A self-hostable AI agent that turns your documents into editable mind maps — your files and API keys never leave your machine.
English · 简体中文
Chat with Hermas, an AI agent that plans, reads your PDFs, and draws editable mind maps live — every tool call visible as it works.
Most AI mind-map tools are cloud SaaS — your documents get uploaded to their servers and your model key lives in their backend. MindGenius runs entirely on infrastructure you control: one Docker image, bring-your-own-key, nothing stored server-side. Built for privacy-conscious knowledge workers, self-hosters, and teams with sensitive documents (legal, research, internal specs) that can't be pasted into a third-party cloud.
🔗 Try the live demo → · bring your own API key (stays in your browser). First load may take ~30s to wake. · ⭐ Star it if it's useful — it helps others find it.
| MindGenius | Typical cloud AI mind-map tools | |
|---|---|---|
| Runs on your own machine | ✅ one Docker image | ❌ their cloud only |
| Your documents stay local | ✅ never leave your box | ❌ uploaded to their servers |
| Model API key | ✅ in your browser only | ❌ held in their backend |
| Bring your own model | ✅ OpenAI · Claude · DeepSeek · Kimi · any compatible endpoint | ❌ locked to their pick |
| Real multi-step agent | ✅ visible tool calls (plan → retrieve → draw → edit) | ❌ one-shot prompt → static tree |
| Edits the existing map | ✅ surgical, by node id | ❌ regenerates the whole thing |
| Source-available & hackable | ✅ BSL 1.1, RAG in ~100 lines | ❌ closed SaaS |
pnpm install
pnpm setup # pick a provider, paste a key → writes apps/server/.env
pnpm dev # web on :5173, api on :3000No key at setup? Skip it and paste one later in the app's Settings (top-right) — it stays in your browser only.
- 🔒 Self-hosted & private — ships as one Docker image you run anywhere; your PDFs are chunked/embedded in memory and your key lives in the browser. Nothing is written to the server's disk, no telemetry on your content.
- 🤖 Real agent, not a prompt wrapper — multi-step tool loop (Vercel AI SDK v5) that decides when to search docs, generate the map, and expand branches, each shown as a live tool card.
- ✏️ Surgical edits — "rename the pricing branch" patches exact nodes by id, so your manual tweaks survive (no full re-render).
- 🤝 Two-way canvas — edit nodes by hand (rename, add, drag, delete) and Hermas sees exactly what you changed on your next message, then builds on it like a collaborator — not just the latest snapshot.
- 📄 Reads your PDFs — attach one or more PDFs (📎): chunk → embed → in-memory retrieval (~100 lines, no LangChain).
rag_querysearches across all attached docs at once. - ⌨️ Keyboard-first canvas —
Tabchild ·Entersibling ·F2rename ·Deldelete ·⌘Zundo · drag a node onto another to re-parent it. - ⌘ Command palette (⌘K) — Cursor-style overlay to fuzzy-search your saved maps + uploaded docs;
↵opens a map or attaches a doc. - 🔁 Import / export — import Markdown / OPML, export PNG · SVG · Markdown · OPML · Mermaid.
- 🔑 Bring your own key — OpenAI / Claude / DeepSeek / Kimi, or any OpenAI-compatible endpoint. Keys never touch the server's disk.
- 🎨 One focused workbench — chat + canvas + doc drawer; persists across refresh; bilingual (EN / 中文); dark, dependency-free UI.
pnpm setup writes apps/server/.env for you. To edit by hand, copy apps/server/.env.example → apps/server/.env.
| Provider | Embeddings (RAG) |
|---|---|
| OpenAI · Claude | ✅ full RAG |
| DeepSeek · Kimi | EMBEDDING_API_KEY to any OpenAI-compatible endpoint to keep it on |
The whole app ships as one image (server serves the web bundle + API):
docker run -p 3000:3000 -e OPENAI_API_KEY=sk-... ghcr.io/xianjianlf2/mindgeniusai
# open http://localhost:3000Or build locally: cp apps/server/.env.example apps/server/.env && docker compose up --build.
Omit the key to let each visitor bring their own from Settings. To run a public demo, set a server key and DEMO_DAILY_LIMIT=5 — visitors get a few free runs per day (per IP) before being nudged to add their own key; BYO-key users are never limited.
sdk: docker / app_port: 3000; set OPENAI_API_KEY as a secret or leave it empty.
| Command | What it does |
|---|---|
pnpm setup |
first-run config wizard |
pnpm dev |
web + server in parallel |
pnpm build |
typecheck & build all packages |
pnpm test |
run vitest workspace |
pnpm lint |
eslint |
apps/web/ React 18 + Vite + Zustand + AntV X6 (custom design system)
apps/server/ Hono + Vercel AI SDK v5 (Hermas agent, RAG, SSE streaming)
packages/shared/ SSE / agent-event protocol shared by both ends
- Agent loop —
streamText+stopWhen: stepCountIs(8), four zod-typed tools (mindmap_generate/node_expand/rag_query/mindmap_edit). - Live edits —
mindmap_editreturnsadd/update/removeops applied to the tree by node id. - Two-way sync — your manual canvas ops share that same op model; they're buffered (coalesced) and sent as
recentEditsso the agent knows what you changed since its last turn, not just the current outline. - RAG — pdf-parse → overlap chunking →
embedMany→ cosine retrieval, in-process. - Multi-provider — resolved per request from request headers; keys never stored server-side.
Full design: docs/REFACTOR_PLAN.md.
Cookieless, count-only (page views + an agent_run event) via Umami / GoatCounter / Cloudflare. Never sends chat, PDFs, or keys. See apps/web/.env.example.
Business Source License 1.1. Free for personal, academic, research, and non-profit use. Commercial use requires a license — contact mark-xian@foxmail.com. Each version converts to Apache-2.0 four years after release (Change Date: 2030-06-29).