AnythingMCP is an open-source, self-hosted MCP gateway whose goal is to convert any REST/OpenAPI, SOAP, GraphQL, OData, or SQL system into MCP tools for use by clients such as Claude, ChatGPT, and Copilot, without having to write your own MCP server. It comes with 474 ready-made connectors covering SAP, Etsy, weclapp, Amazon Seller, and others, 20 of which require no API key.
Purpose and positioning Most companies do not yet have an MCP server, but they already have REST APIs, ERP systems, legacy SOAP services, and databases. AnythingMCP starts at this layer, turning these systems into a set of MCP tools, then serving, scoping, and auditing them behind a single endpoint. If you already have an MCP server and only need federation and a security proxy, the official documentation suggests that a pure gateway may already be sufficient.
Main capabilities · Connect any API, SOAP service, or database: you can install adapters from the catalog, or point to your own OpenAPI specification, WSDL, GraphQL endpoint, Postman collection, or database. Once configured in a workspace, it becomes a connector, and each of its operations is an MCP tool. Tools are registered at runtime, with no restart required. Each connector's {{VAR}} values are interpolated server-side and are not shown to the AI. · Connector catalog: 474 adapters exposing 2,400+ tools, each with a setup guide, available in seven languages. Adapters are single JSON files, so the catalog is large and the barrier to adding one is low. If your ERP is not listed or is a custom-built system, you can connect directly to its REST API, SOAP service, or SQL database (read-only). · Control what the model can see: rename and describe tools in the visual editor, remove fields that should not leave the network, and choose which roles can call which tools. · One URL to hand to AI clients: the MCP server is an endpoint added to Claude, ChatGPT, Copilot, Gemini, or Cursor, exposing only the set of connectors assigned to it.
Security and governance · Everything runs on your own infrastructure, and you decide what may leave it. OAuth2, RBAC, SSO, and SCIM are all available in the self-hosted version, not placed behind a paid tier. · Response mapping: each tool declares which fields reach the model,Customer IBANs, employee salaries, or designated retention fields can be discarded. The editor shows a before-and-after comparison of real responses; in the official example, one adapter went from 12,172 B down to 1,072 B (−91%). If mapping fails, the original response is returned by default; for tools whose fields must never be sent externally, set "fallbackToRaw": false to make the call fail. · Read-only controls: each tool carries MCP annotations (readOnlyHint, destructiveHint), derived from the operation and overridable per tool. Role-based tool allowlists can publish MCP servers that can only read, which is the officially recommended starting approach for ERP. Database query tools execute a single SELECT, blocking writes and stacked statements. · Authentication methods: OAuth2 (PKCE and Client Credentials), Bearer, API key, Basic, HMAC request signing, LOGIN_TOKEN, and OAuth 1.0a. Credentials are encrypted at rest using AES-256-GCM. · Audit logs: every tool call records input, output, duration, and status, stored in your own database, including the full upstream response that the model did not see. · SSO and SCIM: supports Entra ID, Google, Okta, Auth0, or any OIDC provider. Roles are synced from directory groups on every login; disabling someone in the directory invalidates their workspace access and MCP API keys.
How clients use it he same MCP server works with all MCP-compatible clients, so the connector only needs to be built once. · Claude: under Customize → Connectors, add the server URL as a custom connector, then use it in Claude.ai, Claude Desktop, and Claude Code. OAuth 2.0 is supported out of the box. On AnythingMCP Cloud, it can also be added with one click from the Claude Directory. · ChatGPT: Apps in ChatGPT are based on MCP. Add the server under Plugins (+ → Add custom MCP server), or use it as the tool layer for an Apps SDK app. · Meta Muse: in Muse, open Settings → Connectors → Add custom connector, paste the server URL and log in to AnythingMCP. · Copilot, Gemini, Cursor, and other MCP clients: corresponding client setup guides are available.
Knowledge graph and AI skills Forwarding calls leaves the hard parts to the agent: knowing which tool to call next, and what "open orders" means in the business. AnythingMCP learns both and returns them as context rather than as extra tool calls. The knowledge graph is each workspace's map of entities (customers, orders, products) and their relationships across connectors, built from tool definitions and real calls, manually editable, storing only field names and relationships, never values. Each server exposes it through the kg_how_to_obtain tool, so an agent can ask how to get a DHL tracking number from a Shopware order. AI skills turn repeated use into small rules ("today's revenue includes order statuses 2, 3, and 4"), which can be applied, edited, or ignored, and combined into server instructions. AI processing is off by default and uses the user's own OpenAI, OpenRouter, or Anthropic key; the knowledge graph, editor, and MCP tools themselves require no key.
Run it yourself Requires Docker 24+ and openssl; on macOS, start Docker Desktop first. Run it without cloning by visiting http://localhost:3000 to sign up; the first account becomes the administrator. The MCP endpoint is http://localhost:4000/mcp，API Documentation is at http://localhost:4000/api/docs。请保留生成的 .env, ENCRYPTION_KEY is used to decrypt stored credentials; if lost, all connectors must be reconfigured with credentials. On amd64, the image pull takes about 30 seconds, after which the API is healthy in about 24 seconds. The image currently only supports amd64, and the compose file fixes the platform, so it can also run on Apple Silicon through emulation. The quick start is intentionally bound to 127.0.0.1 because there is no component in front of it terminating TLS. For instances accessible to others or to cloud AI clients, clone the repository and run ./setup.sh: it will ask for the domain, obtain certificates through Caddy, generate keys, and set the MCP authentication mode.
Cloud Version AnythingMCP Cloud is the same AGPL code, operated officially in Frankfurt, Germany, and offers a 7-day free trial. You can use it first to try your own API, then migrate it in-house when you no longer want credentials to leave your premises; the connectors are the same in both approaches. A DPA/AVV can be requested via info@helpcode.ai. SSO and SCIM are available only in the self-hosted version.
License and Community The project uses the GNU Affero General Public License v3 (AGPL-3.0-only), and internal commercial use within a company is always included; copyleft obligations only arise if you modify AnythingMCP and provide the modified version to others over a network. The cloud operations code under ee/ is licensed separately and is not required for self-hosting. The official recommendation is to read the Contributing guidelines before submitting a PR; the easiest useful contribution is an adapter (a single JSON file). The launch case is KOCH Freiburg GmbH, used to connect AI assistants with 15+ internal systems (ERP, CRM, SOAP services, and local databases).
Limitations and Details Not Yet Provided The SOAP connector currently does not support signed WS-Security messages. Some adapters (such as certain ERP and e-commerce connectors) are built based on vendors' public API documentation and have not yet been actually verified on real tenants or seller accounts. The image currently only supports amd64. Regarding specific pricing, SLA, cloud version feature boundaries, and complete limitations, the official materials do not yet provide more information.
