Scaffold autonomous AI agent projects pre-wired with safe, read-only bake MCP tooling.

AI Agents (bake agent init)

Overview

Scaffold a complete AI agent project with pre-configured Model Context Protocol (MCP) integrations and secure defaults.

bake agent init creates a production-ready starter workspace designed for AI coding agents (such as Cursor, Claude Desktop, or custom agent runtimes) that interact with Cookie Chain and Solana programs via bake.

Prerequisites

Node.js ≥ 18 (Node.js ≥ 22 recommended if using optional cookie-mcp tools)
bakeacookie installed globally:
bash
```
npm install -g bakeacookie
```

Quick start

bash

# 1. Scaffold your agent project
bake agent init my-agent

# 2. Enter project directory
cd my-agent

# 3. Install dependencies
npm install

Generated project structure

bake agent init generates a structured repository with safety guards, system prompts, and client connection configs:

text

my-agent/
├── README.md               # Project guide and agent workflow instructions
├── package.json            # Dependencies and helper scripts
├── policy.example.json     # Pre-configured bake MCP policy template (read-only by default)
├── prompts/
│   └── system.md           # System prompt containing Cookie Chain & toolchain constraints
└── mcp/
    ├── mcp.json            # Ready-to-use client config for Cursor / Claude Desktop
    └── README.md           # Instructions for connecting AI agent interfaces

Key files explained

File	Purpose
`prompts/system.md`	Injects crucial environment constraints (e.g. read-only defaults, Recipe Book architecture, Windows/WSL requirements) into your agent's context.
`policy.example.json`	Sample policy restricting or granting MCP tool access. Default: `allowWrites: false`.
`mcp/mcp.json`	JSON snippet you can paste directly into Claude Desktop or Cursor MCP settings.

Safe by default (`allowWrites: false`)

Warning: Write operations (deploying programs, rolling back on-chain state) can consume funds and alter contract code. By default, all generated policy configurations set "allowWrites": false.

In read-only mode, the agent has access to query and inspection tools only:

bake_whoami — inspect active wallet and cluster
bake_stats — query program invocation statistics
bake_prove — check on-chain bytecode against Recipe Book hashes
bake_logs — stream or inspect transaction logs
bake_get_history — list on-chain Recipe Book entries

Write tools (bake_deploy, bake_rollback, bake_confirm_action) are completely unregistered from the MCP server until an explicit policy enables them.

Connecting AI clients (Cursor & Claude Desktop)

The generated mcp/mcp.json file provides ready-to-copy configuration for your preferred AI workspace.

Claude Desktop configuration

Add this block to your claude_desktop_config.json:

json

{
  "mcpServers": {
    "bake": {
      "command": "bake",
      "args": ["mcp", "--policy", "./policy.example.json"]
    }
  }
}

Cursor configuration

In Cursor Settings → Features → MCP:

Name: bake
Type: command
Command: bake mcp --policy ./policy.example.json

Note on working directory & policy paths: Always start your agent or configure the policy path relative to your project root. If you move directories or run from a global context, pass an absolute path to the policy file (e.g. --policy /absolute/path/to/my-agent/policy.example.json).

Optional `cookie-mcp` integration

The scaffolded project is pre-configured to optionally integrate with @cookiechain/cookie-mcp for read-only market data and token liquidity lookups.

Security rule: Never paste private keys, seed phrases, or sensitive wallet secrets into chat prompts or agent system prompts. All transaction signing must go through local keypairs managed by bake login or explicit confirmation tokens.

Enabling writes & human confirmation

To grant your agent permission to deploy or rollback programs:

Copy policy.example.json to .bake/mcp-policy.json (or your chosen policy path) and set: