Connect Claude to your backend with MCP

Connect Claude to Anythink and it can create entities and fields, read and write records, build and debug workflows, and manage integrations in your project, using the same commands you would run with the CLI. There are two ways to connect:

• The hosted connector at https://mcp.anythink.cloud/mcp. Nothing to install: add the URL, sign in with your Anythink account and choose a project.
• The local server, anythink-mcp, which your MCP client starts on your machine over stdio. Use it for clients that don't support remote connectors, or to sign in and switch projects from the chat.

For how this relates to the AI panel in the Anythink dashboard, see How AI chat and MCP work.

Connect the hosted connector

You need an Anythink account and a project to connect to.

Claude (web, desktop and mobile)

1. In Claude, open Settings, then Connectors, and select Add custom connector.
2. Name it Anythink and enter the URL https://mcp.anythink.cloud/mcp.
3. Select Connect, sign in with your Anythink account, and choose the project Claude can use, or all of your projects.
4. In a chat, turn on the Anythink connector from the tools menu.

On a Team or Enterprise plan, an owner adds the connector for the organisation first, then each member connects their own Anythink account.

Claude Code

claude mcp add --transport http anythink https://mcp.anythink.cloud/mcp

Then run /mcp in Claude Code and choose Authenticate. Your browser opens so you can sign in and choose a project.

Other MCP clients

The connector signs clients in with OAuth using a client ID metadata document (CIMD), which is how Claude and Claude Code identify themselves. Other clients that support CIMD can use the same URL. Support for dynamic client registration (DCR), which clients such as Cursor and VS Code use, is planned; until then, connect them with the local server below.

One project or all of them

When you sign in, you grant access to one project or to all of your projects.

Tools run as you, so your role in each project decides what a call can do. To stop Claude reaching your projects, disconnect or remove the connector in Claude's settings.

What the hosted connector can do

The hosted connector has a tool for every project command: entities, fields, records, workflows and their steps and jobs, roles, users, secrets, files, integrations, menus, search, payments, API keys and the API docs. Each tool is labelled read-only, write or destructive, so Claude can check with you before changing or deleting anything. For the full list, see MCP server tools reference.

It leaves out tools that only make sense on your own machine: signing in or out, switching saved profiles, opening a browser, reading or writing local files, raw fetch requests and creating API keys. Use the local server or the CLI for those.

Run the server locally

The local server shares its saved profiles with the anythink CLI, so you sign in once for both.

Before you start

• An Anythink account and a project to connect to.
• The anythink CLI on your PATH. The server's cli tool runs your installed anythink binary, so without it every cli call fails with a "CLI not found" error. Homebrew installs both together. See Getting Started with the CLI.

Install the server

Pick one. They all give you the same server.

You can also download an anythink-mcp-<platform> binary from the anythink-cli releases page.

Register it with your client

Every client launches the same command. The examples use npx. If you installed with Homebrew or the .NET tool, use "command": "anythink-mcp" and drop the args.

Claude Code

claude mcp add anythink -- npx -y @anythink-cloud/mcp

With a native install:

claude mcp add anythink anythink-mcp

Or add it to your project's .mcp.json:

{
  "mcpServers": {
    "anythink": {
      "command": "npx",
      "args": ["-y", "@anythink-cloud/mcp"]
    }
  }
}

Claude Desktop

Open Settings, then Developer, then Edit Config. This opens claude_desktop_config.json. Add the server under mcpServers, then restart Claude Desktop.

{
  "mcpServers": {
    "anythink": {
      "command": "npx",
      "args": ["-y", "@anythink-cloud/mcp"]
    }
  }
}

Claude Desktop starts the server without your shell's setup. If it can't find npx, install with Homebrew and use "command": "anythink-mcp" instead.

Cursor

Add the same mcpServers entry to ~/.cursor/mcp.json. Cursor, VS Code (.vscode/mcp.json), Cline and Windsurf all use this shape. Continue and Zed use their own formats, which the anythink-cli README shows.

{
  "mcpServers": {
    "anythink": {
      "command": "npx",
      "args": ["-y", "@anythink-cloud/mcp"]
    }
  }
}

Pin a profile

The server uses the CLI's active profile by default. To pin one saved profile, add --profile to the arguments:

{
  "mcpServers": {
    "anythink": {
      "command": "npx",
      "args": ["-y", "@anythink-cloud/mcp", "--profile", "my-project"]
    }
  }
}

Sign in and pick a project

The server shares its saved profiles with the anythink CLI, so you can sign in either way. Signing in with the CLI keeps your password out of the conversation with the model.

anythink login
anythink accounts use <account-id>
anythink projects use <project-name>

Connect with an API key

If you already have a project's org id and an API key, skip the billing account step. Run this in a terminal so the key never passes through the model:

anythink login --org-id <org-id> --api-key <api-key>

The MCP equivalent is the login_direct tool with orgId and apiKey. Both save a profile named after the org id and make it active.

A key is the safest way to limit what an assistant can do. It holds only the permissions you gave it, so a key with read permissions can look at your data and can't change it.

What the assistant can do

Both servers give Claude a tool for every project command, each labelled read-only, write or destructive. The local server adds tools to sign in and manage saved profiles, billing accounts and projects, plus a general-purpose cli tool that runs any anythink command against the active profile. For every tool and its arguments, see MCP server tools reference.

The assistant can make destructive changes, including deleting entities, records and projects. Read what it plans to run before you approve a write. On the local server, connect with a scoped API key when you only need it to read.

Permissions

MCP calls have no permissions of their own. They run as the signed-in user, or with the API key's permissions. A call succeeds only if that identity could make the same request with the CLI or REST API, and anything else returns 403.

To give an assistant access to look but not change anything, connect the local server with an API key that has only :read permissions on what it needs. See API keys and Control access: roles, permissions and field-level security.

Example prompts

List every entity in this project and the fields on each.

Create an entity called "feedback" with a text field "message" and an
integer field "rating", then add three sample records.

Show me the recent jobs for the "Flag large orders" workflow and tell me
why the failed ones failed.

Connect Slack, then add a step to the "New order" workflow that posts the
order total to #sales.

Limits

Next steps