Skip to main content
You can connect an AI agent to Fluid in four ways. Pick by what the agent needs to do: Most agents use the first two together: the docs MCP server to look up how something works, and the CLI to do it.

Docs MCP server

Fluid’s docs come with a hosted, read-only Model Context Protocol (MCP) server at:
There’s nothing to install, and it needs no API token. Connected agents can:
  • Search the docs, guides, and API reference.
  • Read any docs page or OpenAPI spec in full.
  • Send feedback about a docs page that’s wrong or missing something.
It only reads published documentation. It can’t read or change your company’s data, and it doesn’t accept an API key.

Connect from a docs page

Every docs page has a menu at the top. Use it to copy the MCP server URL or install the server in Cursor or VS Code in one click.

Connect by hand

Run this in any terminal. Claude Code keeps the server in your configuration.
The same setup steps are in the admin, on Settings → Developer. See Developer.

Without MCP

Agents that can’t use MCP can still read the docs:
  • /llms.txt indexes every page for language models.
  • /skill.md is a short guide for AI tools: where to look, Fluid’s naming, and the main conventions.
  • Every page can be copied as Markdown from its page menu.

Fluid CLI

To make changes, give the agent the Fluid CLI. It works on your company’s themes, and its plugins run Fluid’s guided setups, such as connecting a back office, opening a country, or setting up payments onboarding. Every guided-setup command prints JSON, so an agent can read the result of each step. Sign the CLI in one of two ways:
  • Your own sign-in. Run fluid login and enter the code Fluid emails you. Use this when you’re at the keyboard with the agent. An agent with no terminal to prompt in signs in in two steps: fluid login --send-code --email <email> --shop <shop> emails the code, then FLUID_LOGIN_CODE=<code> fluid login --code --email <email> --yes --json redeems it. Use --code - to read the code from stdin instead. See the agent launch guide for the full flow.
  • An API token. Create one at Settings → API Tokens, then sign in with it. Pass it in an environment variable so it stays out of your shell history:
    Use this when the agent runs on its own. See Authentication for token types.
Run fluid theme skills install to give your agent Fluid’s theme skills as well. To set up a whole store with an agent, follow the agent launch guide. The docs MCP server and the CLI are separate. The MCP server never sees your token.

Keep the CLI current

The CLI and its plugins get new commands, options and config keys often, and an older copy doesn’t know them. When a fluid command, option or config key isn’t found, run fluid version. It lists the CLI and each plugin you use with the latest version on npm, and the command that updates each one that’s behind. fluid version --json gives the same as JSON, for an agent. If fluid version itself isn’t found, the CLI predates it. Update the CLI and its plugins where they’re installed: with npm install -g for a global install, with your project’s package manager for a CLI in a project’s node_modules (which fluid shows which one runs), or by updating Mist for the CLI Mist brings. Once a day, the CLI checks npm for a newer CLI and Sites plugin, and prints a short update notice on stderr after a command when one is out. With CLI 0.1.763 or later, a coding agent such as Claude Code, Codex, Cursor or Gemini CLI sees it as one stderr line, at most once a day. Set FLUID_AGENT=1 to mark another agent’s runs. The notice never prints after a command that prints JSON, or in CI. Set FLUID_NO_UPDATE_CHECK=1 to turn it off.

Fluid MCP server

The Fluid MCP server, at https://mcp.fluid.app/mcp, gives a chat client tools that act on your company, such as listing, creating and deploying Sites. In Claude Code:
  • It signs in with your Fluid account, in a browser window, when the client first connects. It doesn’t accept an API token, so an agent with no browser can’t use it. Use the Fluid CLI with a token there instead.
  • Changes need your say-so. Tick Let this app make changes when you sign in, or the client sees only the tools that read.
The Site tools that make changes are limited to Fluid root admins today. See Build and deploy a Site from Claude for the Site tools and who may use each one.

Mist

Mist is Fluid’s own AI assistant, a desktop app that builds and changes things for your company when you chat with it. You sign in with your Fluid account, so it doesn’t need an API token or an MCP setup. The Mist Skills repository has the skills and workflows Mist uses. Many of them also work in other agents, such as Claude Code, outside Mist.

Keep agents from publishing on their own

Some CLI commands change what shoppers, members or reps see the moment they run. Give your agent this rule: publish only when you ask it to publish.
  • Building isn’t publishing. Editing, previewing, fixing, or passing a QA check doesn’t give the agent permission to publish.
  • Each publish needs its own request. An earlier publish doesn’t cover later changes.
  • When the agent can’t tell whether a target is live, it treats it as live.
These are the commands that change live content: These keep changes out of sight until you publish:
  • fluid theme push --unpublished creates a new unpublished theme and pushes to it.
  • fluid theme dev syncs to a development theme, a scratch copy your storefront doesn’t serve. fluid theme push --publish refuses to publish one.
  • fluid portal push without --activate updates the portal’s draft. See Portal CLI reference.
In a terminal, fluid mysite publish asks you to confirm. An agent with no terminal to prompt in can’t answer, so the command refuses to run unless you pass --yes. Only add --yes when you’ve asked for the publish.