Agnai

Within that cluster it has a recognisable niche: it is the one designed from the start around multi-user rooms, and it treats the AI backend as a swappable choice on every chat rather than a fixed runtime. The project lives at github.com/agnaistic/agnai and offers two ways in. The first is the hosted instance at agnai.chat, where anyone can register an account and start chatting in a browser without installing anything. The second is self-hosting: the same codebase ships with a Docker Compose file and a documented Node + MongoDB stack, so a small team can run a private instance on a VPS or a home server with full control over data, branding, and policy. The product centre of gravity is collaboration. A character is not just an object you talk to; it is something you can drop into a room and let several real people share, alongside one or several AI characters at once. That is what underpins the use cases the community gravitates toward: lightweight tabletop sessions where a game master and a few players riff against a shared cast of NPCs, collaborative writing where two authors push a story forward with the help of an AI scene partner, or simply a friend group hanging out with a virtual character in the same channel. None of the closed cloud apps in the C-cluster (Character.AI, Janitor, SpicyChat) ship that pattern as a native feature, and within the open-source cluster only Agnai treats it as a first-class workflow. The backend layer is the other major design choice. Agnai is bring-your-own everywhere, but the bring-your-own is unusually wide. KoboldAI Horde works out of the box as a free distributed inference backend powered by donated community GPUs. Beyond Horde, the same chat can be pointed at OpenAI, Anthropic Claude, NovelAI, an Oobabooga text-generation-webui endpoint, a KoboldCpp local server, or any compatible OpenAI-shaped API. Sampler settings, prompt templates, and model selection live at the chat level as well as globally, which means a single character can use Horde for low-stakes banter and switch to GPT-4 or Claude for a key plot beat without losing memory or persona. A note on positioning: agnai.chat (the hosted instance) follows community-set policy, while self-hosted deployments are unrestricted. The product is "optional-NSFW" rather than NSFW-headlined — that is, adult content is permitted depending on backend and configuration but is not the marketing pitch. The pitch is open source, multi-user, multi-backend, and run-it-yourself if you want to.

The default backend is KoboldAI Horde, which means you can send your first message without filling in an API key — Horde will queue the request against a community worker and stream back the reply. This is the right way to feel out whether the interface and the multi-user model match how you want to chat before you spend anything. Switching backends is the second step, because Horde latency depends on community load. Open Settings → AI Settings, pick the backend you want — OpenAI, Claude, NovelAI, your own Oobabooga URL, a KoboldCpp endpoint, or anything compatible — and paste the API key or URL. Settings can be saved globally or per chat, so it is normal to keep Horde as the default and override individual characters where quality matters more. Creating your first character takes another minute or two. From the left sidebar, choose Characters → Create Character, fill in name, description, an optional persona block, a greeting line, and an avatar image. Two things to know up front: the description field accepts the same structure as Tavern V2 cards (so anything you have read about prompt engineering for SillyTavern-style characters carries over), and if you already have a PNG character card from SillyTavern, Chub, Backyard or Janitor you can just upload it rather than start from scratch. Tavern V2 imports are the most reliable; V3 or proprietary formats may need to be re-exported. Starting a chat is then a single click. Open the character, type, and your selected backend handles the reply. The header lets you swap the backend mid-chat, regenerate or edit any message, and start a new chat thread off the same character if you want a parallel storyline. If you want to invite other people, flip the multi-user toggle on the character or chat: a shareable link is generated, and anyone with the link can join as a participant. Permissions stay with whoever created the room, so they can moderate who can speak as which character.

KoboldAI Horde is free, distributed across donated community GPUs, and the throughput you see depends on how many workers are running the model you picked at the moment you send your message; the more common the model, the better the experience. OpenAI gives the most consistent ceiling on dialogue quality and instruction-following, especially for longer chains of reasoning, but you pay per token. Anthropic Claude is well suited to long-context narrative work and tends to stay on character across multi-turn scenes, also priced per token. NovelAI is a subscription that includes its own narrative-tuned models; it is a fit for users who already pay for NovelAI and want to keep that style inside a chat front-end. Local backends — Oobabooga's text-generation-webui and KoboldCpp — turn a workstation or a GPU server into the inference layer; they remove ongoing costs and give the strongest privacy guarantees, in exchange for hardware investment and setup complexity. Characters in Agnai are Tavern V2 by convention. A character card carries the name, a description block (which doubles as persona and backstory), a greeting, optional example dialogues, and the avatar PNG. The card format is portable: SillyTavern, Backyard AI, RisuAI, Chub AI and Janitor AI all read and write the same PNG-embedded JSON, so a card built in one tool can usually be imported into another with no edits. Inside Agnai, each character can own multiple chat threads, which is how the same persona is reused across different storylines without contaminating history. Per-chat settings include backend, temperature, top-p, top-k, repetition penalty, and the prompt template used to format messages — Alpaca, ChatML, Vicuna, and similar. Multi-user rooms are where Agnai earns its niche. A room can host more than one real participant and more than one AI character. The room creator manages who can join, which characters speak, and which backends are active. Typical use cases are tabletop sessions with a DM plus several players and a small cast of NPCs, collaborative writing where two or three authors push the story forward with an AI as a scene partner, and informal hangouts where a friend group treats an AI character as a member of the channel. For multi-user scenes, stronger backends matter more than they do for single-player chat — coherence across many speakers is harder for smaller models — so the practical pattern is to run rooms on OpenAI, Claude, or a well-tuned local model rather than on Horde during a busy session.

Where money enters the picture is by your choice: a paid backend (OpenAI, Claude, NovelAI) bills you on its own pricing; a VPS or home server hosts the self-hosted version on whatever hardware you pick. Self-hosting is the path for teams that want their own instance, custom branding, no shared moderation policy, or a private deployment for a closed community. The official supported method is Docker Compose. The repo at github.com/agnaistic/agnai contains a docker-compose.yml that brings up the Node app and a MongoDB instance with one command; you supply environment variables for the public URL, optional backend keys, and optional authentication. Hardware-wise, a 1–2 vCPU / 2 GB RAM VPS is enough for a small group as long as inference runs elsewhere; if you also plan to host a local model on the same machine, plan for GPU memory and at least 16 GB of system RAM. Manual deployment without Docker is also documented but is more work: Node 18+, MongoDB 5+, a process manager, and a reverse proxy in front for TLS. Cost-wise, the categories that come up in practice are: free hosted instance plus Horde (zero), hosted instance plus your own OpenAI or Claude key (token-priced by usage), self-hosted on a small VPS plus Horde (the VPS cost only — typically a few US dollars a month), self-hosted on a VPS plus a paid API (the VPS plus the API), and self-hosted on owned hardware with a local backend (electricity and amortised hardware only). Because each backend has its own meter, the same Agnai instance can mix and match — Horde for casual chat, a paid API for important scenes, a local model when privacy is the priority.

e rather than in a browser, which suits anyone who wants total control over the front-end. Agnai gives up some of that depth in exchange for being browser-first, supporting multi-user rooms natively, and offering an official hosted instance for users who do not want to install anything. If you are alone, SillyTavern probably wins on extension breadth; if you intend to share a chat with other humans, Agnai is the path of least resistance. Backyard AI is the desktop counterpart with a polished marketplace and tight local-model integration; it does not target multi-user use, and its identity is a packaged desktop app rather than a self-hostable web service. Oobabooga's text-generation-webui is closer to a backend than a chat product — many Agnai users run Oobabooga as the inference layer rather than using its chat UI directly. RisuAI is the cross-platform lightweight cousin with strong Android support and a friendlier on-ramp; it does not match Agnai on multi-user features or the breadth of supported backends. Against the cloud apps in the C-cluster — Character.AI, Janitor, SpicyChat, Chub — Agnai gives up the polished consumer onboarding and the large catalogues of pre-built characters in exchange for control: you choose the model, you choose the policy, you choose where the data lives, and you can have multiple humans in one chat. The migration pattern is well known. Users who started on Character.AI and chafed at policy tightening, or Janitor users hitting BYO-API friction, often land in the open-source cluster; within that cluster Agnai is the option that keeps the browser convenience and adds collaboration. Agnai is the right choice for two audiences in particular. The first is anyone who wants to chat in a room with friends — a tabletop group, a writing duo, a Discord-style hangout — because no other open-source front-end ships that pattern natively. The second is anyone who is happy with a browser interface, expects to mix free Horde with their own paid API keys, and might later spin up a self-hosted instance for a private community. It is a weaker fit for users who want a packaged desktop app with no setup (Backyard fits better), for users chasing the longest tail of extensions (SillyTavern wins), or for users who prefer a curated mobile-first roleplay storefront with thousands of pre-built characters (a cloud C-cluster app fits better). For everyone else in between, Agnai's blend of open source, multi-backend, multi-user, browser-first, and self-hostable is hard to find anywhere else.