# Connect an agent

> Connect Claude Code, Cursor or Claude Desktop to Fransys through the MCP, with a read-only agent key.

Source: https://docs.fransys.cloud/en/agents/brancher-un-agent/ Checked against the console on 2026-10-23.

## What your agent will be able to do

Once connected to Fransys, your agent reads your projects, services, settings, clusters, logs and metrics, and prepares plans. Each plan waits for a person on your team to approve it in **Plans**: the agent never runs anything itself. See [What an agent can do](https://docs.fransys.cloud/en/agents/ce-qu-un-agent-peut-faire/).

## The steps

Everything happens on the console's **Connect** page.

1. **Check the MCP server**

   At the top of the page, **The MCP server is responding.** confirms that everything is ready, with the number of tools available and the server address.

2. **Create an agent key**

   Give it a name (for example “Claude Code – Kevin”) and choose what it can read. Tick **Also read billing** if the agent needs to see your costs. Click **Create key**.

   The secret is only shown once: copy it now. It's also already in the commands of the next step. The key becomes active 10 to 15 seconds after it's created.

3. **Add Fransys to your agent**

   Copy the command or the configuration for your tool from the page: the server address and your key are already in it.

   **Claude Code**

   In your terminal, at the project root:

   ```sh
   claude mcp add --transport http fransys <server address> --header "Authorization: Bearer <your key>"
   ```

   Then check with `claude mcp list`: “fransys” should show as connected.

   **Cursor**

   In `.cursor/mcp.json` at the project root, or in `~/.cursor/mcp.json` for all your projects:

   ```json
   {
     "mcpServers": {
       "fransys": {
         "url": "<server address>",
         "headers": { "Authorization": "Bearer <your key>" }
       }
     }
   }
   ```

   Reopen Cursor's MCP settings: “fransys” should list its tools.

   **Claude Desktop**

   In Claude Desktop, open **Settings › Developer › Edit Config**, then add the server to `claude_desktop_config.json`:

   ```json
   {
     "mcpServers": {
       "fransys": {
         "command": "npx",
         "args": ["-y", "mcp-remote", "<server address>", "--header", "Authorization:${AUTH_HEADER}"],
         "env": { "AUTH_HEADER": "Bearer <your key>" }
       }
     }
   }
   ```

   Restart Claude Desktop. The `mcp-remote` bridge is started by `npx`: Node.js must be installed on the machine.

   Then click **I've copied it, hide the secret**.

4. **Try it**

   For example, ask it:
   - “Go through my projects and tell me which services are disabled.”
   - “Read my API's logs from the last hour and explain the errors.”
   - “Add a Redis to my project and prepare the plan.”

   When the agent prepares a plan, it lands in **Plans**, for you to approve or refuse.

## One key per agent

Create one key per agent or per person. The page lists your organisation's keys: who created each one, and when it was last used. The activity log says which key did what.

**Remove key** cuts the agent's access right away. The plans it prepared stay in **Plans**.

Don't share a key: it gives read access to your data. If you've lost a secret, create a new key and remove the old one.

## What agents have done

At the bottom of the **Connect** page, **Latest agent activity** lists the plans that were prepared, edited or cancelled, and the tools used. **Full activity log** opens the **Activity** page.