Ganju

Turn your knowledge and integrations into a hosted MCP server that any AI client or chat bot can connect to.

Ganju is an open-source platform for building and hosting Model Context Protocol (MCP) servers. A tenant uploads resources (files, websites, cloud drives), wires up tools (Gmail, Slack, Calendar, web search, or their own HTTP/MCP backends), writes prompts, and gets a ready-to-use MCP endpoint at https://mcp.<domain>/<slug> — reachable from Claude Desktop, ChatGPT, Cursor, and from chat bots on Telegram, Slack, WhatsApp, and Discord.

It runs almost entirely on the Cloudflare developer platform (Workers, Queues, R2, Hyperdrive, Containers, Durable Objects) with Postgres + pgvector for storage and retrieval.

Ganju is live at ganju.ai — dashboard at app.ganju.ai, MCP servers at https://mcp.ganju.ai/<slug>, docs at ganju.ai/docs. The Free plan needs no card. Everything here is Apache-2.0, so running your own instance is a supported path rather than a fallback — self-hosting has no plan caps.

Features

• Hosted MCP servers — each "artifact" becomes a stateless MCP server with its own slug, OAuth, and tool set.
• RAG over your content — upload files, crawl websites, or sync Google Drive / OneDrive folders; content is chunked and embedded for vector search.
• A growing tool catalog — native integrations (Gmail, Outlook, Slack, Google Calendar, Cal.com, Tavily web search) plus two generic, no-code definitions:
  • http-endpoint — expose your own HTTP API as a tool.
  • mcp-proxy — connect a vendor's official MCP server (Notion, GitHub, …) and re-expose its tools.
• Tools you write yourself — when the catalog doesn't cover a case, write a tool in JavaScript and it runs on Workers for Platforms, with connected accounts, resource access and file-sending handed to it as host capabilities so your code never holds a refresh token. Author it in the dashboard's editor or from the ganju CLI, test it against real connections before publishing, and roll back by moving a pointer. See the Functions docs and the examples.
• Chat-channel bots — the same artifact can drive Telegram / Slack / WhatsApp / Discord bots through an LLM tool-calling loop (Gemini, Claude, or OpenAI). They answer voice notes as well as text, can remember each customer between conversations, and cap replies per customer per day.
• Multi-tenant — organizations → projects → artifacts, with invitations and roles.
• Audit & usage trails — every MCP request, tool call, and channel message is recorded.

Architecture at a glance

flowchart LR
  subgraph Clients
    A[MCP clients<br/>Claude / ChatGPT / Cursor]
    B[Chat bots<br/>Telegram / Slack / WhatsApp / Discord]
    C[Dashboard users]
  end

  A -->|MCP over HTTP| MCP[apps/mcp<br/>MCP server Worker]
  B -->|webhooks| API[apps/api<br/>control-plane Worker]
  C --> WEB[apps/web<br/>Next.js dashboard]
  WEB -->|REST| API

  API --> DB[(Postgres + pgvector<br/>via Hyperdrive)]
  MCP --> DB
  API --> Q[[Cloudflare Queues]]
  Q --> API
  API --> RH[apps/resource-handler<br/>Node container]
  MCP --> RH
  API --> R2[(R2 storage)]

See docs/ARCHITECTURE.md for the full picture.

Repository layout

ganju/
├── apps/
│   ├── api/              Control-plane Worker (Hono)
│   ├── mcp/              MCP-server Worker (Hono)
│   ├── resource-handler/ Node container for heavy work
│   └── web/              Next.js dashboard
├── packages/
│   ├── db/               Drizzle schema, migrations, connection
│   ├── utils/            Shared kernel: constants, crypto, oauth, chunking, send helpers
│   ├── ui/               MUI component library
│   ├── containers/       Cloudflare Container class wrapper
│   └── tsconfig/         Shared TypeScript config
├── docs/                 Project documentation (start here)
└── scripts/              Dev tooling

It's an npm workspaces + Turborepo monorepo.

Quick start

npm install
cp .env.example .env      # then fill in the values
npm run migrate-dev       # generate + apply DB migrations
npm run dev               # start all apps via turbo

Default local ports: API 8080, MCP 8081, resource-handler 8082, web 3000.

Full setup (prerequisites, environment variables, database, per-app commands) is in docs/DEVELOPMENT.md.

Tech stack

• Language: TypeScript everywhere
• Edge runtime: Cloudflare Workers, Queues, R2, Hyperdrive, Durable Objects, Containers
• Web framework: Hono (APIs) and Next.js (dashboard)
• Database: Postgres + pgvector, accessed with Drizzle ORM
• Auth: better-auth (social login + OIDC provider for MCP OAuth)
• AI: Gemini, Anthropic, and OpenAI SDKs; Gemini embeddings for RAG
• MCP: @modelcontextprotocol/sdk

Documentation

Contributing

Contributions are welcome — please read CONTRIBUTING.md first.

License

Apache-2.0 — see also NOTICE.