AB Tasty MCP Server

A Model Context Protocol (MCP) server that integrates AB Tasty's Feature Experimentation & Rollouts platform with AI assistants. This server provides tools for feature flag management, campaign retrieval, and guided SDK integration through the MCP protocol.

Features

Tools

• Decision API Integration: Retrieve campaigns and feature flags for visitors

  • decision_api_get_campaigns: Get all campaigns for a visitor with context
  • decision_api_get_campaign: Get a specific campaign by ID for a visitor
  • decision_api_get_flags: Get all feature flags for a visitor
  • decision_api_activate_campaign: Activate a campaign for a visitor

• Resource Loader API: Extract and analyze campaign configurations (v0.2.0+)

  • resource_loader_api_load_webexp_resources: Load Web Experimentation & Personalization resources (use with we_resource_extractor prompt)
  • resource_loader_api_load_featexp_resources: Load Feature Experimentation & Rollout resources (use with fear_resource_extractor prompt)
  • Note: In v0.2.0, the Resource Loader API was split into platform-specific tools for better clarity

Prompts

• Quick Start Guides: Interactive prompts for AB Tasty SDK installation
  • Node.js SDK installation guide
• Resource Extractors: AI-powered resource configuration generators
  • we_resource_extractor: Converts natural language campaign briefs into Web Experimentation resource configurations
  • fear_resource_extractor: Converts natural language campaign briefs into Feature Experimentation resource configurations
  • These prompts generate the JSON needed for the Resource Loader API tools

Resources

• Documentation: Access to AB Tasty Feature Experimentation & Rollout documentation directly through MCP

Prerequisites

• Node.js 18+
• npm or yarn
• AB Tasty credentials:
  • Environment ID
  • API Key
  • (Optional) Account ID and token for Resource Loader features for Web Experimentation

Installation

1. Clone the repository:

git clone https://github.com/flagship-io/mcp-server.git
cd mcp-server

1. Install dependencies:

yarn install

1. Build the project:

yarn build

Usage

Using with npx (Recommended)

The easiest way to use the MCP server is through npx without any installation:

{
  "mcpServers": {
    "ABTasty": {
      "command": "npx",
      "args": [
        "-y",
        "@abtasty/mcp-server",
        "--fear-env-id",
        "YOUR_FEAR_ENV_ID",
        "--fear-api-key",
        "YOUR_FEAR_API_KEY",
        "--resource-loader-fear-account-id",
        "YOUR_FEAR_ACCOUNT_ID",
        "--resource-loader-fear-account-environment-id",
        "YOUR_FEAR_ACCOUNT_ENVIRONMENT_ID",
        "--resource-loader-fear-rca-token",
        "YOUR_FEAR_RCA_TOKEN",
        "--resource-loader-we-account-id",
        "YOUR_WE_ACCOUNT_ID",
        "--resource-loader-we-token",
        "YOUR_WE_TOKEN"
      ]
    }
  }
}

Replace the placeholder values with your actual credentials:

• YOUR_FEAR_ENV_ID: Your AB Tasty Feature Experimentation account environment ID
• YOUR_FEAR_API_KEY: Your AB Tasty Feature Experimentation API key
• YOUR_FEAR_ACCOUNT_ID: Your AB Tasty Feature Experimentation Account ID
• YOUR_FEAR_ACCOUNT_ENVIRONMENT_ID: Your AB Tasty Feature Experimentation account Environment ID
• YOUR_FEAR_RCA_TOKEN: Your AB Tasty Feature Experimentation RCA Token
• YOUR_WE_ACCOUNT_ID: Your AB Tasty Web Experimentation Account ID
• YOUR_WE_TOKEN: Your AB Tasty Web Experimentation token

Alternative: Environment Variables

You can also set credentials via environment variables instead of command-line arguments:

# Feature Experimentation & Rollout
export FEAR_ENV_ID="your_fear_env_id"
export FEAR_API_KEY="your_fear_api_key"

# Resource Loader - Feature Experimentation & Rollout
export RESOURCE_LOADER_FEAR_ACCOUNT_ID="your_fear_account_id"
export RESOURCE_LOADER_FEAR_ACCOUNT_ENVIRONMENT_ID="your_fear_account_environment_id"
export RESOURCE_LOADER_FEAR_RCA_TOKEN="your_fear_rca_token"

# Resource Loader - Web Experimentation & Personalization
export RESOURCE_LOADER_WE_ACCOUNT_ID="your_we_account_id"
export RESOURCE_LOADER_WE_TOKEN="your_we_token"

Then use a simpler npx command:

{
  "mcpServers": {
    "ABTasty": {
      "command": "npx",
      "args": ["-y", "@abtasty/mcp-server"]
    }
  }
}

Running the HTTP Server

Start the server on the default port (3000):

yarn start

Or specify a custom port:

PORT=8080 yarn start

The server will be available at http://localhost:3000/mcp (or your custom port).

Configuration

The server accepts configuration through HTTP headers for each session:

Feature Experimentation & Rollout (FEAR) Headers:

• x-fear-env-id: Your AB Tasty Feature Experimentation & Rollout environment ID
• x-fear-api-key: Your AB Tasty Feature Experimentation & Rollout API key

Resource Loader Headers (optional):

For Web Experimentation & Personalization:

• x-resource-loader-we-account-id: Your AB Tasty Web Experimentation account ID
• x-resource-loader-we-token: Your Web Experimentation resource loader authentication token

For Feature Experimentation & Rollout:

• x-resource-loader-fear-account-id: Your AB Tasty Feature Experimentation & Rollout account ID
• x-resource-loader-fear-account-environment-id: Your AB Tasty Feature Experimentation & Rollout environment ID
• x-resource-loader-fear-rca-token: Your Feature Experimentation & Rollout (Remote Control API) resource loader authentication token

Connecting with MCP Clients

The server supports two modes of operation:

1. Stdio Mode (for Claude Desktop and similar clients)

Use the npx command shown above for stdio-based MCP clients like Claude Desktop.

2. HTTP Mode (Streamable HTTP transport)

For HTTP-based clients, the server uses the Streamable HTTP transport protocol. Configure your MCP client to connect to the server endpoint:

{
  "mcpServers": {
    "ABTasty": {
      "url": "http://localhost:3000/mcp",
      "headers": {
        "x-fear-env-id": "your_env_id",
        "x-fear-api-key": "your_api_key"
      }
    }
  }
}

Example: Using Tools

Once connected through an MCP client, you can use the available tools:

Get campaigns for a visitor:

decision_api_get_campaigns({
  visitor_id: "user123",
  context: {
    age: 25,
    country: "US"
  },
  trigger_hit: true
});

Get feature flags:

decision_api_get_flags({
  visitor_id: "user123",
  context: {
    age: 25,
    country: "US"
  },
  trigger_hit: false
});

Create and load Web Experimentation resources:

Step 1: Use the we_resource_extractor prompt to convert your campaign brief into a resource configuration:

User: Create an A/B test on https://example.com/pricing.
Show a red button for 50% of traffic.

Prompt Output:
{
  "needs_clarification": false,
  "questions": [],
  "resources": [
    {
      "type": "campaign",
      "$_ref": "c1",
      "action": "create",
      "payload": {...}
    }
  ]
}

Step 2: Load the resources using the tool:

resource_loader_api_load_webexp_resources({
  resourceLoaderContent: {
    "needs_clarification": false,
    "questions": [],
    "resources": [...]
  },
  dryrun: false,
});

Create and load Feature Experimentation resources:

Step 1: Use the fear_resource_extractor prompt to convert your campaign brief:

User: Create a feature flag 'new-checkout' for 25% of users

Prompt Output:
{
  "version": 1,
  "needs_clarification": false,
  "questions": [],
  "resources": [
    {
      "type": "project",
      "$_ref": "p1",
      "action": "create",
      "payload": {...}
    },
    {
      "type": "campaign",
      "$_ref": "c1",
      "action": "create",
      "payload": {...}
    }
  ]
}

Step 2: Load the resources using the tool:

resource_loader_api_load_featexp_resources({
  resourceLoaderContent: {
    "version": 1,
    "needs_clarification": false,
    "questions": [],
    "resources": [...]
  },
  dryrun: false,
});

Development

Project Structure

├── src/
│   ├── index.ts              # Main HTTP server entry point
│   ├── cli.ts                # CLI entry point for stdio mode
│   ├── tools/                # MCP tool implementations
│   │   ├── decision-api.ts   # Feature flag tools
│   │   └── resource-loader-api.ts
│   ├── prompts/              # Interactive prompt definitions
│   │   ├── quickstart-guide.ts
│   │   └── resource-loader-prompts.ts
│   └── resources/            # Resource providers
│       └── documentation.ts
├── helpers/                  # Utility functions
│   ├── abtasty-fear.ts      # AB Tasty Feature Experimentation SDK wrapper
│   └── resource-loader.ts   # Campaign extraction utilities
├── types/                   # TypeScript type definitions
├── assistant-prompts/       # Markdown prompt templates
└── build/                   # Compiled output

Development Scripts

# Generate prompt files
yarn generate:prompts

# Build the project
yarn build

# Build and run
yarn dev

Building from Source

The build process:

1. Generates TypeScript files from markdown prompts
2. Bundles the application with esbuild
3. Copies assistant prompt files to the build directory

yarn build

Architecture

Session Management

The server maintains isolated sessions for each client connection:

• Each session has its own credentials (provided via headers)
• Sessions are identified by unique UUIDs
• Sessions are cleaned up when connections close

Transport Layer

Uses the Streamable HTTP transport from the MCP SDK:

• Supports JSON-RPC 2.0 over HTTP
• Session-based connection management
• Automatic session initialization and cleanup

Security Notes

• Credentials are session-scoped and not shared between sessions
• Always use HTTPS in production environments
• Store credentials securely and never commit them to version control

API Documentation

Tools

decision_api_get_campaigns

Retrieves all campaigns for a visitor.

Parameters:

• visitor_id (string): Unique visitor identifier
• context (object): Key-value pairs for targeting (optional)
• trigger_hit (boolean): Send analytics hit (default: false)

Returns: Array of campaigns with variations and modifications

decision_api_get_campaign

Retrieves a specific campaign by its ID for a visitor.

Parameters:

• visitor_id (string): Unique visitor identifier
• campaign_id (string): Campaign ID to retrieve
• context (object): Key-value pairs for targeting (optional)
• trigger_hit (boolean): Send analytics hit (default: false)

Returns: Campaign with variation and modifications

decision_api_get_flags

Retrieves all feature flags for a visitor.

Parameters:

• visitor_id (string): Unique visitor identifier
• context (object): Key-value pairs for targeting (optional)
• trigger_hit (boolean): Send analytics hit (default: false)

Returns: Object with flag keys and their values

decision_api_activate_campaign

Activates a campaign for a visitor.

Parameters:

• visitor_id (string): Unique visitor identifier
• variation_group_id (string): Variation group ID to activate
• variation_id (string): Variation ID to activate

Returns: Activation confirmation

resource_loader_api_load_webexp_resources

Loads Web Experimentation & Personalization resources via the Resource Loader API.

Workflow: Use the we_resource_extractor prompt to generate the resourceLoaderContent from a natural language campaign brief, then pass it to this tool.

Parameters:

• resourceLoaderContent (object): JSON containing resource loader content with needs_clarification, questions, and resources array
• dryrun (boolean): Whether to simulate the request without sending it

Returns: Loaded resources results

resource_loader_api_load_featexp_resources

Loads Feature Experimentation & Rollout resources via the Resource Loader API.

Workflow: Use the fear_resource_extractor prompt to generate the resourceLoaderContent from a natural language campaign brief, then pass it to this tool.

Parameters:

• resourceLoaderContent (object): JSON containing resource loader content with version, needs_clarification, questions, and resources array
• dryrun (boolean): Whether to simulate the request without sending it

Returns: Loaded resources results

Troubleshooting

Server won't start

• Check if port 3000 is already in use
• Verify Node.js version (18+ required)
• Ensure dependencies are installed: yarn install

"Invalid credentials" errors

• Verify your environment ID and API key are correct
• Check that headers are properly set in your MCP client configuration
• Ensure credentials have the necessary permissions in AB Tasty

License

MIT License

Support

For issues related to:

• This MCP server: Open an issue on GitHub
• AB Tasty platform: Contact AB Tasty support

Resources

• MCP protocol: See Model Context Protocol documentation
• AB Tasty SDK: See AB Tasty SDK