AgentForge
What to Build in the Agent Era — AI 에이전트 시대의 선택장애 해결 프레임워크
No API key needed. AgentForge runs entirely through your locally installed Claude Code.
AgentForge helps Claude Code users and AI power-users overcome decision paralysis
about what to build next. It combines personal ontology mapping, local environment
scanning, GitHub trend analysis, and Claude-powered recommendations to produce a
personalized, immediately actionable MVP — all using your existing Claude Code session.
The Problem
You use Claude Code every day. You see the potential. But when it comes to
what to build, you freeze.
- Too many ideas, no framework to evaluate them
- You don't know what gaps exist in the ecosystem
- You don't know which opportunities fit your specific skills
- Even when you pick something, starting from scratch is slow
AgentForge solves all of this in one automated pipeline.
Pipeline Overview
AgentForge Pipeline
===================
[You]
|
v
+-------------------+
| 1. Ontology Survey| 20 questions about skills, domain,
| (20 questions) | pain points, time, risk, motivation
+-------------------+
|
| Personal Ontology JSON
v
+-------------------+
| 2. Env Scanner | pip packages, Claude skills, MCP servers,
| | VS Code extensions, npm globals, git repos
+-------------------+
|
| ToolProfile
v
+-------------------+
| 3. GitHub Search | Trending AI/agent repos across 5 categories
| | (1000+ stars, active in last 90 days)
+-------------------+
|
| RepoInfo[]
v
+-------------------+
| 4. Gap Analysis | Claude analyzes ecosystem for gaps,
| (Claude Code) | pain themes, and opportunities
+-------------------+
|
| GapAnalysis
v
+-------------------+
| 5. Recommender | Combines ontology + tools + gaps
| (Claude Code) | -> Top 5 personalized recommendations
+-------------------+
|
| [User selects #N]
v
+-------------------+
| 6. Data Collector | Fetches GitHub issues, competitive analysis,
| | integration opportunities, MVP scope
+-------------------+
|
| ProjectBrief
v
+-------------------+
| 7. MVP Builder | Claude generates complete working codebase:
| (Claude Code) | Python files, README, tests, Makefile
+-------------------+
|
v
[Runnable MVP in your output directory]
Requirements
- Python 3.11+
- Claude Code installed and logged in
- (Optional)
GITHUB_TOKEN for higher GitHub API rate limits
No ANTHROPIC_API_KEY needed. AgentForge uses your Claude Code session directly.
Quick Install
git clone https://github.com/AlexAI-MCP/AgentForge
cd AgentForge
pip install -e .
Quick Start
# Run the full pipeline (no API key needed!)
agentforge run
# Answer 20 questions, get personalized recommendations, build your MVP
That's it. AgentForge handles everything else using your local Claude Code.
Live Demo Output
Real output from AlexLee's profile (A- · 91/100):
#1 ConstructIQ Korea FIT:9.8 MARKET:8.5 Medium 8weeks
건설 현장의 모든 지식을 AI로 연결하는 건설 특화 GraphRAG 플랫폼
#2 OntologyMe FIT:9.2 MARKET:7.8 Medium 6weeks
한국 최초 퍼스널 지식 그래프 SaaS
#3 MCPHub Korea FIT:8.7 MARKET:8.2 Hard 10weeks
한국어 MCP 스킬 마켓플레이스
#4 GraphRAG Enterprise FIT:9.0 MARKET:9.0 Hard 12weeks
한국 기업 문서를 위한 엔터프라이즈 GraphRAG 레이어
#5 AI Community OS FIT:9.5 MARKET:7.2 Easy 4weeks
1200명 Discord 커뮤니티를 위한 지식 운영체제
Usage
Full Pipeline
agentforge run
Runs all 6 steps interactively. At step 5 you choose which recommendation to
build. All intermediate results are saved so you can re-run individual steps.
Options:
--skip-scan Skip the local environment scan
--skip-github Skip GitHub search (use cached results)
--auto-build Automatically build the top recommendation
Individual Commands
# Run just the 20-question survey and build your Personal Ontology
agentforge survey
# Scan your local development environment
agentforge scan
# Search GitHub for trending AI repos and run gap analysis
agentforge search
agentforge search --days 60 --limit 15 # custom parameters
# Generate recommendations (requires prior survey)
agentforge recommend
# Build MVP for a specific recommendation
agentforge build 1 # build recommendation #1
agentforge build # interactive prompt to choose
How It Works
Step 1: Personal Ontology Survey
Twenty questions covering:
| Category |
Questions |
| Technical Skills |
Primary language, years experience, AI tools used |
| Claude Depth |
How deeply you use Claude Code / Anthropic API |
| Domain Expertise |
Primary and secondary industry knowledge |
| Pain Points |
Biggest workflow frustration + dream tool |
| Availability |
Weekly hours, commitment level |
| Strategy |
Target user, monetization preference, risk tolerance |
| Context |
Team situation, geographic market, open-source stance |
| Motivation |
Why you build, past successes, fears, superpower |
Claude then synthesizes these into a Personal Ontology — a structured JSON
map of your strengths, gaps, opportunities, builder style, and ideal project
traits.
Step 2: Environment Scanner
Non-destructive local scan that detects:
- Python packages (pip list)
- Anthropic, OpenAI, LangChain SDKs
- Claude Code skills (~/.claude/skills/)
- MCP server configurations
- VS Code extensions
- npm global packages
- Nearby git repositories
- Available CLI tools (git, docker, node)
Step 3: GitHub Search
Searches 5 categories across 4 queries each (20 total searches):
agent-frameworks — multi-agent orchestration
llm-tools — prompt engineering, RAG, embeddings
automation — workflow, RPA, computer use
developer-tools — code gen, AI assistants
data-pipelines — vector DBs, knowledge graphs
Returns up to 50 deduplicated repos sorted by stars, filtered to repos
active in the last N days.
Step 4: Gap Analysis
Claude reads the full repo list and identifies:
- Recurring pain themes in open issues
- "Good but incomplete" frameworks
- High-demand / low-quality-supply areas
- Saturated areas to avoid
- Ranked opportunities with potential scores
Step 5: Recommendation Engine
Combines all three inputs (ontology + tools + gaps) and Claude generates 5
personalized recommendations, each with:
- Project name and concept
- Why it fits you specifically (not generic advice)
- Market opportunity score (1-10)
- Technical difficulty score (1-10)
- Estimated MVP timeline in weeks
- Similar existing projects to differentiate from
- Concrete first steps (immediately actionable)
- Monetization path and risk factors
Step 6: Data Collector
Once you choose a recommendation:
- Fetches open issues from similar/competing projects
- Classifies issue themes (missing features, bugs, doc gaps)
- Claude synthesizes a detailed ProjectBrief including:
- Precise MVP scope (what to build first)
- Pain-point evidence from real users
- Competitive landscape with specific weaknesses
- Integration opportunities
- Success metrics
- Risk/mitigation pairs
Step 7: MVP Builder
Claude generates a complete, runnable Python project:
- Full package structure with working code
- CLI entry point
- README with setup instructions
- pyproject.toml / requirements.txt
- .env.example
- Makefile
- pytest test suite
- All files written to a timestamped output directory
Architecture
agentforge/
├── __init__.py
├── cli.py Click CLI with 5 commands
├── config.py Pydantic config, env var loading
├── ontology/
│ ├── survey.py 20-question interactive survey
│ └── builder.py Claude-powered ontology synthesis
├── scanner/
│ └── tools.py Local environment scanner
├── github/
│ ├── searcher.py Async GitHub API client
│ └── analyzer.py Claude-powered gap analysis
├── recommender/
│ └── engine.py Personalized recommendation engine
├── collector/
│ └── data.py GitHub issue research + brief generation
└── mvp/
└── builder.py Claude-powered MVP code generation
All data flows as typed Pydantic/dataclass objects. Claude is called at 4
stages (ontology, gap analysis, recommendations, MVP generation). All
intermediate outputs are persisted as JSON in ./agentforge_output/ so
individual steps can be re-run without repeating expensive API calls.
Output Files
After a full run, ./agentforge_output/ contains:
agentforge_output/
├── survey.json Your survey answers
├── ontology.json Your Personal Ontology
├── tool_profile.json Environment scan results
├── github_repos.json Fetched GitHub repositories
├── gap_analysis.json Ecosystem gap analysis
├── recommendations.json Your 5 recommendations
├── <project>_brief.md Project brief (Markdown)
├── <project>_brief.json Project brief (JSON)
└── <project>_YYYYMMDD_HHMMSS/ Generated MVP codebase
├── README.md
├── pyproject.toml
├── requirements.txt
├── .env.example
├── Makefile
├── <package>/
│ └── *.py
└── tests/
└── *.py
Configuration
All settings via environment variables or .env file:
| Variable |
Required |
Default |
Description |
ANTHROPIC_API_KEY |
Yes |
— |
Your Anthropic API key |
GITHUB_TOKEN |
No |
— |
GitHub PAT (5000 req/hr vs 60) |
AGENTFORGE_MODEL |
No |
claude-opus-4-6 |
Claude model to use |
AGENTFORGE_OUTPUT_DIR |
No |
./agentforge_output |
Output directory |
Get your Anthropic API key at console.anthropic.com.
Get a GitHub token at github.com/settings/tokens
(no permissions needed — public repo read is sufficient).
Requirements
- Python 3.11+
anthropic>=0.40.0
click>=8.1.0
rich>=13.0.0
httpx>=0.27.0
pydantic>=2.0.0
python-dotenv>=1.0.0
jinja2>=3.1.0
For Korean Developers (한국 개발자를 위한 안내)
AgentForge는 Claude Code를 사용하는 한국 개발자들을 위해 특별히 설계되었습니다.
설문조사는 한국어로 진행됩니다 (영어 부제목 포함).
추천 시스템은 한국 시장의 특성을 고려합니다:
- 한국어 우선 시장의 기회
- 국내 개발자 커뮤니티의 미충족 수요
- 글로벌 진출을 위한 영어 도구화 전략
AI 에이전트 시대에 무엇을 만들어야 할지 더 이상 고민하지 마세요.
AgentForge가 당신의 강점, 도구, 시장 갭을 분석하여 가장 적합한 프로젝트를 추천합니다.
Contributing
Contributions are welcome. Please:
- Fork the repository
- Create a feature branch (
git checkout -b feature/your-feature)
- Write tests for new functionality
- Ensure all tests pass (
pytest)
- Run linting (
ruff check . && black --check .)
- Submit a pull request
Development Setup
git clone https://github.com/agentforge/agentforge
cd agentforge
python -m venv .venv
source .venv/bin/activate # or .venv\Scripts\activate on Windows
pip install -e ".[dev]"
cp .env.example .env
# Add your API keys to .env
pytest
License
MIT License — see LICENSE file.
Acknowledgments
Built for Claude Code users who want to spend less time deciding and
more time building. Powered by Anthropic Claude.
AgentForge — Stop deciding. Start building.
더 이상 고민하지 말고, 지금 바로 만드세요.