MindGenius AI

Talk to Hermas — an autonomous agent that plans, retrieves your documents, and draws editable mind maps for you, with every tool call visible in real time.

Why it's different

🤖 A real agent, not a prompt wrapper. Most "AI mind map" tools do a single prompt → markdown → render. Hermas runs a multi-step tool-calling loop (Vercel AI SDK v5): it decides on its own to search your uploaded PDF (rag_query), generate the map structure (mindmap_generate), and expand branches (node_expand) — chained automatically, with each step shown live as an expandable tool card (input/output included).

📄 It reads your documents while drawing. Attach a PDF from the composer (📎): upload → chunk → embed → in-memory vector index, all in ~100 lines without LangChain. Hermas cites retrieved passages and incorporates them into the map.

🔑 Bring your own key, keep your privacy. OpenAI / Anthropic (Claude) / DeepSeek, switchable at runtime. Keys live only in your browser's localStorage and are sent per-request to your own backend — never stored server-side. Custom gateway (base URL) supported.

🔁 Rebuilt without breaking anything. The stack was migrated from Vue/Koa/LangChain to React/Hono/AI SDK incrementally: the legacy SSE envelope {status, data} and every old endpoint still work — agent events are layered inside the existing protocol (packages/shared), not bolted on beside it.

🎨 A focused single workbench. Chat panel + editable X6 canvas + document drawer in one screen. Add/rename/delete nodes, AI-brainstorm any branch, undo/redo, export PNG. Dark, restrained design system with zero UI-framework dependency (custom tokens & components — no antd).

Architecture

├── apps/
│   ├── web/        # React 18 + Vite + Zustand + AntV X6 (custom design system)
│   └── 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), tools defined with zod schemas
• RAG: pdf-parse → overlap chunking (tested) → embedMany → cosine retrieval, in-process
• Multi-provider: resolved per request from Authorization / X-LLM-Provider / OpenAI-proxy headers
• CI: lint → typecheck → test → build on every push; workspace-aware Docker builds

See docs/REFACTOR_PLAN.md for the full design.

Getting Started

Requirements: Node.js >= 20, pnpm >= 9

pnpm install

# configure the server
cp apps/server/.env.example apps/server/.env   # set OPENAI_API_KEY etc.

# start web (5173) + server (3000) together
pnpm dev

Or with Docker:

cp apps/server/.env.example apps/server/.env
docker compose up --build   # web on :8080, api on :3000

Scripts

License

MIT

Star History