Bread MCP Server - Bread Documentation

What is the Bread MCP Server?

The Bread MCP (Model Context Protocol) Server allows AI assistants like Claude Code and Cursor to execute bakes directly on your behalf. Instead of manually running commands, you can describe what you want to your AI assistant and it will handle the baking operations.

New to baking? Start with the Quickstart to understand the basics of bgit and prompt baking first.

How It Works

Once configured, your AI assistant can:

Create and manage repositories
Set up prompts and targets
Run stim, rollout, and bake operations
Check status and retrieve outputs
Adjust configurations based on results

You describe what you want in natural language, and the AI assistant uses the MCP server to execute the actual baking operations.

Installation

Claude Code
Cursor

Prerequisites

API key from Bread (contact us)

Install Claude Code

curl -fsSL https://claude.ai/install.sh | bash

Add Bread MCP Server

claude mcp add bread \
  --transport http https://mcp.bread.com.ai \
  --header "Authorization: Bearer <YOUR-API-KEY>"

Replace <YOUR-API-KEY> with your actual Bread API key.

Verify Installation

claude mcp list

You should see bread in the list of configured MCP servers.

Start Using

Open Claude Code and try asking:

"Create a new Bread repository called 'my-model' and set up a Yoda personality prompt"

Claude will use the MCP server to execute these operations.

Prerequisites

Cursor IDE installed
API key from Bread (contact us)

Configure MCP Server

Open Settings

Click the gear icon in the top right corner of Cursor

Navigate to MCP Settings

In Cursor Settings, click on “Tools & MCP” tab

Add Custom MCP

Click “Add Custom MCP”

Update Configuration

In the mcp.json file that opens, add this configuration:

{
  "mcpServers": {
    "bread": {
      "url": "https://mcp.bread.com.ai",
      "headers": {
        "Authorization": "Bearer <YOUR-API-KEY>"
      }
    }
  }
}

Replace <YOUR-API-KEY> with your actual Bread API key.

Save and Restart

Save the file and restart Cursor if prompted

Start Using

In Cursor’s AI chat, try:

"Using the Bread MCP server, create a repository and start a bake for a Python expert personality"

Usage Examples

Once configured, you can interact with your AI assistant using natural language:

Creating a Repository

"Create a new Bread repository called 'customer-support-bot' 
using Qwen/Qwen3-32B as the base model"

The AI assistant will use the MCP server to create the repository.

Setting Up Prompts

"Set up a teacher prompt that makes the model act as a friendly 
customer support agent. Make the student prompt empty so the 
behavior is always on."

Running a Complete Bake

"Configure a target with 100 AI-generated questions, then run 
stim, rollout, and bake operations. Let me know when it's done."

The AI will:

Create the target configuration
Run stim to generate questions
Run rollout to capture responses
Run bake to train the model
Report back when complete

Checking Status

"What's the status of my current bake? Show me the training loss."

Iterating on Models

"The model is too formal. Update the teacher prompt to be more 
casual and friendly, then run another bake."

Available Operations

The MCP server provides access to all core Bread SDK operations:

Repository Management

Create repositories
List repositories
Get repository details
Configure base models

Prompts

Set teacher prompts (behavior to bake in)
Set student prompts (trigger for behavior)
Update existing prompts
List all prompts

Targets & Stim

Configure targets (stim + rollout settings)
Set up generators (AI-generated, hardcoded, from datasets)
Run stim operations
View stim outputs

Rollout

Run rollout operations
Check rollout status
View rollout outputs (model responses)

Bakes

Configure bake settings (epochs, LoRA, batch size)
Run bake operations
Check bake status and progress
Get training loss metrics
Retrieve baked model names

Best Practices

Be Specific

Provide clear instructions about what you want to bake. The more specific you are, the better the AI can configure the operation.

Check Status

Baking operations take time (30-60 minutes). Ask the AI to check status periodically rather than waiting idly.

Iterate Gradually

Make one change at a time, run a bake, test the results, then iterate. Don’t try to change everything at once.

Review Configurations

Ask the AI to show you the configuration before running expensive operations like bakes.

Troubleshooting

MCP Server Not Found

Symptom: AI assistant says it can’t find the Bread MCP serverSolutions:

Claude Code: Run claude mcp list to verify installation
Cursor: Check Settings → Tools & MCP to ensure the server appears
Verify the JSON syntax is valid (no trailing commas, proper quotes)
Restart Claude Code or Cursor completely
Ensure your API key is correct

Authentication Errors

Symptom: “Invalid API key” or “Unauthorized” errorsSolutions:

Verify your API key is correct in the configuration
Check for extra spaces or newlines in the API key
Ensure the Authorization: Bearer format is exact
Contact support to verify your API key is active

Operations Timing Out

Symptom: Operations fail with timeout errorsSolutions:

Baking operations are long-running (30-60 minutes) - this is normal
Use status checks instead of waiting for completion
Check your network connectivity
Ensure the MCP server URL is correct: https://mcp.bread.com.ai

AI Can't Use MCP Tools

Symptom: AI says it doesn’t have access to Bread toolsSolutions:

Explicitly mention “Use the Bread MCP server” in your request
Check that the MCP server appears in the AI’s tool list
Restart Claude Code or Cursor to reload MCP configurations
Verify the MCP configuration is saved correctly

Advantages of Using MCP

Natural Language Interface

Describe what you want in plain English instead of writing code or running commands

AI-Assisted Configuration

The AI can suggest optimal settings, catch errors, and iterate on your behalf

Faster Iteration

Skip the learning curve - the AI knows the Bread API and best practices

Automatic Status Monitoring

The AI can check operation status and notify you when bakes complete

When to Use MCP vs bgit

	MCP Server	bgit
Best For	Quick experiments, exploring ideas, getting AI suggestions	Production workflows, version control, team collaboration
Interface	Natural language conversation with AI	Git-native commands and YAML files
Version Control	Not built-in (you can ask AI to use git)	Native git integration with full history

Use both! Use MCP for quick experiments and exploration, then transition to bgit when you want proper version control and reproducibility.

Next Steps

Quickstart with bgit

Learn the git-native workflow for production use

Understanding Baking

Learn the theory behind prompt baking

API Reference

Explore the full Bread SDK API

bgit Workflows

Master advanced baking patterns

Usage

​What is the Bread MCP Server?

​How It Works

​Installation

​Prerequisites

​Install Claude Code

​Add Bread MCP Server

​Verify Installation

​Start Using

​Prerequisites

​Configure MCP Server

​Start Using

​Usage Examples

​Creating a Repository

​Setting Up Prompts

​Running a Complete Bake

​Checking Status

​Iterating on Models

​Available Operations

​Best Practices

Be Specific

Check Status

Iterate Gradually

Review Configurations

​Troubleshooting

​Advantages of Using MCP

Natural Language Interface

AI-Assisted Configuration

Faster Iteration

Automatic Status Monitoring

​When to Use MCP vs bgit

​Next Steps

Quickstart with bgit

Understanding Baking

API Reference

bgit Workflows

What is the Bread MCP Server?

How It Works

Installation

Prerequisites

Install Claude Code

Add Bread MCP Server

Verify Installation

Start Using

Prerequisites

Configure MCP Server

Start Using

Usage Examples

Creating a Repository

Setting Up Prompts

Running a Complete Bake

Checking Status

Iterating on Models

Available Operations

Best Practices

Troubleshooting

Advantages of Using MCP

When to Use MCP vs bgit

Next Steps