send a hidden, ephemeral message to a specific character during a group session. They will react to your secret in their next public turn without revealing how they know it.

Speak: Administrators can manually puppet a profile, typing out exactly what they want the character to say to steer the plot.

Global Chat: Take any of your profiles into a private Direct Message for a persistent one-on-one conversation that maintains its own distinct memory separate from server roleplays.

Technical Architecture Under the hood, MimicAI uses Matryoshka Representation Learning (truncated 256-dimensional embeddings) paired with cosine similarity vector searches to handle Long-Term Memory (LTM) and contextual training examples. Conversations are automatically summarised by lightweight background models and stored to disc. When a user sends a message, the system maps the semantic intent, fetches relevant past memories, and injects them via strict XML tagging into the context window to prevent prompt hallucination. The Child Bot system utilises an asynchronous WebSocket Inter-Process Communication (IPC) layer, allowing the primary Hivemind to route complex, multi-chunk generation payloads to blind worker clients with near-zero latency.

And this is just a fraction of what this application is capable of, all while being completely free to use.

The entire codebase is AI-generated. If you find ANY bug, report it to me (@latteboi).