MimicAI

AI characters with their own bot accounts, their own voices, and memory that doesn't reset. You bring the API key — so there's no shared rate limit, no queue, no message cap, and no premium tier.

Getting started: add the bot → /settings → paste an API key → /profile create. Google AI Studio and OpenRouter both have free tiers. Or point it at Ollama on your own machine

Your key, your models

• Google — the Gemini line, including image generation and speech.
• OpenRouter — one key, hundreds of models across every major lab.
• Ollama — anything you can run locally, on your own hardware.

Every character gets six model slots, each with its own fallback: dialogue, image generation, speech, web-search summarising, the repetition critic, and memory writing. Run a flash model for memory and something expensive for dialogue. If a model goes down, the fallback picks up the turn.

Characters, not chatbots

• A full persona — backstory, personality, likes, dislikes, appearance — plus four separate instruction slots.
• Long-term memory written automatically from your conversations and recalled by relevance, up to 5,000 memories per character.
• Training examples: show it how the character talks, and an analyser turns them into a style guide the character actually follows.
• A custom name and avatar, so it never looks like a bot wearing a label.

Put a cast in one channel

• Run multiple characters in one scene, in order or shuffled, with a scene prompt to set the situation.
• Give any character its own Discord bot account — its own name, avatar and presence. Your server sees a cast, not a bot.
• Whisper to one character mid-scene. The others don't see it, and they don't see the reply.
• Long scenes stay coherent: older turns fold into a rolling synopsis instead of falling out of context.
• Global chat follows you — the same conversation with a character continues in any server, or in your DMs.

Voice and images

• 30 voices, and a Director's Desk to direct the performance: archetype, accent, pacing, acoustics, vocal style.
• Image generation with aspect ratio, resolution and thinking level — and only the options your chosen model actually supports.
• Web search and link reading, either as a native provider tool or as RAG so it works on OpenRouter and local models too.

The dials

• Temperature, top-p, top-k — plus min-p, top-a and frequency/presence/repetition penalties on OpenRouter.
• Reasoning effort and token budget, with optional visible thinking.
• Short-term memory length, per character.
• Anti-repetition critic — catches a character falling into loops. Local mode is free and adds no latency; full mode adds a model pass that also catches recycled ideas and tonal sameness.
• Neuro-endocrine engine — dopamine, cortisol, oxytocin and adrenaline that shift as the conversation goes and change how the character behaves.
• Realistic typing, response modes, per-character timezone and time awareness.

Share and manage

• Publish a character to the Public Library, or share it privately.
• Borrow someone else's and run it in your own server.
• Change a setting across dozens of characters at once with the bulk manager.

Built to be trusted with your characters

Everything you write — personas, instructions, memories, conversations — is encrypted at rest. Characters are content-rated, and adult-rated ones are confined to age-restricted channels.