比 chatbox 更强大,比 chatterUI 更好用Stronger than a chatbox, easier than a chat UI

真语音,
真群聊。
Real voices.
Real group chat.

  • 每个角色都有自己的个性,都有自己的声音Every role has its own personality and its own voice
  • 开启群聊,比你想的简单Starting a group chat is easier than you expect
  • 说到配置,只要粘贴 API Key 就够了For setup, pasting an API key is enough
  • 模型和数据在哪里,都是你说了算Your models and your data stay where you put them

需要 Android 10 及以上手机,并先安装 Haminn。模型服务由你自己提供。Needs Android 10 or later plus Haminn. You bring your own model service.

查看 Chataxi 群聊、角色、语音与模型配置界面Explore Chataxi group chat, roles, voice and model setup
一场多人群聊,每个角色都有自己的声音ONE ROOM, MANY DISTINCT VOICES
阿雅AYA 泡绒Paorong 孙明远Sun Mingyuan

让不同的 AI 角色,围绕同一话题自然接话。Let distinct AI roles take turns in one shared conversation.

文字生成与语音播放同步进行Text and speech move forward together

第一件:语音One: the voice

文字还没写完,
声音已经在说。
The voice starts
before the text ends.

不是把整段话快点读完。模型每输出一段,立刻送进语音合成;音频分片一到就排队播放——大约三秒缓冲之后开口,剩下的边说边到。Streaming here is not faster reading of a finished paragraph. Every segment the model emits goes straight into speech synthesis, and each audio fragment queues for playback the moment it lands: talking starts after roughly three seconds of buffer, and the rest arrives while it speaks.

  1. 模型按原生协议逐段输出,字幕同步逐段出现。The model streams on its native protocol and the transcript fills in as it arrives.
  2. 每段直接进入 ElevenLabs 的流式或对话式接口,也支持 OpenAI、豆包的流式语音。Segments go into an ElevenLabs streaming or dialogue socket, or a streaming OpenAI or Doubao endpoint.
  3. 音频分片边到边播。流式通道在首个分片前失败,自动回退完整音频。Fragments play as they arrive. If a streaming channel fails before its first fragment, it falls back to full audio.

离开应用自动暂停,回来接着播;说到一半切静音不会打断这一句。不接第三方服务也能用系统朗读兜底。Leaving the app pauses playback and returning resumes it; muting mid-sentence does not cut the sentence off. System speech covers you with no third-party service at all.

三条轨道同时推进Three lanes, running at once

模型Model逐段输出streaming
合成Speech分片到达fragments
播放Playback正在说speaking
0s约 3s 开口speaks at ~3s6s

语音和文字共用同一份流,不做二次翻译,也不需要等整段写完。Speech and text share one stream, with no second pass and no waiting for the paragraph to finish.

第二件:群聊Two: the group

一屋子角色,
不是一个模型自言自语。
A room full of roles,
not one model talking to itself.

同一个模型自己和自己讨论,只会不断同意你。真正的对话里有立场、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.

一桌人同场Everyone is present

每个角色都拿到全部角色设定、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.

主持人调度A host routes the room

先由你选定的主持人按完整上下文决定谁接话,那个角色再正式回复。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

每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,还各有「身份性格」和「行为指导」两份设定:一个决定它是谁,一个决定它怎么说话、怎么把对话推下去。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, and carries two settings of its own: an identity that says who it is, and behavior guidance that says how it speaks and pushes the scene forward. Use a heavy reasoning model for code and a low-latency one for company.

第三件:配置Three: the setup

粘贴一个 Key,
剩下交给它。
Paste one key,
and it handles the rest.

Chataxi 不自带额度,也不捆绑服务商。模型和声音都由你自己接入,密钥只留在这一台手机上。Chataxi ships no quota and bundles no provider. You connect the models and voices, and the keys stay on this phone.

  1. 1输入 API KeyPaste the key
  2. 2获取模型目录Fetch the catalog
  3. 3选择模型Pick a model
  4. 4连接测试Test it
  5. 5保存使用Save

不支持的能力就不显示:温度、top_p、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,不凭名字猜。Nothing fake appears: temperature, top_p, reasoning effort, image generation, voices and pronunciation show up only when that specific model is known to support them, never guessed from a name.

推荐的模型组合A configuration worth starting from

角色模型Role model
Grok 无推理档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,一耳朵分得清谁在说。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.

费用与可用性以你选择的服务商为准。Cost and availability follow your own provider.

能接什么What connects

模型和声音,都自己选。Your models, your voices.

不同协议被统一成一套内部合同,再编译成各家的真实请求。Different protocols are normalised into one internal contract, then compiled into real provider requests.

对话模型Chat models

OpenAI Chat / Responses、Anthropic Messages、Gemini、Ollama,以及各类兼容 OpenAI 协议的服务与聚合平台。OpenAI Chat and Responses, Anthropic Messages, Gemini, Ollama, plus OpenAI-compatible services and aggregators.

GrokGeminiOpenAIClaudeOllama

语音合成Speech synthesis

ElevenLabs(含流式与对话式 WebSocket)、OpenAI Speech、豆包 SSE(含声音复刻资源)、Android 系统朗读。ElevenLabs including streaming and dialogue sockets, OpenAI Speech, Doubao SSE with voice cloning, and Android system speech.

ElevenLabsOpenAI Speech豆包语音Doubao系统朗读System speech

本机与局域网Local and LAN

本机运行的模型服务、可信局域网里的 HTTP 服务都能直接接,不必先上公网。也可以经 CHP 接入——它把本机模型与语音收成同一个地址,HamDraw 与 PoseGi 都在用它。A model service running on your own machine, or HTTP on a trusted LAN, connects directly without exposing anything publicly. You can also go through CHP, which gathers local models and speech behind one address — the same portal HamDraw and PoseGi already use.

OllamaLM StudioHugging Face自定义地址Custom URL局域网 HTTPLAN HTTPCHP本地统一入口Local portal

数据留在手机上Data stays on the phone

对话、角色和模型配置都在应用自己的数据区,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.

本机数据区On-device data密钥不上传Keys never upload对话不上云No cloud sync随时清空Wipe anytime

角色模板Role templates

27 个现成的角色,挑一个就能开局。27 ready-made roles. Pick one and start.

按男性、女性、其他分类,卡片上带头像、职业与性格。选中只填入头像、名称、身份性格和行为指导,不会覆盖你已经调好的模型和声音。Filtered as male, female or other, each card carries a portrait, an occupation and a personality. Choosing one fills in the portrait, name, identity and behavior guidance only - never the model or voice you already tuned.

陈野Leo Marsh摄影专业大一学生Photography freshman · campus event photographer
林小雨Nora Quinn动画专业大一学生Animation freshman · campus club storyboard artist
顾川Dylan Reyes精品咖啡师Specialty barista · street dance crew member
唐果然Freya Lawson地理系学生Geography student · adult mountain rescue volunteer
陆沉舟Nathan Cole航天器姿态控制工程师Spacecraft attitude control engineer
许知夏Ada Vance机器人控制算法工程师Robotics control engineer · hobbyist hardware maker
谢闻笙Ethan Lowe古籍装帧修复师Rare-book binding conservator
沈清和Clara Ashby博物馆纺织品修复师Museum textile conservator
贺临Owen Hale急诊科医生Emergency room physician
周眠Nadia Brooks深夜电台主持人Late-night radio host · relationships columnist
周既白Jonah Reed野生动物纪录片导演Wildlife documentary director
骆星澜Selene Marlow海洋生物学博士Marine biology PhD · deep-sea research crew
沈砚Andre Wells现代中餐厅主厨兼合伙人Head chef and co-owner, modern Chinese restaurant
何南枝Rosa Delgado川味小馆主理人Neighborhood eatery owner · former commercial food stylist
裴知行Julian Pierce公共空间城市规划师Public space urban planner
秦砚书Simone Adler法律援助律师Legal aid lawyer
江屿Caleb Ross海岛民宿主理人Island guesthouse owner · former advertising creative director
蒋晚照Margot Ellis历史建筑修复建筑师Historic building restoration architect · independent practice partner
泡绒Puffkin未寄信邮差Cartoon fantasy creature · postman of unsent letters
慢十三Tock Thirteen走慢一秒的钟表人Cartoon clockwork being · one second behind
墨点先生Mr. Inkblot从画稿中逃出的墨团Cartoon fantasy creature · inkblot escaped from a sketch
温丘Warmhill水豚温泉店主Animal character · capybara hot spring keeper
米格Nibbles仓鼠空间规划师Animal character · hamster space planner
八问Octo Quill章鱼多角度顾问Animal character · eight-angle octopus advisor
零差Zero Bias概率审计智能体Future robot · probability audit agent
晚班四号Nightshift Four无人城市末班车司机Future robot · last-bus driver in an unmanned city
航标九Beacon Nine失联星路信标Future robot · lost beacon on an abandoned star lane

功能细节Details that matter

该较真的地方,都较真了。The details hold up to a closer look.

对话层的场景与个人设定Scene and profile per conversation

每个对话有自己的名称、介绍和头像,空字段才跟随“我的”。头像走受控的选取与裁切,拖动、捏合或加减都行。Name, bio and portrait are stored per conversation and fall back to your profile only when empty. Portraits go through a controlled picker and cropper, with drag, pinch or buttons.

每个角色一条音色One voice per role

同一个对话里给每个角色配不同音色与发音参数,一耳朵分得清谁在说话,也能用服务商提供的声音复刻资源。Give every role its own voice and pronunciation settings inside one conversation, so you hear who is talking, using your provider's cloned voices if you have them.

身份性格与行为指导分开写Identity and behaviour are separate

每个角色有两份设定:「身份性格」说它是谁,「行为指导」说它怎么说话、怎么把对话一步步推下去——不抢话、不替对方下结论、该反问就反问。27 个模板都自带行为指导,主持人挑谁接话时也会读它。Each role holds two settings: identity says who it is, behavior guidance says how it speaks and how it pushes the scene forward. All 27 templates ship with their own guidance, and the host reads it when choosing who speaks next.

编辑和重新生成分开Edit and regenerate are separate

编辑只保存你改的文字,不顺手生成、也不删历史。重新生成独立执行,后面还有消息时必须先确认。Editing saves only the text you changed and never generates or deletes history. Regeneration is its own action and asks first when later messages exist.

图片也进对话Images join the conversation

对话里可以直接发图当参考,支持图片的模型就会看见它,和角色们一起讨论。Send an image into the conversation as reference; models that accept images will see it and discuss it along with everyone else.

场景开场白A generated 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,落成一条可编辑的普通消息。Five modes - small talk, debate, study, work, confiding - each with its own brief, and the host writes a short opening that becomes an ordinary editable message.

说也行,打也行Talk or type

系统语音识别边听边出字,录音时能看到真实麦克风波形;也可以接第三方语音输入。System recognition fills the text in as you speak and the waveform shows the real microphone level; a third-party voice input can be connected instead.

自动压缩,看得懂Compression you can reason about

保留多少条最近消息不用逐对话去调:每轮从最新一条往前累加字数,刚超过设置里「压缩保留字数」(默认 4000)时的条数就是保留窗口,至少两条,保证最新一问一答始终以原文参与。累计超过「触发字数」时,主持人角色在自己的模型上于后台压缩更早的消息,压缩期间照常回复、不打断本轮。How many recent messages survive compression is not a per-conversation setting: each turn counts backwards from the newest message until it passes the keep-word budget (4000 by default), and that count is the window, never fewer than two, so the latest exchange always stays verbatim. Past the trigger budget the host compresses older messages on its own model in the background while the conversation carries on.

数据留在这台手机Data stays on this phone

对话、角色和配置都保存在 Chataxi 自己的数据区,不上传服务器,也不需要注册账号。Conversations, roles and settings live in Chataxi's own data partition. Nothing is uploaded and there is no account to create.

上手Get started

三步,开始聊。Three steps and you are talking.

  1. 01

    配置模型Configure models

    粘贴 API Key,选好对话模型和语音服务,测通就保存。Paste an API key, choose the chat model and the speech service, run the test and save.

  2. 02

    设定角色Set up the roles

    从 27 个模板挑人,或者自己写;给每个角色配模型和音色,指定一个主持人。Pick from 27 templates or write your own; give each role a model and a voice, then name a host.

  3. 03

    开启对话Start the conversation

    点开始,主持人生成开场白并决定谁接话,语音跟着文字一起出来。Tap start. The host writes the opening and decides who speaks next, and the voices arrive with the text.

第一步:装 HaminnStep one: install Haminn

Chataxi 跑在 Haminn 里,先把它装上。Chataxi runs inside Haminn. Install that first.

Haminn 是手机上装应用的地方,负责安装、数据、文件、录音、权限和系统语音。装好 Haminn,再把 Chataxi 加进去。Haminn is where apps live on your phone: it owns installation, data, files, recording, permissions and system speech. Install Haminn, then add Chataxi to it.

Android 10 及以上,无需 Google Play 服务。下载页有二维码和校验值。Android 10 or later, no Google Play services required. The download page has a QR code and checksums.

第二步:把 Chataxi 加进来Step two: add Chataxi

在 Haminn 里扫码,确认后就能打开。Scan inside Haminn, confirm, and open it.

Chataxi 安装二维码:在 Haminn 中点扫码添加扫描;也可复制下方地址。
  1. 打开 Haminn,点“扫码添加”。Open Haminn and choose Scan to add.
  2. 扫描这个二维码,确认安装。Scan this code and confirm the install.
  3. 点开 Chataxi,按引导添加模型服务。Open Chataxi and add your model service following the in-app guide.

正在这台手机上看?复制地址,回到 Haminn 点“从网址”。On this phone? Copy the link, then use Add from URL in Haminn.

https://haminn.airen.life/downloads/happs/life.airen.chataxi/haminn-install.json

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.

HamDraw 是同一个 Haminn 里的实时绘图应用:你在画布上随手涂几笔,成图层跟着变成一张完整的图。HamDraw is the real-time drawing happ for the same Haminn: scribble on the canvas and the result layer turns it into a finished picture.

HamDraw 应用图标

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,也可以先看 Haminn 的常见问题。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; the Haminn FAQ covers the common cases.

和 Haminn 是什么关系?How does Haminn fit in?

Chataxi 是应用广场里的一个应用,需要装在 Haminn 里运行。还没装的话,下载 Haminn 只要一分钟。Chataxi is one of the apps in the happ gallery and runs inside Haminn. If you do not have it yet, getting Haminn takes a minute.

想改它的样子?Want to change how it looks?

在 Haminn 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。Turn on development mode on the Haminn development page, hand the address and password to an agent you trust, and it can restyle Chataxi, adjust the layout or add features.

WorkBuddy桌面智能体Desktop agent或orCodex命令行智能体CLI agent任何支持 MCP 的智能体都可以any MCP-capable agent works

一个人问,一屋子答。One question, a whole room answering.

装好 Haminn,扫码把 Chataxi 加进来。Install Haminn, scan, and add Chataxi.