There is no fixed catalogue here: you import a spec or point at a database, decide in a visual editor what the model sees, and the endpoint serves it. That makes the parts a plain forwarder skips the interesting ones — a Knowledge Graph that tells an agent how to chain tools across two systems, tool-level whitelisting per role, and an audit log holding the input and output of every call. Do the first-account step immediately after setup; whoever registers first becomes Admin.
An open-source MCP gateway you run on your own infrastructure. Point it at an API or a database and the resulting tools are registered at runtime and served from one MCP endpoint — no SDK, no restart, no code. Five connector types: REST, SOAP, GraphQL, Database (PostgreSQL, MySQL, MariaDB, MSSQL, Oracle, MongoDB, SQLite) and an MCP-to-MCP bridge. Six import formats: OpenAPI/Swagger, Postman, cURL, WSDL, GraphQL introspection and custom JSON. It ships with 175+ pre-built adapters. AGPL-3.0; the cloud-operator code under `ee/` is separately licensed and not needed for self-hosting.
- One MCP endpoint at `http://localhost:4000/mcp` serving tools generated from your connectors, registered at runtime with no restart
- A visual tool editor: map parameters to path, query, body and headers, and rename and describe tools so the model reads intent rather than endpoint names
- 175+ ready-to-use adapters across logistics, ERP, e-commerce, HR, public data, banking and messaging — Deutsche Bahn, DHL, DPD, GLS, weclapp, Xentral, Etsy, Shopware 6, WooCommerce, Personio, VIES VAT, UK Companies House, N26, Wise and more — each with its own setup guide
- A per-workspace Knowledge Graph of how your connectors' data relates, built from tool names, parameters and the input and output of real calls, editable by hand in a graph editor. It stores entity and field names and relationship metadata, never the values
- `kg_how_to_obtain` — the graph served back over MCP, so the agent can ask how to reach something and get chaining hints across connectors
- AI skills: business rules captured from how the tools are actually used, reviewed by a human, then composed into the server's instructions so they guide the agent without adding tool calls
- Auth on the connector side — OAuth2 with PKCE and Client Credentials, Bearer, API Key, Basic, WS-Security, client certificates, LOGIN_TOKEN and OAuth 1.0a handshakes
- Auth on the client side — `MCP_AUTH_MODE` selects `none`, `legacy`, `oauth2` or `both` and defaults to `oauth2`, alongside per-user MCP API keys
- Every tool call logged with input, output, duration and status, plus tool-level whitelisting per custom role
- Per-connector `{{VAR}}` interpolation for environment variables, hidden from the model, and credentials encrypted AES-256-GCM at rest
Docker 24+, `bash` and `openssl`; on macOS start Docker Desktop first. Clone the repository and run `./setup.sh`, which walks through deployment mode, domain and HTTPS (automatic Let's Encrypt through Caddy), secrets, MCP auth mode and optional SMTP and Redis. The Web UI comes up on `http://localhost:3000`, the MCP endpoint on `http://localhost:4000/mcp` and Swagger docs on `http://localhost:4000/api/docs`. Three environment variables are required: `DATABASE_URL` (a PostgreSQL connection string), `JWT_SECRET` (at least 32 characters) and `ENCRYPTION_KEY` (exactly 32 characters). Register the first account the moment setup finishes — the first user to register becomes Admin — and bind the UI to `127.0.0.1` or firewall it until you have. The AI passes for graph enrichment and skill generation are off by default and need both a global env flag and a per-workspace switch; they run against OpenAI, OpenRouter or Anthropic. The graph, manual editing and the MCP tool work with no LLM key at all. A hosted demo endpoint is published at `https://cloud.anythingmcp.com/mcp/demo`.
Build from source — clone the repository and build it, then deploy it and point your client at the endpoint