Put intelligence to work.

让智能真正工作。Execution you can trace, verify and take over.

Faven Studio 是一个可追溯、可校验的 AI 执行基座。从一个想法到发布上线 —— 分析、拆解、落地、交付编排在同一条链路上,全程自动跑得起来,也能在任何一环把决定权交回给人。 Faven Studio is a traceable, verifiable substrate for AI execution. From an idea to a shipped release — analysis, decomposition, implementation and delivery compose onto one chain that runs end to end unattended, and hands the decision back to a person at any link.

黑盒式 Agent 把执行压成一段文字流:做了什么只能事后读,想改只能重说一遍。Faven 把执行放在一个能留痕、能回放、能落闸的基座上 —— 对话只是它的一种形态。 A black-box agent flattens execution into a stream of text: you read what happened after the fact, and to change it you say it again. Faven puts execution on a substrate that records, replays and gates it — chat is only one of its surfaces.

看一次执行长什么样See what one execution looks like
需求与思路 The idea 一句话说清要做什么 State what to build 分析研判 Analysis 读代码 · 出方案 Read, then propose 运行中 · 轨迹可查 Running · traced 模块 Modules 按职责切分 Split by concern 文件 Files 落到具体路径 To real paths 人工确认 Sign-off 方案过目才动代码 Approve before code 编码与自测 Build & test 按拆解逐文件实现 File by file 构建发布 Ship it 打包、发版、拉起服务 Release and launch

← 左右滑动看全链← swipe to follow the chain

一条链One chain=思路 · 分析 · 拆解 · 确认 · 落地 · 发布idea · analysis · decomposition · sign-off · build · ship 每一格Each cell=一次可追溯的执行one traceable step 绿色Green=链路停下来等人the chain waits for a person 末端The end=真的拉起服务,不止于产出文件real software starts, not just files

差别The difference

问题不在对话,在没有底座 The problem is not chat. It is the missing substrate.

对话本身没错 —— Faven 自己也有对话形态。出问题的是执行之后什么都没留下:改不动、复现不了、审计无从查起。 Nothing is wrong with chat — Faven has a chat surface of its own. What fails is that nothing survives the run: you cannot amend it, reproduce it, or audit it.

黑盒式 AgentBlack-box agent
Faven Harness
形态Surface
只有一种 —— 一段对话,做什么都塞在里面One only — a conversation, with everything crammed inside
对话、可视链路、无人值守,按任务选;底座是同一个Chat, visual chain, unattended runs — chosen per task, on one substrate
过程留痕Trace
工具调用混在正文里,读完才知道做了什么Tool calls buried in prose; you only know afterwards
每一步的输入、输出、工具调用与耗时各自留痕Inputs, outputs, tool calls and timings recorded per step
介入方式Intervention
只能打断、重说、从头再来Interrupt, restate, start over
在任一步接管:改参数、改提示词、单步重跑,下游照旧Take over at any step: change the params or prompt, re-run that step alone
可复现Reproducibility
重新提问,两次结果没法逐项对比Ask again; two runs cannot be compared item by item
同一条链路重跑,输入与产物逐项对齐Re-run the same chain; inputs and artifacts line up
换模型 / 换工具Swapping parts
产品给什么用什么,换一家就得重来You use whatever the product ships; switching vendors means starting over
在基座上换零件,链路和留痕口径都不动Swap parts on the substrate; the chain and its records stay put

Faven Harness

一个基座,多种形态 One substrate, many surfaces

对话是一种形态,摊开的执行链路是另一种,定时无人值守是第三种。形态按任务挑,基座不换 —— 留痕、回放、回滚与权限的口径也就不会跟着换。 Chat is one surface. An execution chain laid out in full is another. A scheduled unattended run is a third. Pick the surface per task; the substrate stays, and so do the rules for tracing, replay, rollback and approval.

对话会话 Chat session 一句话交代目标, 过程照样逐条留痕 State the goal once; every step still recorded 可视链路 Visual chain 把执行摊开来编排, 任一步都能接管重跑 Lay execution out; take over at any step 无人值守 Unattended 定时点火,闸门与 判断接住每个分支 Fires on a schedule; gates catch each branch 开发推进 Dev planning 按模块先给方案, 确认之后才动代码 Proposes per module; touches code once agreed 事实问答 Grounded Q&A 断言必须带出处, 查不到就答没有 Claims carry sources; no source, no answer Faven Harness 基座 The substrate 上下文 Context 工具 Tools 控制 Control 可观察 Observability 五种形态,同一份执行记录 —— 留痕、回放、回滚与权限的口径完全一致。 Five surfaces, one execution record — identical rules for tracing, replay, rollback and approval.
对话会话Chat session一句话交代目标,过程照样逐条留痕State the goal once; every step still recorded
可视链路Visual chain把执行摊开来编排,任一步都能接管重跑Lay execution out; take over at any step
无人值守Unattended定时点火,闸门与判断接住每个分支Fires on a schedule; gates catch each branch
开发推进Dev planning按模块先给方案,确认之后才动代码Proposes per module; touches code once agreed
事实问答Grounded Q&A断言必须带出处,查不到就答没有Claims carry sources; no source, no answer

基座由四件事撑起来Four things hold the substrate up

上下文Context

项目即边界。执行记录、会话、线程与资产同在一个项目里,随项目整体同步到云端。The project is the boundary. Execution records, sessions, threads and assets live in one project and sync to the cloud with it.

工具Tools

能力可插拔。代码库读写、命令执行、联网搜索、子代理、技能与 MCP 服务器,按需接上。Capabilities plug in. Repository access, command execution, web search, sub-agents, skills and MCP servers attach as needed.

控制Control

人可落闸。权限审批、权限预设、检查点与整轮回滚,决定权始终在你这边。A person can shut the gate. Approvals, permission presets, checkpoints and whole-round rollback keep the decision with you.

可观察Observability

过程可见。思考、工具调用、耗时与状态逐条留痕,随时回放、随时核对。The process is visible. Reasoning, tool calls, timings and status are recorded item by item, ready to replay and check.

可追溯Traceability

每一个结论,都能指回它的来源 Every claim carries its receipt

交付物能不能用,取决于能不能查。Faven 把「查」做成默认行为,不是事后补的日志。 Whether the output is usable comes down to whether it can be checked. Faven makes checking the default, not a log bolted on afterwards.

事实纪律Grounded answers

接入数据库副本后,每条断言必须带上 [记录 id · 标题] 的出处;查不到就回答「数据库中没有该信息」,不用记忆补全,数字一律走计算工具。Once a database replica is attached, every claim must carry a [record id · title] source. If it is not there, the answer is "not in the database" — no filling in from memory, and every number goes through a calculator tool.

检查点与回滚Checkpoints

每轮会话开始前自动打检查点。这一轮改坏了,整轮改动一键退回,不必逐个文件找补。A checkpoint is taken before every round. If a round goes wrong, the whole round reverts in one step — no file-by-file salvage.

全程留痕Full trace

思考过程与工具调用轨迹逐条可查,运行指标随会话保存。静默超过 1 分钟直接在运行行标出「已 N 秒无事件」,不让任务默默挂死。Reasoning and tool-call traces are inspectable item by item, and run metrics are saved with the session. Go quiet for more than a minute and the run line says so — a task never dies in silence.

权限落闸Approvals

越界动作停下来等人点头。权限预设按场景切换,从「全程确认」到「放手跑」之间自己定刻度。Anything out of bounds stops for a nod. Permission presets switch by situation — you set the notch between "confirm everything" and "let it run".

自动化Autonomy

能整条链无人值守,也能把人放回关键那一环 Fully automated, or human-in-the-loop — same chain

自动化与人工不是两套系统。同一条链路上,哪一步交给机器、哪一步留给人,摆一个环节就切换。 Automation and human work are not two systems. On one chain, which step goes to the machine and which stays with a person is settled by placing one element.

无人值守Unattended

定时触发按一次 / 间隔 / Cron 点火,闸门、判断、序列器、互斥与计数把控制流接起来。批量按条目并行生产,上游未就绪时自动递归执行到可跑。Timers fire once, on an interval or on cron; gates, branches, sequencers, mutexes and counters wire the control flow together. Batches run per item in parallel, and an unready upstream is executed recursively until it can run.

定时触发Timers 闸门Gates 条件分支Branches 序列器Sequencers 互斥Mutex 计数Counters

人在环上Human in the loop

需求等待环节让链路停在需要判断的那一步;智能环节可以主动提问、请求授权。人处理完,链路从原处继续,跑过的一步都不重来。A wait element holds the chain at the step that needs judgement; an intelligent element can ask a question or request authorisation. Once the person is done, the chain resumes where it stopped — nothing already run is repeated.

需求等待Wait for input 提问确认Ask the user 权限审批Approvals 检查点回滚Rollback 终点判定Outcome gate 单步重跑Step re-run

一条链路里可以放什么What can go on a chain

代码库读写Repository access 模块架构拆解Architecture decomposition 子任务规划Sub-task planning 命令执行Command execution 智能任务Agent tasks 数据库检索Database lookup 资料与素材输入Source intake 文本生成Text generation 图像 / 视频 / 语音生成Image, video and speech 网络收发Network I/O 定时触发Scheduled triggers 条件分支Conditional branches 人工闸口Human gates 批量并行Parallel batches 子流程收纳Sub-chains 结果落盘Persisted output

多人项目Team projects

项目在云端,团队在同一张画布上 Projects live in the cloud — the team works on one canvas

建项目就像建一个代码仓库:项目存在云端,邀请就能一起做,版本管理在底层运转 —— 界面上不需要任何人懂 Git。 Creating a project is like creating a repository: it lives in the cloud, an invite is all it takes to work together, and version control runs underneath — nobody on the team needs to know Git.

换台电脑接着做Pick up anywhere

画布、会话与文件都在项目里,随项目存到云端。换台电脑登录就能接着做;每同步一次就是一个版本,改坏了随时回到之前。Canvases, sessions and files all belong to the project, and the project lives in the cloud. Sign in on another machine and carry on; every sync is a version you can go back to.

邀请与权限Invites and roles

所有者、可编辑、只读三种身份。成员一同步,新版本和成员变动经实时通道推到每个人,不用来回问「你那边最新了吗」。Owner, editor or viewer. When someone syncs, the new version and any membership change reach everyone over a live channel — no more "are you on the latest?"

合并不打架Merges that don't fight

按字段做三方合并,各改各的自动收拢;两个人改到同一处时,AI 给出合并建议,你确认了才落地。Field-level three-way merges fold separate edits together automatically. When two people touch the same spot, AI proposes a merge — nothing lands until you confirm it.

个人空间Personal space

当前文件、历史版本与社区作品各占多少,一目了然;空间不够按档位扩容。See exactly how much your current files, past versions and community works take up; add room by plan when you need it.

共创社区Community

每一个工作流,都可以成为别人的起点 Every workflow can be someone else's starting point

做好的东西不必锁在自己的项目里。发布出去,别人一键「做同款」就建进他自己的项目,改完再发,署名一路跟着走。 What you build doesn't have to stay locked in your project. Publish it, and anyone can remix it straight into their own project with one click — then publish again, with credit carried along.

就地发布Publish in place

在节点、文件或整个项目上右键就能发布,标题与封面替你预填。提交前自动检查密钥、本机路径、手机号这类不该公开的信息。Right-click a node, a file or a whole project to publish; the title and cover are filled in for you. Before submitting, keys, local paths, phone numbers and the like are checked automatically.

做同款与改编Remix

看中的作品一键「做同款」,直接建进自己的项目。改完再发布,会自动标明「改编自」并保留原作者署名。Remix any work straight into your own project. Publish your version and it is marked "remixed from", with the original author's credit kept.

新版本与审核Updates and review

原作发了新版本,用过它的人会收到提醒。上架前有审核、上架后有抽查,任何人都可以举报。When a work gets a new version, everyone who used it is told. Works are reviewed before listing and spot-checked after, and anyone can report one.

能发布什么What you can publish

工作流Workflows 图片Images 视频Video 音频Audio 文档Documents 网页 / 小游戏Web pages and mini-games 单个节点Single nodes 一组节点 / 超级节点Node groups and super nodes 项目模板Project templates

模型Models

模型是可替换的零件,不是绑死的地基 Models are swappable parts, not a foundation you are locked into

推理、代码、图像、视频、语音,在同一条链路里混着用。换模型不用重搭流程 —— 换的是那一步里的一个下拉框。 Reasoning, code, image, video and speech mix on one chain. Switching models does not mean rebuilding the flow — it is a dropdown inside one step.

推理 · 代码Reasoning & code
OpenAI Anthropic Gemini DeepSeek xAI Qwen
图像Image
GPT Image Gemini Image Midjourney Seedream Qwen Image FLUX Grok Imagine
视频Video
Seedance Kling Hailuo Sora Veo Vidu
音乐 · 语音Audio
Suno MiniMax OpenAI TTS

模型由官方统一维护与下发,客户端不再各自配服务商、也不再各自存密钥。换一个模型只是节点里的一个下拉框,链路和留痕口径都不动。 Models are maintained and delivered centrally — no per-client providers, no per-client keys. Switching one is a dropdown inside a step; the chain and its records stay put.

数据边界Data boundary

项目只给你邀请的人,调用走统一网关 Your projects reach only the people you invite. Model calls go through one gateway.

边界画得很清楚:项目只有你和你邀请的成员看得到,密钥不落在任何一台电脑上,消耗掉的东西统一入账。The line is drawn plainly: only you and the members you invite can see a project, keys never land on anyone's computer, and what you consume is metered in one place.

存在你的云端空间In your cloud space

项目存在你的云端空间里,只有你和你邀请的成员看得到,身份分所有者、可编辑、只读。本机留一份工作目录,点「同步」时才把改动传上去。A project lives in your cloud space, visible only to you and the members you invite — as owner, editor or viewer. A working copy stays on your machine; changes go up when you sync.

日志不带内容Logs without content

操作日志只记名字、状态和耗时,不记内容与密钥。远程排查要你主动开启、把接入码发给开发者才能连上,关掉即断。The activity log records names, states and timings — never content or keys. Remote diagnostics only connect once you switch them on and hand a developer the access code, and end when you switch them off.

一个网关,一份用量账One gateway, one ledger

官方模型统一经 Faven 网关调用:一个地址、一套凭据,密钥不落到任何一台客户端。哪一步用了哪个模型、花了多少,逐笔记在同一份账上。Official models are called through one Faven gateway: a single endpoint and one set of credentials, with keys never landing on any client. Which step used which model, and what it cost, lands line by line on the same ledger.

下载Download

装上就能跑 Install and run

桌面端原生应用Native desktop app · v1.0.0 · macOS 13 或更高macOS 13 or later · Windows 10 或更高Windows 10 or later