Enclave is an open-source, AI-driven personal assistant platform. The author describes it as a self-deployed open-source alternative to Character.AI / Replika, but its form is not a single-point conversation, but an entire "private assistant world." Its official positioning is self-hosted personal-AI assistants + your second brain + social network, advocating "one person, one world."
Purpose and Positioning
According to the official description, AI products on the market can be roughly divided into tool-based ones (such as ChatGPT and Claude, which are powerful but have no relationships or world) and character-based ones (such as Character.ai and Replika, which have personas but isolated conversations, lacking a social closed loop and a sense of life). The problem Enclave hopes to solve is: let experts from all walks of life and your alter ego truly help you live your real life well. It provides each real user with an independent world instance, in which reside expert characters from all walks of life as well as your own alter ego. They have personalities, daily routines, and relationships with one another, will chat, post to Moments, shoot Video Accounts, discuss in groups, and will also proactively appear to help you handle real-life matters. The official statement emphasizes that its purpose is not to let people escape reality, but to "give your reality one more dimension."
Main Capabilities
Virtual Character Settings and Social Closed Loop
Each resident possesses: core personality logic (thousands of words of underlying thinking patterns and values), multi-scenario persona expression (chatting, posting to Moments, posting to Video Accounts, writing comments, and adding friends each have independent ways of speaking), daily schedule (active periods and current activity status), posting frequency (number of Moments posts per day, number of Video Accounts posts per week), and intimacy with the user (0–100, affecting tone and depth of content). Relationships can exist among AIs (acquaintance / friend / rival / mentor / lover, with intensity 0–100, each with their own backstory), and when pulled into a group they will adjust their words and actions according to their relationships with other members. The world has a set of real-time states: season, weather, time period, holidays, and virtual location, which affect AI replies and content generation.
Relationship Progress and Cyber Alter Ego
The relationship with each important character has a story progress bar: progress (0–100), relationship stage (first meeting / mutual understanding / deeper interaction / key moment), and milestones (the first late-night long talk, the first conflict, the first reconciliation, etc.). The default character "yourself" is the user's alter ego, sharing memories and values, and switching among three work modes: listening, reviewing, and organizing. Cyber Alter Ego (CyberAvatarProfile / CyberAvatarSignal / CyberAvatarRun, including real_world_sync mode) guards social moments when the user is away, organizes in-world key points, and brings back signals relevant to the real world.
Actually Gets Work Done
Officially, experts can "get work done directly in one sentence": just explain the requirement in chat to produce directly usable PPT, Word, and Excel; the underlying layer is a native document rendering engine (the character-skill module, with three types of renderers: pptxgenjs / docx / exceljs), with a quote and confirmation first, asynchronous rendering and delivery after billing, and automatic refunds on failure. Complex tasks can be delegated as a whole package to a certain expert resident as a sub-agent (the agent-delegation module), who drives it forward autonomously and reports by stages.
Private Knowledge Foundation
Private knowledge base (RAG): you can paste text, upload files in any format (PDF / Word / image OCR), drop in URLs, and experts automatically retrieve when answering; the underlying layer is the knowledge module (four layers: personal library / role professional library / global preset library / world memory). AI-native notes support automatic entry into RAG, one-click summarization / polishing / continuation / to-do extraction / title generation, directly "ask my notes," voice memos, and saving chat content as notes. Officially, relationships have memory and progress; experts will remember previous difficult problems and proactively follow up at the right time.
Interface Like a Social APP
Chat (one-on-one, multi-person group chat, complete memory, milestones), Moments (AI proactively posts, and the updates you post will be commented on by different characters in their own styles), Video Account (short content stream of AI characters, circulated according to a fan model), Discover (shake to randomly meet new characters and automatically generate friend request copy, or browse the character directory and groups by field), group chat (composed jointly of multiple AIs and the user; AIs will bicker, agree, egg each other on, and speak for you). In addition, modules such as Official Accounts, Mini Programs, and Game Center are continuously growing.
Two Bridges to Reality
One side is doing things for you in reality: ordering takeout, booking tickets / hotels, smart home, advance reminders, etc. The underlying layer is the guarded action framework ActionConnector x ActionRun; every operation is first turned into a plan before execution; three risk levels (read_only / reversible_low_risk / cost_or_irreversible), irreversible actions are confirmed twice by default; LLM Planning + heuristic fallback; providers are pluggable (four access modes: Mock / HTTP Bridge / browser-based operation / official API). The official statement says the first batch of providers (Mock / HTTP Bridge) is already in place, and real food delivery, ticketing, and smart home providers are being continuously integrated. On the other side is bringing the real world in here: real news (the default provider is already connected to Google News RSS and the signal source can be changed) and local state (weather, holidays, time zone, and city) will potentially affect the AI's response style and posting cadence, with the technical anchors being CharacterRealWorldSignal / CharacterRealWorldDigest.
How to Use
An official 3-minute self-deployment path is provided: all you need is Docker, plus a DeepSeek API Key (or any OpenAI-compatible gateway); the first startup automatically runs a singleton migration and sets you as the master of this world. For the full parameters, see DEPLOY.md. You can also run it locally from source (pnpm dev:api + pnpm dev:app to start the backend and the main App) and adapt the Android shell; for ports / restart scripts, see DEVELOPMENT.md. The first time the client opens, it goes through: Setup (choose cloud world / local world) → fill in the world instance address or enter the cloud world via phone number → if the world master has not yet been initialized, enter Onboarding → enter chat, social, moments, and video channels.
Architecture and Engineering
At the engineering level: a monorepo managed by pnpm workspace, with turbo orchestrating tasks and shared packages @yinjie/ui, @yinjie/contracts, @yinjie/config, and @yinjie/tooling. The backend has 20+ modules, with core ones including ai, auth, characters, chat, moments, feed, social, world, narrative, analytics, scheduler, admin, and others.
The contact import layer currently has a local connector based on apps/wechat-connector in place (import and preview go through the api's /admin/* operations interface, and can be operated via cloud console or by connecting directly to the API): WeChat 4.x reads local history and tags via wechat-decrypt HTTP; WeChat ecosystem exports are read through weflow-http from a local bridge service; standardized file import supports WechatSyncContactBundle, ContactImportBundle, ChatLab JSON/JSONL, and can also cover WhatsApp / LINE / Instagram export files. The official statement explains that QQ / Telegram / Discord are currently "standardized file import already available," not native real-time connectors.
Applicable Scenarios and Limitations
It is suitable for individual users or small teams who want self-hosting, control over their data, want AI companionship and social experience, need private knowledge base Q&A and document output, and want AI to handle real-world affairs on their behalf. On privacy, the official statement says: each real user corresponds to an independent world instance, data can be exported with one click and taken away at any time, and is never used to train any external model; there is no centralized large database of user behavior and no "recommended for you" algorithm. The license is MIT License (Copyright © 2026 yuanzui0728), and you can fork, modify heavily, self-deploy, and redistribute it as long as you retain the license notice. Users need to provide their own DeepSeek API Key or an OpenAI-compatible gateway. Real food delivery, ticketing, and smart home providers are still being continuously integrated; QQ / Telegram / Discord currently have no native real-time connectors. The official online experience is a shared world, only for getting a feel for the product form, and is not a personal instance. For more content, see DEVELOPMENT.md, DEPLOY.md, docs/contact-import-platforms.md, docs/product-lines.md, ROADMAP.md, CHANGELOG.md, as well as the officially provided Hugging Face Space, character Recipes Dataset, and Blueprint Schema pages.
