MCP Tool Forge

A Python CLI tool that extracts tools from MCP servers and converts them into 5 formats: boto3, AWS CLI, OpenAPI schemas, AgentCore Gateway configs, and Claude Code / Kiro-CLI Skills. Includes 9 ready-to-use AWS Skills for Claude Code and Kiro-CLI.

한국어 버전은 아래에 있습니다 / Korean version below

Why?

The MCP Token Cost Problem

MCP is powerful, but every tool definition and request/response consumes LLM tokens.

Token consumption per conversation:

┌─────────────┐     JSON Schema (tool defs)      ┌─────────┐
│  LLM Agent  │ ◄──────── ~2,000 tokens/tool ──── │   MCP   │
│             │ ──────── request (function call) ► │  Server │
│             │ ◄──────── response (result) ────── │         │
└─────────────┘                                   └─────────┘

Real-world measurement (Kiro-CLI):

Scenario	Credits	Token Usage
4 MCP servers ON	0.71	Schema loading + response
4 MCP servers OFF (Skills only)	0.27	Selective skill loading
Savings	62%

Loading all 792 tools from 67 AWS MCP servers would consume ~1.5M tokens just for tool definitions.

How MCP Tool Forge Solves This

Extract once → Convert to native code → Run without MCP

  MCP way:    Agent ←→ MCP Server ←→ AWS API  (tokens every time)
  Skill way:  Agent → boto3/CLI code → AWS API  (zero tokens, zero latency)

Comparison	MCP Approach	Skill Approach
Tool definition tokens	~2,000/tool (per conversation)	0 (embedded in code)
Request/response tokens	JSON-RPC roundtrip	0 (direct execution)
Server dependency	Always required	Not required
Latency	MCP server roundtrip	0
Offline support	No	Yes

Skills Installation (3 lines)

No MCP servers, Python packages, or Bedrock access required. Just AWS credentials and you're ready.

Claude Code

git clone https://github.com/whchoi98/mcp-tool-forge.git
mkdir -p .claude/skills
cp -r mcp-tool-forge/.claude/skills/aws-* .claude/skills/

Kiro-CLI

git clone https://github.com/whchoi98/mcp-tool-forge.git
mkdir -p .kiro/skills
cp -r mcp-tool-forge/.claude/skills/aws-* .kiro/skills/

Global Installation (all projects)

# Claude Code
mkdir -p ~/.claude/skills
cp -r mcp-tool-forge/.claude/skills/aws-* ~/.claude/skills/

# Kiro-CLI
mkdir -p ~/.kiro/skills
cp -r mcp-tool-forge/.claude/skills/aws-* ~/.kiro/skills/

Install Specific Skills

# Claude Code
cp -r mcp-tool-forge/.claude/skills/aws-iam .claude/skills/
cp -r mcp-tool-forge/.claude/skills/aws-cost .claude/skills/
cp -r mcp-tool-forge/.claude/skills/aws-network .claude/skills/

# Kiro-CLI
cp -r mcp-tool-forge/.claude/skills/aws-iam .kiro/skills/
cp -r mcp-tool-forge/.claude/skills/aws-network .kiro/skills/

Natural language triggers activate skills automatically:

Request	Skill
"List IAM users"	aws-iam
"Show this month's costs"	aws-cost
"Check CloudWatch alarms"	aws-cloudwatch
"List Lambda functions"	aws-infra
"List SQS queues"	aws-messaging
"Check VPC network"	aws-network
"List Bedrock models"	aws-ai
"Query DynamoDB tables"	aws-data
"Run security audit"	aws-security

9 AWS Skills

Skill	Trigger	Operations
aws-iam	"IAM users/roles/policies"	list_users, create_role, attach_policy, simulate_policy
aws-cloudwatch	"Logs/metrics/alarms/audit"	Logs Insights, get_metric_data, describe_alarms, CloudTrail
aws-cost	"Cost/billing/pricing"	get_cost_and_usage, cost_forecast, pricing lookup
aws-infra	"Resources/stacks/containers"	Cloud Control API, CloudFormation, EKS, ECS, Lambda
aws-messaging	"Queues/topics/messages"	SNS publish, SQS send/receive, MQ, Step Functions
aws-network	"VPC/subnet/TGW/VPN"	VPC, Transit Gateway, Cloud WAN, Network Firewall, VPN, Flow Logs
aws-ai	"Bedrock/SageMaker/Kendra"	Bedrock Converse, Knowledge Bases, Agents, SageMaker, Kendra, Q Business
aws-data	"DB/cache/query"	DynamoDB, Aurora, Redshift, ElastiCache, Neptune
aws-security	"Account/credentials/audit"	get_caller_identity, credential audit, MFA check

Skills Coverage

The 9 hand-crafted skills cover 52 of 67 MCP servers (78%).

Uncovered 15 servers:

Category	Uncovered Servers	Reason
Core / Essential	aws-api, core-mcp, aws-mcp	MCP proxy/planning — not a specific service
Documentation	aws-documentation, aws-knowledge	Doc search only — not a boto3/CLI target
Developer Tools	aws-diagram, aws-msk, code-doc-gen, frontend, git-repo-research, synthetic-data	Dev tools (diagrams, Kafka, code docs, etc.)
Healthcare	aws-healthomics, healthimaging, healthlake	Specialized healthcare services
Cost & Operations	aws-managed-prometheus	Prometheus monitoring

Uncovered servers can still be used through 792 auto-generated skills via mcp-tool-forge convert.

Skills Test Results

Tested on a real AWS account (Seoul region, IAM role auth):

#	Skill	Test	Result
1	aws-security	`sts get-caller-identity`	Pass
2	aws-iam	`list-users`	Pass (2 users)
3	aws-iam	`list-roles`	Pass (13 roles)
4	aws-cloudwatch	`describe-log-groups`	Pass
5	aws-cloudwatch	`describe-alarms`	Pass
6	aws-infra	Cloud Control EC2	Pass
7	aws-infra	Cloud Control S3	Pass (1 bucket)
8	aws-infra	CloudFormation stacks	Pass
9	aws-cost	`get-cost-and-usage`	Pass ($1,093)
10	aws-cost	Cost by service (top 5)	Pass
11	aws-security	MFA check (boto3)	Pass (2 NO MFA)
12	aws-security	Access key age (boto3)	Pass (23 days)
13	aws-security	Account summary (boto3)	Pass (Root MFA: NO)
14	aws-messaging	SNS topics	Pass
15	aws-messaging	SQS queues	Pass
16	aws-data	DynamoDB tables	Pass
17	aws-infra	Lambda functions	Pass

17/17 passed — All skills work with standard AWS credentials.

Token Economics: MCP vs Skills

Why Does MCP Consume So Many Tokens?

When MCP servers load, all tool JSON Schemas are injected into the LLM context. This is necessary for the LLM to "know" available tools, but most tools go unused.

Example: IAM MCP server loading

  29 tools x ~2,000 tokens/tool = ~58,000 tokens (definitions only)
  + request/response JSON-RPC roundtrip = ~500 tokens/call

  All 67 servers loaded:
  792 tools x ~2,000 tokens = ~1,584,000 tokens (~1.5M)

How Skills Save Tokens

Skills use selective loading — only the relevant skill is loaded per request.

Example: "List IAM users" request

  MCP way:   Load all 29 tool schemas (58,000 tokens) -> use 1
  Skill way: Load aws-iam SKILL.md (~3,000 tokens) -> execute directly

Item	MCP (1 server)	MCP (all 67)	Skill
Initial loading	~58,000 tokens	~1,584,000 tokens	~3,000 tokens
Per tool call	~500 tokens	~500 tokens	0 tokens
Server processes	1 required	67 required	0
Cold start	2-5s	30s+	0s

Cost Estimation (Claude Opus 4.6, $15/1M input tokens)

Scenario	Input Tokens	Cost
MCP 1 server/conversation	58,000	$0.87
MCP 10 servers/conversation	580,000	$8.70
Skill 1/conversation	3,000	$0.045
Skill all 9/conversation	27,000	$0.41

Skills reduce token costs by 95%+ compared to MCP.

Quick Start (CLI Tool)

If you only need Skills, skip this section. Skills Installation above is all you need.

# Install
pip install -e ".[dev]"

# List servers
mcp-tool-forge list-servers
mcp-tool-forge list-servers --category "Data & Analytics"

# Check tools (connects to real MCP server)
mcp-tool-forge list-tools --server aws-iam-mcp-server

# Convert to all formats
mcp-tool-forge convert --server aws-iam-mcp-server --output all

# LLM mapping (Bedrock Claude)
mcp-tool-forge convert --server amazon-cloudwatch-mcp-server --output all --llm-assist

# Multi-profile support (set default profile in generated code)
mcp-tool-forge convert --server aws-iam-mcp-server --output boto3 --aws-profile prod-account

# Register skills to Claude Code
mcp-tool-forge register --server aws-iam-mcp-server -d output

# Register skills to Kiro-CLI
mcp-tool-forge register --server aws-iam-mcp-server -d output --target kiro

Multi AWS Profile Support

Use --aws-profile to set a default AWS profile in generated boto3 code. Useful for AWS Organizations + SSO environments with multiple accounts.

# Generated without --aws-profile (default)
def list_users(profile_name: str | None = None, **kwargs) -> dict:
    session = boto3.Session(profile_name=profile_name)
    client = session.client('iam')
    ...

# Generated with --aws-profile prod-account
def list_users(profile_name: str | None = "prod-account", **kwargs) -> dict:
    session = boto3.Session(profile_name=profile_name)
    client = session.client('iam')
    ...

# Override profile at call time
list_users()                                  # Uses default profile
list_users(profile_name="staging-account")    # Switch to another account

Output Formats

Format	File	Purpose
boto3 (.py)	`output/*/boto3/tools.py`	Direct invocation from AgentCore Gateway Lambda
AWS CLI (.sh)	`output/*/cli/tools.sh`	Shell-based agents, automation scripts
Schema (.json)	`output/*/schema/tools.json`	OpenAPI-compatible tool definitions
AgentCore (.json)	`output/*/agentcore/tool_config.json`	Bedrock AgentCore Gateway toolSpec
Skill (.md)	`output//skill/.md`	Claude Code / Kiro-CLI skills

Architecture

┌──────────────────────────────────────────────────────┐
│                    mcp-tool-forge                     │
├──────────────────────────────────────────────────────┤
│                                                      │
│  Phase 1: Schema Extraction                          │
│  ┌───────────┐  tools/list  ┌────────────────┐      │
│  │ MCP Server ├────────────►│ Schema Cache    │      │
│  │ (stdio)    │  MCP SDK    │(~/.mcp-tool-forge)│    │
│  └───────────┘              └───────┬────────┘      │
│                                      │               │
│  Phase 2: Static Mapping             │               │
│  ┌──────────────┐                    │               │
│  │ YAML Mappings │◄──────────────────┘               │
│  │ (iam, dynamo) │                                   │
│  └──────┬───────┘                                   │
│         │ unmapped                                   │
│  Phase 3: LLM Inference                              │
│  ┌──────▼───────┐                                   │
│  │ Bedrock Claude│  tool schema → boto3 mapping      │
│  │ (Opus 4.6)    │                                   │
│  └──────┬───────┘                                   │
│         │                                            │
│  Code Generation                                     │
│  ┌──────▼───────┐                                   │
│  │  Generators   │  boto3 / cli / schema /           │
│  │  (Jinja2)     │  agentcore / skill                │
│  └──────────────┘                                   │
└──────────────────────────────────────────────────────┘

Extract — MCP SDK stdio_client connects to server, tools/list extracts schemas. Cached in ~/.mcp-tool-forge/cache/.
Static Map — Known mappings from mappings/*.yaml (IAM 29 + DynamoDB 6 = 35 tools).
LLM Map — Unmapped tools sent to Bedrock Claude Opus 4.6 for inference. --llm-assist flag.

Extraction Results

Metric	Value
Registered servers	67
Servers connected	55 / 67 (82%)
Tools extracted	792
boto3 functions	480+
Skills generated	792
Syntax pass rate	91.4% (after auto-fix)

Supported Servers (67)

Category	Count	Servers
Data & Analytics	18	DynamoDB, Aurora, Redshift, ElastiCache, Neptune
Infrastructure & Deployment	11	EKS, ECS, CDK, CloudFormation, Network, Terraform, Serverless
AI & Machine Learning	10	Bedrock, SageMaker, Kendra, Nova Canvas
Cost & Operations	8	CloudWatch, CloudTrail, Cost Explorer, Billing
Developer Tools & Support	7	IAM, MSK, Diagram, Code Doc Gen
Integration & Messaging	5	SNS/SQS, MQ, Step Functions, Location
Healthcare & Lifesciences	3	HealthOmics, HealthImaging, HealthLake
Core	2	AWS API, Core MCP
Documentation	2	AWS Documentation, Knowledge
Essential Setup	1	AWS MCP (unified proxy)

Project Structure

mcp-tool-forge/
├── .claude/skills/             # 9 AWS Skills (portable)
│   ├── aws-iam/                # IAM users, roles, policies
│   ├── aws-cloudwatch/         # Logs, metrics, alarms, CloudTrail
│   ├── aws-cost/               # Cost Explorer, billing, pricing
│   ├── aws-infra/              # CloudFormation, EKS, ECS, Lambda
│   ├── aws-messaging/          # SNS, SQS, MQ, Step Functions
│   ├── aws-network/            # VPC, Transit Gateway, Cloud WAN, VPN
│   ├── aws-ai/                 # Bedrock, SageMaker, Kendra, Q Business
│   ├── aws-data/               # DynamoDB, Aurora, Redshift, Neptune
│   └── aws-security/           # Account info, security audit
├── src/mcp_to_cli/
│   ├── cli.py                  # Click CLI entry point
│   ├── pipeline.py             # 3-phase orchestrator
│   ├── connector.py            # MCP SDK stdio_client
│   ├── registry.yaml           # 67 server configs
│   ├── llm_mapper.py           # Bedrock Claude Opus 4.6
│   ├── validator.py            # Generated code fixer
│   ├── generators/             # 5 output generators
│   ├── mappings/               # Static YAML (IAM, DynamoDB)
│   └── templates/              # 6 Jinja2 templates
├── tests/                      # 38 pytest tests
└── docs/                       # Architecture & plans

Development

pip install -e ".[dev]"      # Install
python -m pytest tests/ -v   # Test (38 tests)

Requirements

Item	Skills Only	Full CLI Tool
Python >= 3.11	Not required	Required
AWS credentials	Required	Required
uvx / npx	Not required	Required (MCP server execution)
Bedrock access	Not required	Optional (LLM mapping)

License

MIT

MCP Tool Forge (한국어)

MCP 서버의 도구를 추출하여 boto3, AWS CLI, OpenAPI 스키마, AgentCore Gateway, Claude Code / Kiro-CLI Skill 5가지 형식으로 변환하는 Python CLI 도구. 9개 AWS Skills을 포함하며 Claude Code와 Kiro-CLI 모두 지원합니다.

English version above / 영어 버전은 위에 있습니다

왜 필요한가?

MCP의 토큰 비용 문제

MCP 프로토콜은 강력하지만, 모든 도구 정의와 요청/응답이 LLM 토큰으로 소모됩니다.

매 대화마다 발생하는 토큰 소비:

┌─────────────┐     JSON Schema (도구 정의)      ┌─────────┐
│  LLM Agent  │ ◄──────── ~2,000 토큰/도구 ────── │   MCP   │
│             │ ──────── 요청 (함수 호출) ────────► │  Server │
│             │ ◄──────── 응답 (결과) ──────────── │         │
└─────────────┘                                   └─────────┘

실제 측정 결과 (Kiro-CLI 기준):

시나리오	Credits	토큰 사용량
MCP 서버 4개 ON	0.71	도구 스키마 로딩 + 응답
MCP 서버 4개 OFF (Skills만)	0.27	필요한 스킬만 선택 로딩
절감률	62%

67개 AWS MCP 서버에서 792개 도구를 전부 로딩하면, 도구 정의만으로 ~150만 토큰이 소모됩니다.

MCP Tool Forge의 해법

한 번 추출 → 네이티브 코드로 변환 → MCP 없이 직접 실행

  MCP 방식:   Agent ←→ MCP Server ←→ AWS API  (매번 토큰 소비)
  Skill 방식:  Agent → boto3/CLI 코드 → AWS API  (토큰 0, 지연 0)

비교	MCP 방식	Skill 방식
도구 정의 토큰	~2,000/도구 (매 대화)	0 (코드에 내장)
요청/응답 토큰	JSON-RPC 왕복	0 (직접 실행)
서버 의존성	항상 실행 필요	불필요
지연 시간	MCP 서버 왕복	0
오프라인 동작	불가	가능

Skills 설치 (3줄이면 끝)

MCP 서버 설치, Python 패키지, Bedrock 접근 모두 불필요. AWS 자격 증명만 있으면 바로 사용할 수 있습니다.

Claude Code

git clone https://github.com/whchoi98/mcp-tool-forge.git
mkdir -p .claude/skills
cp -r mcp-tool-forge/.claude/skills/aws-* .claude/skills/

Kiro-CLI

git clone https://github.com/whchoi98/mcp-tool-forge.git
mkdir -p .kiro/skills
cp -r mcp-tool-forge/.claude/skills/aws-* .kiro/skills/

글로벌 설치 (모든 프로젝트에 적용)

# Claude Code
mkdir -p ~/.claude/skills
cp -r mcp-tool-forge/.claude/skills/aws-* ~/.claude/skills/

# Kiro-CLI
mkdir -p ~/.kiro/skills
cp -r mcp-tool-forge/.claude/skills/aws-* ~/.kiro/skills/

특정 스킬만 설치

# Claude Code 예시
cp -r mcp-tool-forge/.claude/skills/aws-iam .claude/skills/
cp -r mcp-tool-forge/.claude/skills/aws-cost .claude/skills/
cp -r mcp-tool-forge/.claude/skills/aws-network .claude/skills/

# Kiro-CLI 예시
cp -r mcp-tool-forge/.claude/skills/aws-iam .kiro/skills/
cp -r mcp-tool-forge/.claude/skills/aws-network .kiro/skills/

설치 후 자연어로 요청하면 스킬이 자동 활성화됩니다:

요청 예시	활성화 스킬
"IAM 사용자 목록 보여줘"	aws-iam
"이번 달 비용은?"	aws-cost
"CloudWatch 알람 확인"	aws-cloudwatch
"Lambda 함수 목록"	aws-infra
"SQS 큐 목록"	aws-messaging
"VPC 네트워크 확인"	aws-network
"Bedrock 모델 목록"	aws-ai
"DynamoDB 테이블 조회"	aws-data
"보안 감사 실행"	aws-security

9개 AWS Skills

Skill	트리거	주요 작업
aws-iam	"IAM 사용자/역할/정책"	list_users, create_role, attach_policy, simulate_policy
aws-cloudwatch	"로그/메트릭/알람/감사"	Logs Insights, get_metric_data, describe_alarms, CloudTrail
aws-cost	"비용/청구/가격"	get_cost_and_usage, cost_forecast, pricing lookup
aws-infra	"리소스/스택/컨테이너"	Cloud Control API, CloudFormation, EKS, ECS, Lambda
aws-messaging	"큐/토픽/메시지/워크플로"	SNS publish, SQS send/receive, MQ, Step Functions
aws-network	"VPC/서브넷/TGW/VPN"	VPC, Transit Gateway, Cloud WAN, Network Firewall, VPN, Flow Logs
aws-ai	"Bedrock/SageMaker/Kendra"	Bedrock Converse, Knowledge Bases, Agents, SageMaker, Kendra, Q Business
aws-data	"DB/캐시/쿼리"	DynamoDB, Aurora, Redshift, ElastiCache, Neptune
aws-security	"계정/자격증명/보안감사"	get_caller_identity, credential audit, MFA check

Skills 커버리지

9개 수동 스킬은 67개 MCP 서버 중 **52개 (78%)**를 커버합니다.

미커버 15개 서버:

카테고리	미커버 서버	사유
Core / Essential	aws-api, core-mcp, aws-mcp	MCP 프록시/플래닝 — 특정 서비스가 아님
Documentation	aws-documentation, aws-knowledge	문서 검색 전용 — boto3/CLI 대상 아님
Developer Tools	aws-diagram, aws-msk, code-doc-gen, frontend, git-repo-research, synthetic-data	개발 도구 (다이어그램, Kafka, 코드 문서화 등)
Healthcare	aws-healthomics, healthimaging, healthlake	헬스케어 전문 서비스
Cost & Operations	aws-managed-prometheus	Prometheus 모니터링

미커버 서버도 mcp-tool-forge convert로 자동 생성된 792개 개별 스킬을 통해 사용 가능합니다.

Skills 테스트 결과

실제 AWS 계정에서 테스트 (서울 리전, IAM 역할 인증):

#	Skill	테스트	결과
1	aws-security	`sts get-caller-identity`	Pass
2	aws-iam	`list-users`	Pass (2 users)
3	aws-iam	`list-roles`	Pass (13 roles)
4	aws-cloudwatch	`describe-log-groups`	Pass
5	aws-cloudwatch	`describe-alarms`	Pass
6	aws-infra	Cloud Control EC2	Pass
7	aws-infra	Cloud Control S3	Pass (1 bucket)
8	aws-infra	CloudFormation stacks	Pass
9	aws-cost	`get-cost-and-usage`	Pass ($1,093)
10	aws-cost	Cost by service (top 5)	Pass
11	aws-security	MFA check (boto3)	Pass (2 NO MFA)
12	aws-security	Access key age (boto3)	Pass (23 days)
13	aws-security	Account summary (boto3)	Pass (Root MFA: NO)
14	aws-messaging	SNS topics	Pass
15	aws-messaging	SQS queues	Pass
16	aws-data	DynamoDB tables	Pass
17	aws-infra	Lambda functions	Pass

17/17 통과 — 모든 스킬이 표준 AWS 자격 증명으로 동작합니다.

토큰 경제학: MCP vs Skills

왜 MCP는 토큰을 많이 소모하는가?

MCP 서버가 로딩되면 모든 도구의 JSON Schema가 LLM 컨텍스트에 주입됩니다. 이는 LLM이 어떤 도구를 사용할 수 있는지 "알기 위해" 필요하지만, 대부분의 도구는 사용되지 않습니다.

예: IAM MCP 서버 로딩 시

  29개 도구 x ~2,000 토큰/도구 = ~58,000 토큰 (도구 정의만)
  + 요청/응답 JSON-RPC 왕복 = ~500 토큰/호출

  67개 서버 전체 로딩 시:
  792개 도구 x ~2,000 토큰 = ~1,584,000 토큰 (약 150만)

Skills는 어떻게 토큰을 절약하는가?

Skills는 필요한 스킬만 선택적으로 로딩합니다.

예: "IAM 사용자 목록 보여줘" 요청 시

  MCP 방식:  29개 도구 스키마 전체 로딩 (58,000 토큰) -> 1개 사용
  Skill 방식: aws-iam SKILL.md 1개 로딩 (~3,000 토큰) -> 바로 실행

항목	MCP (서버 1개)	MCP (67개 전체)	Skill
초기 로딩	~58,000 토큰	~1,584,000 토큰	~3,000 토큰
도구 호출	~500 토큰/회	~500 토큰/회	0 토큰
서버 프로세스	1개 필요	67개 필요	0개
콜드 스타트	2~5초	30초+	0초

비용 환산 (Claude Opus 4.6 기준)

시나리오	입력 토큰	비용 ($15/1M 토큰)
MCP 1개 서버/대화	58,000	$0.87/대화
MCP 10개 서버/대화	580,000	$8.70/대화
Skill 1개/대화	3,000	$0.045/대화
Skill 전체/대화	27,000	$0.41/대화

Skill 방식은 MCP 대비 토큰 비용을 95% 이상 절감합니다.

빠른 시작 (CLI 도구)

Skills만 사용하려면 이 섹션은 건너뛰어도 됩니다. 위의 Skills 설치만으로 충분합니다.

# 설치
pip install -e ".[dev]"

# 서버 목록 조회
mcp-tool-forge list-servers
mcp-tool-forge list-servers --category "Data & Analytics"

# 도구 확인 (실제 MCP 서버 연결)
mcp-tool-forge list-tools --server aws-iam-mcp-server

# 모든 형식으로 변환
mcp-tool-forge convert --server aws-iam-mcp-server --output all

# LLM 매핑 (Bedrock Claude)
mcp-tool-forge convert --server amazon-cloudwatch-mcp-server --output all --llm-assist

# 멀티 AWS 프로필 지원 (생성 코드에 기본 프로필 설정)
mcp-tool-forge convert --server aws-iam-mcp-server --output boto3 --aws-profile prod-account

# Claude Code에 스킬 등록
mcp-tool-forge register --server aws-iam-mcp-server -d output

# Kiro-CLI에 스킬 등록
mcp-tool-forge register --server aws-iam-mcp-server -d output --target kiro

멀티 AWS 프로필 지원

--aws-profile 옵션으로 생성되는 boto3 코드에 AWS 프로필을 설정할 수 있습니다. AWS Organizations + SSO 환경에서 여러 계정을 사용하는 경우 유용합니다.

# --aws-profile 없이 생성 (기본)
def list_users(profile_name: str | None = None, **kwargs) -> dict:
    session = boto3.Session(profile_name=profile_name)
    client = session.client('iam')
    ...

# --aws-profile prod-account 으로 생성
def list_users(profile_name: str | None = "prod-account", **kwargs) -> dict:
    session = boto3.Session(profile_name=profile_name)
    client = session.client('iam')
    ...

# 호출 시 프로필 오버라이드 가능
list_users()                                  # 기본 프로필 사용
list_users(profile_name="staging-account")    # 다른 계정으로 전환

출력 형식

형식	파일	용도
boto3 (.py)	`output/*/boto3/tools.py`	AgentCore Gateway Lambda에서 직접 호출
AWS CLI (.sh)	`output/*/cli/tools.sh`	Shell 기반 에이전트, 자동화 스크립트
Schema (.json)	`output/*/schema/tools.json`	OpenAPI 호환 도구 정의
AgentCore (.json)	`output/*/agentcore/tool_config.json`	Bedrock AgentCore Gateway toolSpec
Skill (.md)	`output//skill/.md`	Claude Code / Kiro-CLI 스킬

아키텍처

┌──────────────────────────────────────────────────────┐
│                    mcp-tool-forge                     │
├──────────────────────────────────────────────────────┤
│                                                      │
│  Phase 1: 스키마 추출                                 │
│  ┌───────────┐  tools/list  ┌────────────────┐      │
│  │ MCP Server ├────────────►│ Schema Cache    │      │
│  │ (stdio)    │  MCP SDK    │(~/.mcp-tool-forge)│    │
│  └───────────┘              └───────┬────────┘      │
│                                      │               │
│  Phase 2: 정적 매핑                    │               │
│  ┌──────────────┐                    │               │
│  │ YAML Mappings │◄──────────────────┘               │
│  │ (iam, dynamo) │                                   │
│  └──────┬───────┘                                   │
│         │ unmapped                                   │
│  Phase 3: LLM 추론                                   │
│  ┌──────▼───────┐                                   │
│  │ Bedrock Claude│  tool schema → boto3 mapping      │
│  │ (Opus 4.6)    │                                   │
│  └──────┬───────┘                                   │
│         │                                            │
│  코드 생성                                            │
│  ┌──────▼───────┐                                   │
│  │  Generators   │  boto3 / cli / schema /           │
│  │  (Jinja2)     │  agentcore / skill                │
│  └──────────────┘                                   │
└──────────────────────────────────────────────────────┘

추출 — MCP SDK stdio_client로 서버 연결, tools/list로 스키마 추출. ~/.mcp-tool-forge/cache/에 캐시.
정적 매핑 — mappings/*.yaml에서 알려진 매핑 조회 (IAM 29 + DynamoDB 6 = 35개).
LLM 매핑 — 미매핑 도구를 Bedrock Claude Opus 4.6에 전송. --llm-assist 플래그.

추출 결과

지표	값
등록 서버	67개
연결 성공	55 / 67 (82%)
도구 추출	792개
boto3 함수	480+
생성된 스킬	792개
구문 통과율	91.4% (자동 수정 후)

지원 서버 (67개)

카테고리	수	서버
Data & Analytics	18	DynamoDB, Aurora, Redshift, ElastiCache, Neptune
Infrastructure & Deployment	11	EKS, ECS, CDK, CloudFormation, Network, Terraform, Serverless
AI & Machine Learning	10	Bedrock, SageMaker, Kendra, Nova Canvas
Cost & Operations	8	CloudWatch, CloudTrail, Cost Explorer, Billing
Developer Tools & Support	7	IAM, MSK, Diagram, Code Doc Gen
Integration & Messaging	5	SNS/SQS, MQ, Step Functions, Location
Healthcare & Lifesciences	3	HealthOmics, HealthImaging, HealthLake
Core	2	AWS API, Core MCP
Documentation	2	AWS Documentation, Knowledge
Essential Setup	1	AWS MCP (unified proxy)

프로젝트 구조

mcp-tool-forge/
├── .claude/skills/             # 9개 AWS Skills (이식 가능)
│   ├── aws-iam/                # IAM 사용자, 역할, 정책
│   ├── aws-cloudwatch/         # 로그, 메트릭, 알람, CloudTrail
│   ├── aws-cost/               # 비용 탐색기, 청구, 가격
│   ├── aws-infra/              # CloudFormation, EKS, ECS, Lambda
│   ├── aws-messaging/          # SNS, SQS, MQ, Step Functions
│   ├── aws-network/            # VPC, Transit Gateway, Cloud WAN, VPN
│   ├── aws-ai/                 # Bedrock, SageMaker, Kendra, Q Business
│   ├── aws-data/               # DynamoDB, Aurora, Redshift, Neptune
│   └── aws-security/           # 계정 정보, 보안 감사
├── src/mcp_to_cli/
│   ├── cli.py                  # Click CLI 진입점
│   ├── pipeline.py             # 3단계 오케스트레이터
│   ├── connector.py            # MCP SDK stdio_client
│   ├── registry.yaml           # 67개 서버 설정
│   ├── llm_mapper.py           # Bedrock Claude Opus 4.6
│   ├── validator.py            # 생성 코드 검증/수정
│   ├── generators/             # 5가지 출력 생성기
│   ├── mappings/               # 정적 YAML (IAM, DynamoDB)
│   └── templates/              # 6개 Jinja2 템플릿
├── tests/                      # 38개 pytest 테스트
└── docs/                       # 아키텍처 및 설계 문서

개발

pip install -e ".[dev]"      # 설치
python -m pytest tests/ -v   # 테스트 (38개)

요구사항

항목	Skills만 사용	CLI 도구 전체
Python >= 3.11	불필요	필요
AWS 자격 증명	필요	필요
uvx / npx	불필요	필요 (MCP 서버 실행)
Bedrock 접근	불필요	선택 (LLM 매핑용)

라이선스

MIT

Name		Name	Last commit message	Last commit date
Latest commit History 43 Commits
.claude		.claude
.kiro		.kiro
docs		docs
output_haiku_backup		output_haiku_backup
scripts		scripts
src/mcp_to_cli		src/mcp_to_cli
tests		tests
tools		tools
.gitignore		.gitignore
CLAUDE.md		CLAUDE.md
README.md		README.md
pyproject.toml		pyproject.toml

Folders and files

Latest commit

History

Repository files navigation

MCP Tool Forge

Why?

The MCP Token Cost Problem

How MCP Tool Forge Solves This

Skills Installation (3 lines)

Claude Code

Kiro-CLI

Global Installation (all projects)

Install Specific Skills

9 AWS Skills

Skills Coverage

Skills Test Results

Token Economics: MCP vs Skills

Why Does MCP Consume So Many Tokens?

How Skills Save Tokens

Cost Estimation (Claude Opus 4.6, $15/1M input tokens)

Quick Start (CLI Tool)

Multi AWS Profile Support

Output Formats

Architecture

Extraction Results

Supported Servers (67)

Project Structure

Development

Requirements

License

MCP Tool Forge (한국어)

왜 필요한가?

MCP의 토큰 비용 문제

MCP Tool Forge의 해법

Skills 설치 (3줄이면 끝)

Claude Code

Kiro-CLI

글로벌 설치 (모든 프로젝트에 적용)

특정 스킬만 설치

9개 AWS Skills

Skills 커버리지

Skills 테스트 결과

토큰 경제학: MCP vs Skills

왜 MCP는 토큰을 많이 소모하는가?

Skills는 어떻게 토큰을 절약하는가?

비용 환산 (Claude Opus 4.6 기준)

빠른 시작 (CLI 도구)

멀티 AWS 프로필 지원

출력 형식

아키텍처

추출 결과

지원 서버 (67개)

프로젝트 구조

개발

요구사항

라이선스

About

Topics

Resources

Uh oh!

Stars

Watchers

Forks

Releases

Packages 0

Uh oh!

Contributors

Uh oh!

Languages

Packages