一桌人同场Everyone is present
每个参与角色都收到全部角色设定、个人设定和共享上下文，不是各自开小窗。别人刚说的话，下一位真的看得见。Every participant receives the full role list, the personal setting and the shared context. The next speaker genuinely sees what was just said.
真正的多角色群聊A real group chat, not a prompt box
让几个 AI，
一起把话聊完。Let several AIs
finish the thought together.
chataxi 是装进 Hermit 的 AI 对话应用。它不是“换一个模型接着问”，而是一桌人同时在场：每个角色有自己的模型、人设和声音，有主持人决定谁接话，语音在文字还没写完时就已经开口。chataxi is an AI chat happ for Hermit. It is not "swap the model and ask again". Several roles are in the room at once: each with its own model, persona and voice, a host that decides who speaks next, and speech that starts before the sentence is finished.
chataxi · 0.6.15 · 由 Hermit 安装chataxi · 0.6.15 · installed by Hermit
需要 Android 10 及以上手机，并先安装 Hermit。模型服务由你自己提供。Needs Android 10 or later plus Hermit. You bring your own model service.
一问一答，聊不出好结果One model, one voice, one opinion
为什么必须是群聊。Why it has to be a group.
同一个模型自己和自己讨论，只会不断同意你。真正有用的对话里有立场、有反驳、有分工——那需要一个房间里不止一个人。A single model talking to itself mostly agrees with you. Useful conversations contain positions, rebuttals and division of labour, which needs more than one participant in the room.
主持人负责调度A host routes the room
由你选定的主持人先按完整上下文输出该谁说话，再由那个角色正式回复。预选不写进消息历史，不会污染对话。Your chosen host reads the whole context, names the next speaker, and only then does that role reply. The routing step never enters the message history.
角色是独立个体Roles are individuals
一个角色可以绑定自己的模型、温度、推理强度、音色和发音参数。写代码的用强推理模型，陪聊的用低延迟无推理模型。A role can bind its own model, temperature, reasoning effort, voice and pronunciation settings. Use a heavy reasoning model for code and a low-latency one for company.
和同类工具比Compared with the usual options
更像 chatbox，还是更像 chatterUI？Closer to a chatbox, or closer to a chat UI?
两边各有一半：chatbox 这类工具接了更多模型，但会话形态通常还是“一个人回答一切”；chatterUI 这类应用界面更像聊天软件，却通常只服务一个模型。chataxi 要的是两者的交集——一个模型池，加上一屋子角色。Half of each. Tools like chatbox connect plenty of models but still let one voice answer everything. Apps like a chat UI feel like a messenger yet usually serve a single model. chataxi takes the intersection: a pool of models, and a room full of roles.
| 能力Capability | chataxi | chatbox 类工具chatbox-style tools | chatterUI 类应用chat-UI apps |
|---|---|---|---|
| 会话形态Conversation shape | 多角色群聊，一桌人同时在场Multi-role group chat | 通常是一个模型回答一切Usually one model answers | 以单模型连续对话为主Mostly one-to-one chats |
| 角色与音色绑定Per-role model and voice | 每个角色独立绑定模型、参数与音色Independent per role | 一般整会话共用一个模型Typically session-wide | 一般整会话共用一个模型Typically session-wide |
| 主持人调度Host routing | 有，且预选不写入历史Yes, kept out of history | 没有No | 没有No |
| 流式语音（边生成边播）Streaming speech | 模型流式 → 语音流式 → 分片播放Model → speech → playback, all streamed | 多以整段朗读为主Mostly read after finishing | 多以整段朗读为主Mostly read after finishing |
| 逐对话静音Per-conversation mute | 有：这一个对话单独静音Yes, per conversation | 通常只有全局开关Usually global only | 通常只有全局开关Usually global only |
| 语音输入Voice input | 系统识别实时部分结果 + 真实录音电平Live partials plus real input level | 多依赖系统输入法Depends on the keyboard | 多依赖系统输入法Depends on the keyboard |
| 场景开场白Scene opening | 有：五种模式自动生成主持人开场Yes, five generated modes | 没有No | 没有No |
| 上下文管理Context handling | N 滑杆 + 概要边界，保留最近的完整消息N slider with a summary boundary | 多为固定窗口或手动清理Often a fixed window | 多为固定窗口或手动清理Often a fixed window |
| 对话数据Conversation data | 只留在这台手机里Stays on this phone | 取决于部署与服务商Depends on deployment | 取决于部署与服务商Depends on deployment |
对比依据各自公开说明与常见用法整理，不同版本可能有差异，仅作参考。chataxi 侧的能力均可在应用内直接使用。Compiled from public documentation and common usage; versions differ, so treat it as a reference. Everything listed for chataxi is available in the app itself.
真正优秀的流式语音Speech that streams
文字还在写，
声音已经出来了。The voice starts
before the text ends.
“流式”在 chataxi 里不是把整段话快点读完。语言模型的原生流式输出会被切成段落、通过 WebSocket 或 SSE 直接送进语音合成，回来的音频分片立刻进入 Web Audio 队列——缓冲大约三秒就开始说话，后面的内容边说边到。Streaming here does not mean reading the finished paragraph faster. The model's native stream is segmented and pushed straight into speech synthesis over WebSocket or SSE; each audio fragment goes into the Web Audio queue immediately, so playback starts after roughly three seconds of buffer and the rest arrives while it talks.
- 语言模型按原生协议逐段输出，界面同步逐段显示。The model streams on its native protocol and the transcript fills in as it arrives.
- 段落进入 ElevenLabs 的 Text-to-Dialogue / TTS WebSocket，或 OpenAI、豆包的流式接口。Segments go into an ElevenLabs TTS or Text-to-Dialogue socket, or a streaming OpenAI or Doubao endpoint.
- 音频分片直接排队播放。某个流式通道在首个分片前不可用时，自动回退完整音频。Fragments queue straight into playback. If a streaming channel fails before its first fragment, it falls back to full audio.
应用退到后台立即暂停网页音频或停止系统朗读，回到前台自动继续；正在播放时手动切换静音不会打断这一句。Android 系统朗读作为默认兜底，不配置第三方服务也能用。Leaving the app pauses web audio or stops system speech, and returning resumes it; muting mid-sentence does not cut the sentence off. Android system speech is the default fallback, so it works without any third-party service.
对话与语音的建议组合A configuration worth starting from
角色模型Role model
Grok 无推理档（fast / non-reasoning）Grok non-reasoning tier
群聊拼的是“接话速度”。无推理模型首字延迟低、不把预算烧在思考上，角色扮演、闲聊、头脑风暴都更自然。A group chat is won on turn-taking. A non-reasoning tier answers sooner and spends nothing on hidden thinking, which suits role play, small talk and brainstorming.
语音Speech
ElevenLabs
Flash v2.5 延迟最低，适合实时接话；Multilingual v2 音质更稳；Eleven v3 适合对话式演绎。给每个角色配不同音色，一耳朵就能分清谁在说。Flash v2.5 is the lowest latency and best for live turn-taking; Multilingual v2 sounds steadier; Eleven v3 suits dialogue delivery. Give each role a different voice and you hear who is talking.
主持人Host
用一个更快的模型Use a faster model
主持人只负责决定谁说话，用最低延迟的模型即可，把贵模型留给真正要输出内容的角色。The host only chooses a speaker, so give it the fastest model and save the expensive one for roles that actually produce content.
模型与音色在 chataxi 里按“API Key → 获取目录 → 选择模型 → 连接测试 → 保存”接入，密钥只保存在这台手机上。费用与可用性以你选择的服务商为准。Models and voices are added in-app through key, catalog, selection, connection test and save, and keys stay on this phone. Cost and availability follow your own provider.
值得单独说的功能Features worth naming
细节多到你会开始较真。Enough detail to get picky about.
27 个角色模板27 role templates
按全部、男性、女性、其他分类，人物卡显示头像、职业、年龄和性别。选中只填入头像、名称和提示词，不会覆盖你已经调好的模型与声音。Filter by all, male, female or other; cards show portrait, occupation, age and gender. Choosing one fills in portrait, name and prompt only, never your tuned model or voice.
一张卡只代表一个模型One card, one model
复制模型卡就能在同一账号下快速加另一个模型，凭据通过本机引用复用，供应商目录单独缓存。不把多个模型塞进一张卡，也不猜能力。Duplicate a card to add another model on the same account; credentials are reused by local reference and provider catalogs are cached separately. No merged cards and no guessed capabilities.
场景设定与开场白Scene setting and opening
闲聊、思辨、学习、工作、倾诉五种模式各有独立策划提示，主持人生成约一百字的开场白。开场白落成一条可编辑的普通消息，和其他消息一样参与压缩。Five modes - small talk, debate, study, work, confiding - each with its own brief, and the host writes a short opening. The opening becomes an ordinary editable message that ages like any other.
看得懂的上下文压缩Context you can reason about
每个对话一根 5–50 的 N 滑杆，表示压缩后至少保留的最近完整消息数。触发压缩时只把最近 N 条之前的消息与上次概要合并，概要里记着精确覆盖的消息 ID 与累计数量。Each conversation has an N slider from 5 to 50: the number of recent full messages kept after compression. Only what precedes those N messages is merged into the summary, which records the exact message IDs it covers.
每个对话长成自己的样子Conversations get their own look
对话级的名称、介绍和头像分别保存，空字段才跟随“我的”。头像走 Hermit 受控的选取与裁切，拖动、双指缩放或加减按钮都行。Name, bio and portrait are stored per conversation and only fall back to your profile when empty. Portraits go through Hermit's controlled picker and cropper, with drag, pinch or buttons.
编辑与重新生成分开Editing and regeneration are separate
编辑只保存你改的文字，不会顺手生成或删除历史。重新生成独立执行，目标回复之后还有消息时必须先确认。Editing saves only the text you changed and never touches history. Regeneration is its own action and asks first when later messages exist.
图片、系统语音与第三方语音Images, system and third-party speech
对话里可以发图；系统语音识别与系统朗读开箱可用；也可以接第三方语音输入，用真实麦克风波形观察录音状态。Send images into the conversation; system speech recognition and system read-aloud work out of the box; third-party voice input shows the real microphone waveform while recording.
数据留在手机上Data stays on the phone
对话、角色、模型配置都保存在 chataxi 自己的数据区，完整密钥不写进页面 HTML 或日志，密钥编辑器只从系统剪贴板主动粘贴。Conversations, roles and model configs live in the happ's own data partition. Full keys never reach page HTML or logs, and the key editor only pastes from the system clipboard on request.
能接什么What connects
模型和声音，都自己选。Your models, your voices.
chataxi 不自带额度，也不捆绑服务商。它把不同协议统一成一套内部合同，再编译成各家的真实请求。chataxi ships no quota and bundles no provider. It normalises protocols into one internal contract and compiles real provider requests from it.
对话模型Chat models
OpenAI Chat / Responses、Anthropic Messages、Gemini Interactions / GenerateContent、Ollama Chat，以及兼容 OpenAI 协议的各类服务与聚合平台。OpenAI Chat and Responses, Anthropic Messages, Gemini Interactions and GenerateContent, Ollama Chat, plus OpenAI-compatible services and aggregators.
语音合成Speech synthesis
ElevenLabs（Multilingual v2 / Flash v2.5 / Eleven v3，支持流式与对话式 WebSocket）、OpenAI Speech、豆包 V3 SSE（含声音复刻资源）、Android 系统朗读。ElevenLabs (Multilingual v2, Flash v2.5, Eleven v3 over streaming and dialogue sockets), OpenAI Speech, Doubao V3 SSE including voice cloning, and Android system speech.
局域网与本地LAN and local
可信局域网里的 HTTP 服务可以直接接入，自建网关或本机运行时都能用，不必先上公网。HTTP services on a trusted LAN connect directly, so a self-hosted gateway or a local runtime works without exposing anything publicly.
不受支持就不显示Nothing fake is shown
温度、最大输出、top_p、top_k、推理强度、图片生成、音色与发音参数都按当前具体模型的已知能力出现；不确定的模型归入“未知能力”，不凭名字开启高级参数。Temperature, max output, top_p, top_k, reasoning effort, image generation, voices and pronunciation appear only when the specific model is known to support them. Unknown models stay unknown rather than being guessed from their name.
把 chataxi 装进手机Install chataxi
先用 Hermit 扫码，确认后就能打开。Scan with Hermit, confirm, and open it.
- 打开 Hermit，点“扫码添加”。Open Hermit and choose Scan to add.
- 扫描这个二维码，确认安装。Scan this code and confirm the install.
- 点开 chataxi，按连接引导添加模型服务。Open chataxi and add your model service following the in-app guide.
chataxi 用于连接你选择的 AI 服务，本身不附送模型额度。密钥只填在 chataxi 的服务配置里。chataxi connects the AI services you choose and includes no quota. Keys belong in its own service settings.
同一位作者的下一个作品The next one from the same workshop
让它一边聊，一边把画面画出来。Let it draw while it talks.
VibeDraw 是同一套 Hermit 里运行的实时绘图应用：你在画布上随手涂几笔，成图层跟着变成一张完整的图。和 chataxi 一起装，讨论和视觉可以同步推进。VibeDraw is the real-time drawing happ for the same Hermit: scribble on the canvas and the result layer turns it into a finished picture. Install both and the discussion can move with the visuals.
装好了却不能聊天？Installed but it will not talk?
chataxi 只负责连接，不提供模型额度。打开应用后按引导添加服务，填入服务商的 API Key 并选择模型，再创建角色开始聊天。网络与费用以你选择的服务为准。chataxi only connects; it provides no quota. Add a service in the app, paste your provider key, pick a model, then create roles and start chatting. Network and cost follow that provider.
和 Hermit 是什么关系？How does Hermit fit in?
chataxi 是一个 happ，需要装在 Hermit 里运行。Hermit 负责安装、数据、文件、录音、权限与系统语音，chataxi 负责对话本身。更多问答在 Hermit 的常见问题里。chataxi is a happ and runs inside Hermit. Hermit owns installation, data, files, recording, permissions and system speech; chataxi owns the conversation. More answers live in the Hermit FAQ.
想改它的样子？Want to change how it looks?
在 Hermit 的开发页开启开发模式，把地址和密码交给你信任的智能体，就可以让它给 chataxi 换配色、调布局、加功能。看看怎么连接。Turn on development mode in Hermit, hand the address and password to an agent you trust, and it can restyle chataxi, adjust the layout or add features. See how to connect.
一个人问，一屋子答。One question, a whole room answering.
装好 Hermit，扫码把 chataxi 加进来。Install Hermit, scan, and add chataxi.