Private beta · An agent-native canvas私有内测 · 智能体原生画布
Muse is an AI agent that plans, reads your PDFs, and draws on a shared canvas — skeleton first, details grafting in live. Edit any node by hand; on your next message Muse sees exactly what you changed, and builds on it. Muse 是一个会自主规划的 AI 智能体：读你的 PDF，在一张你们共用的画布上落笔——骨架先行，细节实时长出。任何节点你都可以上手改；下一条消息，Muse 看得见你具体改了哪，顺着你的思路继续画。
A real agent, drawing in the open真·智能体，当着你的面画
Muse decides when to search your documents, when to draw, and when to edit — each step streams in as a live tool card. The map lands skeleton-first, so you watch it grow instead of staring at a spinner. Muse 自己决定什么时候检索文档、什么时候画图、什么时候改图——每一步都以实时工具卡片流进来。出图骨架先行，你看着它长出来，而不是盯着转圈。
Reads your goal, decides which tools the job needs, and says so out loud.读懂你的目标，决定这活儿需要哪些工具，并把打算说出来。
Chunk → embed → in-memory retrieval, across all attached docs at once. ~100 lines, no LangChain.切块 → 向量化 → 内存检索，一次跨所有附加文档。约 100 行实现，不依赖 LangChain。
Root and main branches land immediately; detail grafts in as it streams.根和主枝立即出现，细节随流式陆续长上去。
“Rename the pricing branch” patches exact nodes by id. Your manual tweaks survive.「把定价分支改名」按节点 id 精准打补丁，你手动改过的地方不被冲掉。
…and you hold the pen: auto, confirm-risky, or review-everything — rejected ideas are never suggested twice. 而且笔在你手里：全自动、危险先问、逐条审——被你拒掉的提议，它不会再提第二次。
A workbench, not a chat log是工作台，不是聊天记录
Edit nodes by hand; on your next message Muse sees exactly what you changed and builds on it, like a collaborator.手动改节点，下条消息 Muse 能看到你具体动了哪，像协作者一样顺势接着做。
Muse asks one sharp question before drawing — the structure comes out of your head, not the model's.Muse 先问你一个尖锐的问题再落笔——结构从你脑子里长出来，而不是模型的。
Tab child, Enter sibling, F2 rename, drag to re-parent; a ⌘K palette over your maps and docs.Tab 子节点、Enter 兄弟、F2 改名、拖拽改父；⌘K 面板模糊搜索导图和文档。
Fold branches away or hoist one to read alone; a gallery keeps every map, auto-titled.分支可折叠、可单独拎出来读；画廊存下每张图，标题自动精炼。
In: Markdown, OPML. Out: PNG, SVG, Markdown, OPML, Mermaid — or a read-only share link.进：Markdown、OPML。出：PNG、SVG、Markdown、OPML、Mermaid——或一条只读分享链接。
A warm paper canvas, hand-drawn nodes and edges, handwriting labels — a map that feels sketched, not generated.暖纸画布、手绘风节点连线、手写体标签——像亲手勾出来的图，而不是机器吐出来的。
The honest comparison摆在明面上比
"AI mind map" usually means: type a prompt, get a static tree, regenerate and lose your work. Inkling is built around one idea — the canvas is alive, and you and the agent draw on it together.「AI 思维导图」通常意味着：敲一句 prompt，得到一棵静态树，一重新生成就把你的心血冲掉。Inkling 只围绕一件事构建——画布是活的，你和智能体在同一张画布上一起画。
| Inkling | Typical AI mind-map tools常见 AI 思维导图工具 | |
|---|---|---|
| How the map is born导图怎么诞生 | Drawn live, skeleton first实时画出，骨架先行 | A spinner, then a wall of nodes转圈，然后糊你一脸节点 |
| How edits work改图方式 | Patches exact nodes by id按节点 id 精准补丁 | Regenerates the whole map整图重新生成 |
| Your hand edits你的手动改动 | Seen, kept, and built upon被看见、被保留、被接着发挥 | Overwritten on the next run下一次生成就被覆盖 |
| The agent's work智能体的工作过程 | Every tool call visible每步工具调用可见 | A black box黑盒 |
| Model choice模型选择 | OpenAI · Claude · DeepSeek · Kimi · any compatible endpointOpenAI · Claude · DeepSeek · Kimi · 任意兼容端点 | Locked to their pick锁死它们选的 |
| Deployment部署形态 | Private, in your environment私有部署，跑在你的环境 | Their multi-tenant cloud它们的多租户云 |
Private beta私有内测
We're onboarding early users in small batches — leave your email and we'll write to you once, when it's your turn. Tell us nothing else; we count interest, not people.我们正在小批量邀请早期用户——留下邮箱，轮到你时我们写一封邮件给你。不用填别的：我们统计的是意向，不是个人信息。
FAQ
Those tools render a tree from a prompt — one shot, take it or leave it. Inkling's canvas is a shared workspace: the agent draws in visible steps, edits exact nodes instead of regenerating, and reads your hand edits back into the conversation. It's closer to pairing with a collaborator than prompting a generator.那些工具是「一句 prompt 渲染一棵树」，一次成型、爱要不要。Inkling 的画布是一个共同工作区：智能体一步步当面画、按节点精准改而不是重新生成，你的手动改动还会被读回对话里。它更像和一位协作者结对，而不是向生成器要图。
Always. Export PNG, SVG, Markdown, OPML, or Mermaid; import Markdown and OPML back in; or share any map as a read-only link.随时可以。导出 PNG、SVG、Markdown、OPML、Mermaid；Markdown 和 OPML 也能导回来；任何图都能生成只读分享链接。
OpenAI, Claude, DeepSeek, Kimi out of the box, plus any OpenAI-compatible gateway — including a fully local endpoint, so documents and inference both stay on your network.开箱支持 OpenAI、Claude、DeepSeek、Kimi，以及任意 OpenAI 兼容网关——包括纯本地端点，文档和推理都不出你的内网。
Yes — Inkling deploys privately, and works with any OpenAI-compatible endpoint including fully local ones. The model needs function-calling support.能——Inkling 支持私有部署，任意 OpenAI 兼容端点都能接，包括纯本地端点。模型需支持 function calling。
Inkling is in private beta. Join the waitlist and we'll email you once, when your invite is ready — or try the live demo right now.Inkling 目前私有内测中。加入等候名单，邀请就绪时我们发一封邮件给你——也可以现在就试试在线 demo。