How AI chat and MCP work

Anythink gives AI models access to your project in two ways. The AI chat in the Anythink dashboard answers questions and makes changes through tools. The MCP server, anythink-mcp, lets Claude Code, Claude Desktop, Cursor or another Model Context Protocol client work on your project from outside the dashboard.

Both can act, not just answer. They can create entities and fields, write and delete records, and build workflows, and both do it with your permissions. For calling a model as a step inside an automation instead, see the Integration step in Step types.

AI chat or MCP

Use the chat for quick questions and changes while you're in the dashboard. Use MCP when you want the assistant in your editor, working on your code and your backend together.

The moving parts

How a chat message runs

1. You send a message in the AI panel, optionally with image or PDF attachments.
2. Anythink picks the provider from the model name. Models starting gpt-, o1, o3 or o4 go to OpenAI, grok- models go to Grok, gemini- models go to Gemini, and everything else, including all Claude models, goes to Claude.
3. It picks the credentials. In Bring Your Own Key mode it uses your enabled connection for that provider. In Anythink AI mode, or if you have no connection for that provider, it uses the platform key and checks your AI token allowance first.
4. It runs the conversation loop. The model replies, and if it asks for a tool, Anythink runs the tool as you and feeds the result back. This repeats until the model stops asking for tools, up to 10 rounds.
5. The reply streams back as it's generated, with a marker each time a tool runs.
6. The conversation is saved so you can reopen it from Chat history in the panel.

Before the first round, the model gets a short summary of your project: how many entities it has and the names of the 10 most recently created. Anything more, it has to fetch with a tool.

What the chat can do

When the model asks for a tool, the panel sends the command to a copy of the CLI's MCP server that Anythink hosts. That server runs the command in your project with the same sign-in you're using in the dashboard. The commands it supports are:

• entities: list, get, create and delete entities.
• fields: list, add and delete fields.
• data: list, get, create, update and delete records.
• workflows: list, get, create, update, delete, enable, disable and trigger workflows, add and edit their steps, and read their jobs.
• users and roles: list and get.
• secrets: list, create and delete.
• files: list, get and delete.
• fetch: make an authenticated request to any REST endpoint in your project.
• docs: read the API reference.

Because fetch can call any endpoint, the chat can do more than the commands above, within your permissions. Integrations, menus, payments, migrations and search run only through a local MCP server.

The chat acts as you, and a tool call runs as soon as the model makes it. The panel doesn't pause to ask you first. Ask it to explain a change before it makes one, and use a role with narrow permissions for people who shouldn't be able to delete things. See Control access: roles, permissions and field-level security.

Anythink AI or Bring Your Own Key

Choose the mode under Settings, then AI Settings, then AI Provider.

In Bring Your Own Key mode, the panel lists only models that work with your own keys. If you pick a model whose provider has no saved key, Anythink falls back to the platform key and your allowance. To use OpenAI, Grok or Gemini in chat, save a key for that provider first. The Usage card in AI Settings shows your token consumption for the last 30 days.

The same provider connections power the Integration workflow step. See How integrations work.

The MCP server

anythink-mcp is published from the public anythink-cli repository. Your MCP client starts it as a local process. It reads the same saved profiles as the anythink CLI, and its cli tool runs your installed anythink binary, so the assistant can run any CLI command you could. To set it up, see Connect Claude to your backend with MCP.

The server has 16 tools: sign-in (signup, login, login_google, login_direct, logout), profiles (config_show, config_use, config_remove), billing accounts (accounts_list, accounts_create, accounts_use), projects (projects_list, projects_create, projects_use, projects_delete) and cli.

The server also has an --http mode. That's a plain REST service that the dashboard's chat uses, not an MCP endpoint, and you can't connect your own MCP client to it. Always run anythink-mcp in its default stdio mode.

Permissions and safety

The AI chat and the MCP server act as you, or as the API key you connected with. There's no separate AI permission that widens access: a tool call succeeds only if your role allows the same action through the CLI or REST API.

Tool calls need the permission for what they touch. Listing entities needs anythink_entities:read, and creating a workflow needs anythink_workflows:create. Row-level security applies too, so a request only sees the records its user can see.

Ways to keep an assistant in bounds:

• Give it a scoped key. With MCP, connect with an API key that has only the permissions the task needs, and :read permissions for an assistant that should only look. See API keys.
• Keep credentials out of the chat. Sign in with the CLI, and create keys in the dashboard or the CLI, so a key or password never passes through a model. The dashboard chat can't create API keys.
• Read before you approve. Most MCP clients ask before running a tool. Check writes, and especially deletes, before you allow them.

Limits

Next steps