AI_README MCP Server
@Draco-Cheng
AI_README MCP Server について
AI_README MCP Server is a Model Context Protocol (MCP) server that helps AI assistants understand your project conventions through dedicated AI_README.md guide files. It automatically discovers, routes, and manages these files so AI can generate consistent, high-quality code that
基本情報
設定
以下の設定を使って、このサーバーを MCP 対応クライアントに追加してください。
{
"mcpServers": {
"ai-readme-manager": {
"command": "npx",
"args": [
"-y",
"ai-readme-mcp@latest"
]
}
}
}ツール
6Scan the project and discover all AI_README.md files. Returns an index of all README files with their paths, scopes, and coverage patterns.
ALWAYS CALL THIS FIRST - NO EXCEPTIONS. WHEN TO CALL (check these FIRST before responding): - User asks: optimize, improve, analyze, review, suggest, think, opinion. - User mentions: file name, component name, project, codebase, backend, frontend, infra, devops, CI/CD, docker, config. - User asks: "what can", "what should", "how to", "can you", "help me". - User wants to: create, build, add, implement, setup, configure, deploy, migrate, optimize, refactor, improve, analyze, review, suggest, think, opinion. - User selects an option that leads to code changes (e.g., choosing from a list of approaches). - User says: "next step", "next phase", "next section", "continue", "proceed", "go ahead", "start", "do it", "yes" (confirming action). - Before: Write, Edit, TodoWrite, Bash (for build/deploy), suggesting ANY code changes. - When executing a plan, todo, or multi-phase task: call BEFORE each file modification. - After user approves a plan or todo, call this BEFORE the each Write/Edit. IF ANY ABOVE = TRUE → CALL get_context_for_file IMMEDIATELY. MANDATORY: Call this tool BEFORE every code-related response. CALL EVERY TIME - EVERY response that involves code/suggestions MUST start with get_context_for_file. CALL EVERY TIME - even if you called it earlier in this conversation. CALL EVERY TIME - even if working on the same file. CALL EVERY TIME - even within the same section/phase - multiple calls are expected and correct. CALL EVERY TIME - context may have changed or sub-directories may have additional conventions. CALL EVERY TIME - when user confirms or selects an option that will result in code changes. What this tool does: Gets AI_README context showing project conventions that MUST be followed. Why it matters: - Without: Wrong tech (suggest Tailwind when project uses CSS Modules). - With: Suggestions fit existing codebase perfectly. - During plan/todo/phase execution: Context may differ per directory; call before EACH file change. DETECT CONFLICTS: After reading context, if user wants X but AI_README says Y: - This is ARCHITECTURAL DECISION. - Workflow: get_context → update_ai_readme → get_context → Write/Edit. RECORD DECISIONS: When you make architectural decisions during planning or implementation: - Design patterns, API structure, naming conventions, new abstractions. - Call update_ai_readme to record decisions that affect multiple files. - Future code (yours or others) will follow these recorded conventions.
CALL THIS to record DECISIONS and CONVENTIONS. WHEN TO CALL: A. CONFLICT RESOLUTION — STOP IMMEDIATELY when any of these occur: - User says: "don't use X", "use Y instead", "prefer", "switch to". - During planning: user's request or your proposal differs from AI_README conventions. - During planning: user approves a plan that contradicts AI_README. - User overrides a convention mid-task (even casually, e.g. 'just use X here'). - DO NOT continue planning or coding. Call update_ai_readme first, then resume. B. ARCHITECTURAL DECISIONS (during planning/implementation): - You chose a design pattern (e.g., repository pattern, factory, singleton). - You decided on API structure (REST paths, error format, response shape). - You established naming conventions (files, functions, variables). - You created new abstractions (utilities, hooks, services, types). - You set up error handling strategy or validation approach. - You introduced a new dependency or integration pattern. C. IMPLEMENTATION PATTERNS (after writing code): - You created a reusable pattern others should follow. - You established a file/folder structure for a new feature. - You made decisions that affect future development. D. MISSING / UNDOCUMENTED (during get_context or code review): - AI_README is missing a convention that is ALREADY USED in 2+ existing files. - A pattern exists in code but not in AI_README — record it so future code follows it. - Do NOT record one-off choices or speculative future patterns. RULE: If a decision will affect MORE THAN ONE FILE or FUTURE CODE → RECORD IT. WORKFLOW: 1. get_context (read current conventions). 2. Make decision or detect conflict. 3. update_ai_readme (record the decision). 4. Continue with implementation. Content Rules: - Extremely concise (< 400 tokens). - Only actionable conventions (tech, naming, patterns, infrastructure patterns, testing patterns). - NO explanations or examples
Validate all AI_README.md files in a project. Checks token count, structure, and content quality. Returns validation results with suggestions for improvement.
Compress an AI_README.md file using deterministic filler-language removal (no LLM call). WHEN TO CALL: - validate_ai_readmes reports 'filler-language' warnings. - validate_ai_readmes reports token count is too high. - After init_ai_readme, to tighten up generated content. - Any time you want to reduce AI_README token footprint without losing information. WHAT IT DOES (pure text transforms, deterministic): - Removes filler: just, really, basically, actually, simply, essentially - Shortens verbose phrases: 'in order to' → 'to', 'utilize' → 'use', 'make sure to' → 'ensure' - Removes hedging: 'you should', 'remember to', 'it might be worth', 'please note that' - Removes fluff connectives: furthermore, additionally, in addition, moreover - NEVER modifies: code blocks (``` fenced), inline code (`...`), headings, file paths, URLs, commands - Output may contain sentence fragments — this is intentional. Fragments are valid token-efficient format. USE dryRun:true FIRST to preview changes before writing.
Initialize and populate empty AI_README files within a project. When to use: - First-time setup when no AI_README exists. - get_context_for_file reports empty or missing AI_README files. - Newly created directories need conventions recorded. - Multiple directories require conventions in one pass. What it does: - Scans for missing or empty AI_README documents. - Creates a root-level AI_README if none is present. - Provides directory-specific prompts to gather conventions. - Guides you through documenting tech stack, patterns, and naming. Workflow: - Call init_ai_readme. - Follow the step-by-step instructions to inspect each directory. - Use update_ai_readme to record the conventions. - Run validate_ai_readmes to check for problems. - Fix any warnings (remove redundant content, add Cross-directory dependencies section). - Re-run get_context_for_file to confirm coverage before coding.
概要
What is AI_README MCP Server?
The AI_README MCP Server is a Model Context Protocol (MCP) server that helps AI assistants understand and follow a project’s coding conventions through dedicated AI_README.md guide files. It works with GitHub Copilot, Claude Code, Cursor, OpenClaw, and other MCP-compatible AI tools.
How to use AI
「メモリとナレッジ」の他のコンテンツ
Docs MCP Server
araboldGrounded Docs MCP Server: Open-Source Alternative to Context7, Nia, and Ref.Tools
Mcp Knowledge Graph
shanehollomanMCP server enabling persistent memory for Claude through a local knowledge graph - fork focused on local development
Anytype MCP Server
anyprotoAn MCP server enabling AI assistants to interact with Anytype - your encrypted, local and collaborative wiki - to organize objects, lists, and more through natural language.
Memory Bank MCP Server
alioshrA Model Context Protocol (MCP) server implementation for remote memory bank management, inspired by Cline Memory Bank.
Semantic Scholar MCP Server
YUZongminA FastMCP server implementation for the Semantic Scholar API, providing comprehensive access to academic paper data, author information, and citation networks.
コメント