AnythingMCP

Give Claude, ChatGPT and Copilot safe access to the software your company already runs.
189 ready adapters, any REST/SOAP/GraphQL/SQL system without code, on your own infrastructure — and it learns how your systems connect.

Claude answering a question no chatbot could answer, because the data lives in a field-service system that speaks REST, not MCP:

Run it yourself — three lines, no clone, details below:

mkdir anythingmcp && cd anythingmcp
curl -fsSLo docker-compose.yml \
  https://raw.githubusercontent.com/HelpCode-ai/anythingmcp/main/docker-compose.quickstart.yml
printf 'JWT_SECRET=%s\nENCRYPTION_KEY=%s\n' "$(openssl rand -hex 32)" "$(openssl rand -hex 32)" > .env
docker compose up -d   # → http://localhost:3000

Three words appear throughout and mean three different things:

• an adapter is one of the 189 JSON definitions that ship in this repo — DATEV, weclapp, DHL, Deutsche Bahn, Shopware, Personio, Handelsregister and the rest. 20 of them need no API key at all; the others ask for your credentials at import.
• a connector is an adapter, or your own OpenAPI spec / Postman collection / WSDL / GraphQL endpoint / database, once you have configured it in your workspace. Anything you can point at, in minutes, without writing an MCP server.
• an MCP server is the URL you hand to Claude. It exposes the connectors you assign to it, and nothing else.

Everything runs on your infrastructure, so you decide what leaves it. Per-tool response mapping declares which fields ever reach the model; credentials are AES-256-GCM at rest; the audit log keeps the full upstream response on your side. OAuth2, RBAC, SSO and SCIM are in the self-hosted build, not held back for a paid tier.

In production at KOCH Freiburg GmbH, where it connects AI assistants to 15+ internal systems — ERP, CRM, SOAP services and on-prem databases. AnythingMCP was extracted from that system by helpcode.ai in Freiburg, Germany, and open-sourced because an adapter catalog grows faster as a community than as a product.

Run it yourself

The recommended path, and the one measured below. Requires Docker 24+ and openssl; on macOS, start Docker Desktop first.

mkdir anythingmcp && cd anythingmcp
curl -fsSLo docker-compose.yml \
  https://raw.githubusercontent.com/HelpCode-ai/anythingmcp/main/docker-compose.quickstart.yml
printf 'JWT_SECRET=%s\nENCRYPTION_KEY=%s\n' "$(openssl rand -hex 32)" "$(openssl rand -hex 32)" > .env
docker compose up -d

Open http://localhost:3000 and register — the first account becomes admin.

Keep the generated .env. ENCRYPTION_KEY is what decrypts the credentials you store. Lose it and every connector has to be re-credentialed; back it up wherever you keep your other secrets.

Measured on amd64: 31 s to pull the image, 24 s to a healthy API and a login page. The published image is amd64 only for now — the compose file pins the platform so it also runs on Apple Silicon under Docker Desktop's emulation, where the same boot took 24 s on an M-series laptop but can take a few minutes on older hardware.

The quickstart binds to 127.0.0.1 on purpose: nothing in front of it terminates TLS. For an instance other people or a cloud AI client can reach, clone the repo and run ./setup.sh — it asks for a domain, gets certificates through Caddy, generates the secrets and sets the MCP auth mode. See the Deployment Guide.

What it connects, governs and learns

Connect

• 5 connector types — REST, SOAP, GraphQL, Database, MCP-to-MCP bridge. Seven database engines: PostgreSQL, MySQL, MariaDB, MSSQL, Oracle, MongoDB, SQLite.
• Import from what you already have — OpenAPI/Swagger, Postman, cURL, WSDL, GraphQL introspection, or tool discovery straight from a running MCP server.
• The adapter catalog — see what ships.
• Visual tool editor — map parameters to path, query, body and headers; rename and describe tools so the AI reads them the way you meant.
• Dynamic MCP server — tools register at runtime, no restart. Per-connector {{VAR}} interpolation, hidden from the AI.

Govern

• Response shaping — declare per tool exactly which fields reach the model, with a live before/after preview.
• Read-only where it matters. Every tool carries MCP annotations (readOnlyHint, destructiveHint), derived from the operation and overridable per tool, so a client can show the difference between reading an invoice and issuing a credit note. Role-based tool whitelisting lets you publish an MCP server that can only read — which is how most people should start with an ERP.
• Every auth scheme you will meet — OAuth2 (PKCE and Client Credentials), Bearer, API Key, Basic, WS-Security, client certificates, LOGIN_TOKEN and OAuth 1.0a.
• Audit logging — every tool call recorded with input, output, duration and status, in your own database.
• SSO and SCIM — Entra ID, Google, Okta, Auth0, generic OIDC. Roles sync from your directory groups on every sign-in; disable someone in the directory and their workspace access and MCP API keys die with it (self-hosted only).

Learn

• Knowledge Graph — a per-workspace, PII-safe map of how your connectors' data relates, served back to the agent as an MCP tool so it chains calls across systems correctly.
• AI skills — recurring usage turned into small reusable rules and composed into the server's instructions, so they guide the agent without adding a tool call (optional, opt-in).

How AnythingMCP compares

The projects it gets compared with are mostly MCP gateways: they federate, scope and secure MCP servers you already have. AnythingMCP starts one step earlier, because most companies have no MCP servers at all — they have a REST API, a SOAP service from 2009 and a database nobody wants to expose. Every project below solves a real problem; they just don't solve the same one.

Side-by-side pages with the full feature tables: anythingmcp.com/vs.

Knowledge Graph & AI skills

Forwarding calls leaves the hard part to the agent: knowing which tool to call next, and what your business actually means by "open order" or "active customer". AnythingMCP learns both — how the data in your connectors relates, and how your team really uses the tools — then feeds that back to the AI client as context rather than as extra tool calls.

• Knowledge Graph — a per-workspace map of entities (customers, orders, products…) and their relationships. It builds itself from tool names, parameters and the input/output of real calls; an optional AI pass infers the cross-connector links heuristics miss. It stays PII-safe: it stores entity/field names and relationship metadata, never the values.
• Build it visually — a graph editor lets you create, edit and delete entities and connections by hand, add descriptions, and curate what the AI proposed.
• Served over MCP — each server exposes a kg_how_to_obtain tool, so the customer's agent can ask "how do I get from a Shopware order to a DHL tracking number?" and receive chaining hints across connectors.
• AI skills, written from real usage — with intent capture on, each tool call can record why it was made. An AI pass turns recurring patterns into small reusable rules (e.g. "today's revenue includes order statuses 2, 3 and 4"). You Apply, Edit or Dismiss each one, or let auto-apply take the high-confidence ones unattended. Applied skills are composed into the MCP server's instructions at serve time, so they guide the agent without adding a single tool call. The knowledge your team builds up by using the system stops living in someone's head.

The AI passes are off by default — opt in with a global env flag and a per-workspace switch, using OpenAI, OpenRouter or Anthropic. The graph, manual editing and the MCP tool work with no LLM key at all.

➡️ Knowledge Graph & AI skills guide →

Control what the model sees

Every tool can declare exactly which fields leave your infrastructure. The mapping is attached per tool and applied on the way out, so the AI client — and the third-party model behind it — only ever receives the shape you approved.

• Drop what should never travel. List the paths to remove and they are stripped before the response reaches the agent: a customer's IBAN, an employee's salary, an access token an API hands back alongside the data.
• Or declare the whole output. A select template names the fields to keep and what to call them; a JMESPath expression covers the reshaping a template can't express. Where an agent is better served by a stable shape, swap the value for a placeholder ("iban": "= [redacted]") instead of removing the field.
• See it before you save it. The editor runs the mapping against a real response and shows the before/after side by side, with the size difference. A shipped adapter measures 12,172 B → 1,072 B (−91%) on a four-train result.
• It fails open by default, and this says so. If a mapping breaks at runtime, the raw response is returned and a warning is logged, so one bad expression cannot take a working tool offline. That default is wrong for fields that must never travel: set "fallbackToRaw": false on those tools and a broken mapping fails the call instead of leaking through it.

Two payoffs at once: sensitive fields never reach the model, and every field you drop is a field you don't pay for in the context window.

{
  "transform": {
    "mode": "select",
    "fallbackToRaw": false,
    "exclude": ["customer.iban", "customer.taxId"],
    "select": { "order": "$.id", "total": "$.amounts.gross", "status": "$.state" }
  }
}

The audit log still records the full upstream response inside your own database. Shaping what the agent sees never costs you the evidence of what the API actually returned.

➡️ Response mapping reference →

Build custom Claude connectors — no code

Claude supports custom connectors: remote MCP servers you add once in Settings → Connectors, and that work across Claude.ai, Claude Desktop and Claude Code. AnythingMCP creates that connector from any API you already have — without writing an MCP server:

1. Import your API spec, or pick a pre-built adapter
2. Adjust tool names, descriptions and parameters in the visual editor — what the AI sees is up to you
3. Add your MCP server's URL to Claude as a custom connector (OAuth 2.0 supported out of the box)

Your credentials stay on your infrastructure, every tool call lands in the audit log, and role-based access controls which users see which tools. Step-by-step guide →

Turn your API into a ChatGPT app

Apps in ChatGPT are built on MCP, and AnythingMCP gives you that MCP backend without writing one. Point it at your REST, SOAP, GraphQL or database endpoint and you get a ChatGPT-ready connector: add it in ChatGPT's settings (or use it as the tool layer of an Apps SDK app) and ChatGPT can read and act on your business data.

The same connector works simultaneously in Claude, ChatGPT, Gemini, Copilot and Cursor — build once, connect everywhere. ChatGPT setup guide →

Why AnythingMCP

AI clients speak MCP, but your systems speak REST, SOAP, GraphQL and SQL. Writing and maintaining a bespoke MCP server per system — with auth, audit and access control — takes weeks each. AnythingMCP is the no-code layer in between:

What people actually build with it

The adapter catalog

189 adapters, exposing 1,800+ tools. 20 need no API key; the rest ask for your credentials at import and the tools are available immediately. Every one has a setup guide on anythingmcp.com/guides, in seven languages.

An adapter is a single JSON file. That is why the catalog is this size, and why adding one is a reasonable first contribution. Missing yours? Request it — we prioritise by 👍 — or build it.

Guides, client setup & FAQ

➡️ docs/guides.md — Claude / ChatGPT / Gemini / Copilot / Cursor setup · REST / SOAP / GraphQL / Database / MCP-bridge connector guides · API reference & deployment docs · FAQ.

Looking for a specific service? Every adapter has a step-by-step guide at anythingmcp.com/guides.

Community & support

• 💬 Questions & discussions — GitHub Discussions — vote on the next adapter, share what you've built
• 🐛 Bugs / 💡 features — Issues · 🆘 SUPPORT.md
• 🔐 Security — please do not open a public issue; follow SECURITY.md
• 🏢 Built by helpcode.ai in Freiburg, Germany. AI-assisted development, human-reviewed — AUTHORS.md says which parts and how.

Contributing

Read the Contributing guide before opening a PR. The easiest useful contribution is an adapter: one JSON file, and there is a walkthrough issue for it.

License

Open source under the GNU Affero General Public License v3 (AGPL-3.0-only). Commercial use inside your own company is included and always was; the copyleft obligation only starts if you modify AnythingMCP and offer the modified version to others over a network. Cloud-operator code under ee/ is separately licensed and is not required for self-hosting — see the License FAQ.

⭐ If this saved you a week of writing MCP servers, star it.
Stars are how the next person finds it — and how we decide which adapter to build next.

Star history