Skip to main content
Version: 5.x
Supported on
Snapchat
Spectacles

Developer Mode Using Claude Code

This guide will walk you through setting up Lens Studio's MCP server with Claude Code, allowing you to use AI-powered tools to control Lens Studio directly from the command line.

This guide is provided to support developers using Claude Code, as we've seen its growing popularity across our developer community. For official Claude Code MCP documentation, please refer to Claude Code's MCP Documentation.

Prerequisites​

Setup Steps​

1. Start the MCP Server in Lens Studio​

Go to Lens Studio and start the MCP server:

  1. In the Lens Studio menu bar, click AI Assistant > AI Model Context Protocol (MCP) > Configure Server

  2. In the MCP Server Configuration popup window that appears, click Start Server

2. Copy the MCP Configuration​

Click on Copy MCP Config in the Lens Studio MCP configuration window.

This will copy configuration details that look like this:

{
  "servers": {
    "lens-studio": {
      "headers": {
        "Authorization": "Bearer [your-auth-token]"
      },
      "type": "http",
      "url": "http://localhost:[your-port]/mcp"
    }
  }
}

You will need the URL and Authorization token from this configuration for the next step.

3. Add MCP Server via Claude Code CLI​

Option 1: Add MCP Server using Claude​

You can now provide the copied configuration directly to Claude Code to add the MCP server:

Please add the following MCP server: [your copied configuration].

Option 2: Add MCP Server manually​

Open your terminal and run the following command, replacing [your-auth-token] and [your-port] with the token and port from your copied configuration:

claude mcp add --transport http lens-studio http://localhost:[your-port]/mcp \
  --header "Authorization: Bearer [your-auth-token]"

For example, if your copied configuration includes:

"Authorization": "Bearer FbyXG5Yq7-IJhLKrCVaTgvG-wbYM5iTM6o4EnrJcR_UqESBr_d4NR1uP-IVPB0h7"
"url": "http://localhost:8732/mcp"

Your command would be:

claude mcp add --transport http lens-studio http://localhost:8732/mcp \
  --header "Authorization: Bearer FbyXG5Yq7-IJhLKrCVaTgvG-wbYM5iTM6o4EnrJcR_UqESBr_d4NR1uP-IVPB0h7"

By default, the MCP server is added to your local scope (current project only). You can specify different scopes:

  • --scope local (default): Available only to you in the current project
  • --scope project: Shared with everyone in the project via .mcp.json file
  • --scope user: Available to you across all projects

4. Verify the Connection​

You can verify that the Lens Studio MCP server is connected by running:

claude mcp list

This will show all configured MCP servers, including lens-studio.

Once inside Claude Code, you can also use the /mcp command to check the status of your MCP connections:

/mcp

5. Start Using Lens Studio with Claude Code​

Launch Claude Code and start interacting with your Lens Studio instance. Make sure your MCP server is still running in Lens Studio.

claude

Managing Your MCP Connection​

View Server Details​

To see details about your Lens Studio MCP configuration:

claude mcp get lens-studio

Remove the Server​

If you need to remove the Lens Studio MCP server:

claude mcp remove lens-studio

Update the Authorization Token​

When the authorization token expires or changes (for example, after restarting Lens Studio's MCP server), you'll need to update your configuration:

  1. Remove the existing server:

    claude mcp remove lens-studio

  2. Copy the new configuration from Lens Studio

  3. Add the server again with the new token:

    claude mcp add --transport http lens-studio http://localhost:[new-port]/mcp \
      --header "Authorization: Bearer [new-auth-token]"

Using a Custom Agent​

Claude Code supports custom agents that can help improve the AI's performance when working with Lens Studio projects.

  1. Create a lens-studio-agent.md file in a .claude/agents directory at the root of your project:

    mkdir -p .claude/agents
    touch .claude/agents/lens-studio-agent.md

  2. Add your custom instructions to this file.

For optimal results, consider using the Example Prompt to help improve Claude Code's performance with Lens Studio. This prompt includes guidelines for working with Lens Studio's TypeScript API and MCP tools.

  1. Now when you send requests to Claude Code it can reference your agent instructions.

Next Steps​

Now that you've connected Claude Code to Lens Studio, you can ask Claude Code to interact with your Lens Studio project. Here are some example prompts and tasks:

  • "What objects are in my scene?"
  • "Create a new 3D object called 'Cube' in my scene"
  • "Add a ScreenTransform component to the selected object"
  • "Search the Asset Library for face effects"
  • Request modifications to objects and components
  • Generate scripts and add interactivity
  • Debug issues in your project
  • Search and install assets from the Asset Library

The MCP server must be running in Lens Studio for the connection to work.

You might need to reconnect Lens Studio MCP in Claude if the plugin system has been restarted (for example: when adding a new ChatTool). You can do this by sending /mcp to Claude and ask it to reconnect, or sending it the Lens Studio MCP config again and asking the AI to add it.

Coordinating Multiple Agents with Agent Teams​

Claude Code supports an experimental agent teams feature that lets you run multiple Claude Code sessions working in parallel, coordinated by a lead agent. This can be useful for larger projects where different parts of the work are independent—for example, when building a Snap experience that involves both Lens Studio and a backend web service.

In a Snap Cloud project, you might spawn a team where one agent focuses on the Lens Studio side (interacting with the scene via MCP), another handles the web service layer, and a third writes integration tests—all working simultaneously and sharing a task list.

Create an agent team to build a new AR try-on feature:
- A "lens-studio" teammate to set up the scene, shaders, and scripts in Lens Studio using MCP
- A "web-service" teammate to update the API endpoints that serve the try-on assets
- A "qa" teammate to write integration tests covering the full flow

Each teammate works in its own context window and can message the others directly. The lead coordinates tasks, synthesizes results, and manages the shared task list.

Agent teams are experimental and disabled by default. For full details on enabling the feature, display modes, best practices, and current limitations, see the official Agent Teams documentation.

Was this page helpful?
Yes
No