Connect your AI assistant

Who this is for: people who want their AI assistant to create, test and fix workflows in Liitora.

Before you start: you need the address of your Liitora MCP server, for example https://mcp.liitora.example/mcp, and a Liitora API token that may do more than read, for example one that may also Create and change workflow drafts. Only admins can create API tokens: see Create API tokens.

Liitora works with any assistant that supports remote MCP servers (the Model Context Protocol) over HTTP. The assistant sends your token in an Authorization: Bearer header with every request. The token gives it access to one company only.

To use Claude.ai or ChatGPT, which sign in instead of sending a token, see Connect Claude or ChatGPT.

Pick what the token may do

Under What the token may do, keep the reading permissions that are ticked to start with, and tick what you want the assistant to do as well:

So that the assistant can	Tick
Look at workflows, runs, connections and secret names	the reading permissions (ticked to start with)
Create and change drafts, run their tests and try them in the test environment	Create and change workflow drafts and See runs and their results
Set up connections	Manage connections
Ask a person for a secret	Create and change workflow drafts (the assistant never sees or sets the value)
Create webhook endpoints	Manage webhooks
Switch on and pause workflows that a person approved	Start, pause and switch on approved workflows

Without a permission, the tool that needs it answers forbidden and names the missing permission. No token can approve a workflow or show a secret value: a person always does that.

Claude Code

Open a terminal.

Run this command, with your MCP address and your token:

claude mcp add --transport http liitora https://mcp.liitora.example/mcp \
  --header "Authorization: Bearer liitora_your_token_here"

Start Claude Code and run /mcp to check that liitora is connected.
Ask: “Ping Liitora.”

GitHub Copilot in VS Code

In your project, create the file .vscode/mcp.json.

Paste this, with your MCP address. VS Code asks for the token the first time and stores it for you, so the token is never written in the file:

{
  "inputs": [
    {
      "type": "promptString",
      "id": "liitora-token",
      "description": "Liitora API token",
      "password": true
    }
  ],
  "servers": {
    "liitora": {
      "type": "http",
      "url": "https://mcp.liitora.example/mcp",
      "headers": { "Authorization": "Bearer ${input:liitora-token}" }
    }
  }
}

Select Start above the liitora server in the file, and paste the token when VS Code asks.
Open Copilot Chat in agent mode, check that the Liitora tools are selected, and ask: “Which Liitora workflows failed this week?”

Other MCP clients

Any client that can call a streamable HTTP MCP server with a custom header works. Use the MCP address as the server URL and add the header Authorization: Bearer <your token>.

An AI assistant client for your company

If you prefer that your assistant signs in as its own identity instead of with a person’s API token, we can set up an AI assistant client for your company. It has the AI assistant role: it builds and tests workflows, but never approves, never sees secret values and can’t manage members. Use Report a problem in the user menu (top right) in the app.

What the assistant can and cannot do

It can search the action library, draft workflows with tests, validate them, run tests and test runs, read logs and explain failures.
It asks you for credentials with a secure link. You type the value in Liitora, never in the chat, and the assistant never sees it.
It can’t approve a workflow. A person approves every new version through the Approve & activate link the assistant shows you.
It can’t turn on a workflow that needs approval, even when it uses your own sign-in. You turn it on with the same link.
It can’t save a script with a password, key or token typed into it. It asks you to store the value as a secret instead.
The assistant’s work counts as the assistant’s, even with your API token or your own sign-in, so you may review and approve it yourself. The assistant itself never approves.
Every change it makes is in the audit log, marked as made by an API token or an AI assistant.

The assistant also gets four guided flows (MCP prompts): new-scheduled-report, new-webhook-automation, fix-failing-workflow and explain-workflow. In clients that show prompts, pick one to start.

If a tool call fails, the assistant gets a message that says what went wrong and how to fix it. The MCP tools reference lists every tool.

If it doesn’t connect

401 or “unauthorized”: the token is wrong, revoked or expired, or it is missing the Bearer prefix. See unauthorized.
The assistant can read but not save: the token is read-only. Create a new one with Read-only turned off.
“forbidden”: the person who created the token doesn’t have the permission the tool needs. See forbidden.