AI Code Toolkit

Enforce AI coding agents your team's conventions and existing practices. Setup once, work across multiple AI tools.

Why This Exists

As projects grow from MVP to production, you accumulate patterns, conventions, components, and style guides. Your AGENTS.md, CLAUDE.md and rule files keep growing — consuming precious context window and documentation maintenance becomes difficult.

This toolkit encodes your team's conventions in a centralized, shareable location with hierarchical inheritant. Instead of preloading AI agents with plain-text documentation, you encode your best practices and guideline in yaml, and our tools with extract the configs to get relevant patterns and enforce boundaries before and after AI agent writing code in progressive discovery manner.

If you use multiple AI tools to assist your development, this toolkit enable your team to encode your knowledge and workflow once; and reuse it across AI tools seamlessly.

Quick Start

Requirements: Node.js >= 18, MCP-compatible agent (Claude Code, Cursor, Gemini CLI)

1. Initialize

Use aicode-toolkit to help you setup the project quickly. This includes download templates, setup MCPs and spec tool to assist with your development.

Creates templates/ with scaffold definitions, patterns, and rules. Your development knowledge stayed within templates folder, and is linked to actual project via project.json's sourceTemplate setting.

2. Configure MCP

We recomment to use MCP for plug-and-play capabilities. For folks who don't liked context hog problem with MCP, our libraries also have cli commands equivalent to the MCP's tools; or you can use our one-mcp to support progressive discovery.

The init command configures MCP automatically. For manual setup:

Claude Code (.mcp.json):

Cursor: Same config in .cursor/mcp.json

Flags:

• --admin-enable - Allow template creation

• --design-pattern-tool claude-code - AI-powered pattern analysis

• --review-tool claude-code - AI-powered code review

3. Verify

When the MCPs are setup, you can interact with the agent using natural language.

Ask your agent: "What boilerplates are available?"

Should call list-boilerplates. If not recognized, restart the agent.

Architecture

We recommend to use scaffold-mcp, architect-mcp and style-system together for full-stack development. This will help you creating new project quickly, add new feature at the right place, write code which follow your conventions and design guideline.

scaffold-mcp

Add new apps, libraries, or features that follow your company conventions. Generates minimal boilerplate code and uses guided generation to fill in the blanks.

Admin tools (--admin-enable):

architect-mcp

Pre-flight suggestions to ensure AI-generated code follows best practices and design patterns based on your file structure. Post-check with RULES.yaml to enforce styles and patterns using an LLM as a judge.

Admin tools (--admin-enable):

style-system (NEW)

Design system operations for theme management, CSS class discovery, and component visualization. Helps AI agents use existing design tokens and components instead of creating duplicates.

NOTE: This package hasn't been integrated to aicode-toolkit yet.

Workflow

We suggest to plan your task in advanced. Then simply give the task to agent to run autonomously. In copilot mode, you can prompt the agent using natural language or troubleshoot with the packages' cli commands.

Creating Projects

Before Editing Files

Adding Features

After Editing

Styling Components

Template Structure

scaffold.yaml

This is where you configure the template for scaffolding. We use structured output (supported by mcp) to generate minimal files with Comment header to guide the AI tool to fill-in the blank.

architect.yaml

Before the AI tool actually write the code, architect.yaml is where you steer the AI by giving it clues how to actually write the code. Don't wait until the code violates your conventions, steer it first.

RULES.yaml

There is no guarantee the AI follow your guidance. This file defines enforceable rules that the AI must, must not or should follow. The toolkit provide capability for a different AI agent to do code review and enforce the rules and not being affected by noisy context.

First, we extract the rules based on the file pattern; then we give the file diff plus rules to another agent to identify violation. Anything in must_do or must_not_do rules violations we explicitly ask the agent to fix (and other agent also provide fixing recommendation). You can inherit global RULES.yaml file in templates/RULES.yaml for maximum reusability.

Built-in Templates

Below templates are examples; you can clone the repo and start exploring how the mechanism works using existing templates.

Custom Templates

We suggest to build your own template from your existing repo. It's quite simple by using slash command:

Use this slash command and reference your directory to create template. This will create scaffold.yaml with boillerplate config and relevant files extracted from your production application.

After boilerplate is generated, you can now use this command to add feature scaffolding. Think of feature as a group of files that generated together per your requirement (new page, new service, etc...)

The scaffold-mcp will automatically add this new template to the discovery.

NOTE: MPC's prompts are added as commands in Claude Code; other tools might not have the same implementation. We plan to fix it with aicode-toolkit soon.

Custom Design Pattern

add-design-pattern is the tool from architect-mcp (with --admin-enable flag) that help you add a new design pattern to architect.yaml the template. Simply ask the AI agent to add a design pattern to a template by giving it a source file reference.

Custom Rule

add-rule is the tool from architect-mcp (with --admin-enable flag) that help you add a new rule to the RULES.yaml in template. Simply ask the AI agent to add a new rule to a template by giving it a source file reference and your rule requirement.

Project Types

The toolkit exists because we had scaling problem with mono-repo. Mono-repo has first-citizen support in this toolkit. Monolith is also support and we plan to make it more robust on single-purpose project ASAP!

Monorepo

Mono-repo can be complex for root level blob matching. Considering you can have multiple apis built with different languages, or same language but support different design patterns. Then you need to create duplicated rule files just to match a file within particular project. Eg:

Backend lib utils.

Frontend lib utils.

A slightly diverted pattern requires your team to write a different rule file. With our toolkit, your package/project reference template in project.json; so the match can be collocated per project. (We will support architect and rules override per project soon if you have edge cases).

Monolith

Single purpose project follow same template approach but simplier. Configuration in toolkit.yaml is enough for most use cases.

Auto-detected based on config files.

Token Optimization

For daily development work, the MCP context hog lies in the json-schema input definition. We create one-mcp to help you save context token by progressively disclose tools' schema per request.

If you want to use Anthropic skills for agent to automatically invoke commands for you on tools which does not support skills. One-mcp also support that. By the end of the day, skill is just a tool call.

See @agiflowai/one-mcp.

Supported Agents

Packages

Contributing

See CONTRIBUTING.md.

License

AGPL-3.0

Issues · Discord · Website