使用说明
Test Pilot
Automated MCP server generation, deployment, and AI-agent testing — from GitHub repos to live endpoints in one click.
Test Pilot takes GitHub repository URLs containing OpenAPI/Swagger specs, generates MCP (Model Context Protocol) servers, deploys them and the upstream APIs to Blaxel, then runs an AI agent to validate everything end-to-end.
Repository Structure
.
├── generate.py # CLI: standalone spec → MCP generator
├── pipeline/ # Core pipeline library
│ ├── ingest.py # Parse OpenAPI 3.x / Swagger 2.x / Postman v2.1
│ ├── mine.py # Discover MCP tools from API endpoints
│ ├── safety.py # Safety classification & execution policy
│ ├── codegen.py # LLM-powered MCP server code generation
│ ├── models.py # Shared data models
│ ├── reasoning.py # LLM reasoning helpers
│ └── logger.py # Logging setup
├── blaxel-swagger-finder/ # Backend API server
│ ├── backend.py # FastAPI + SSE pipeline orchestrator
│ ├── scanner.py # Blaxel sandbox repo cloner & spec extractor
│ ├── agent_tester.py # AI agent end-user testing engine
│ └── upstream_services/ # Auto-generated upstream API deployments
├── Columbia-Hackathon-Test-Pilot-Frontend/ # React frontend (Vite + TailwindCSS)
│ ├── src/
│ │ ├── components/pipeline/ # Pipeline UI (sidebar, stepper, step content)
│ │ ├── hooks/usePipeline.ts # SSE-driven pipeline state management
│ │ └── pages/Pipeline.tsx # Main pipeline page
│ └── vite.config.ts
├── examples/ # Sample OpenAPI specs
├── output/ # Generated MCP server output
└── .env # API keys (not committed)
How It Works
GitHub Repo URLs
│
▼
┌──────────────────┐
│ 1. CLONE │ Clone repos into Blaxel sandbox
└──────┬───────────┘
▼
┌──────────────────┐
│ 2. EXTRACT │ Find & extract OpenAPI/Swagger specs
└──────┬───────────┘
▼
┌──────────────────┐
│ 3. INGEST │ Parse specs → endpoints, schemas, metadata
└──────┬───────────┘
▼
┌──────────────────┐
│ 4. DISCOVER │ Mine MCP tool capabilities from endpoints
└──────┬───────────┘
▼
┌──────────────────┐
│ 5. POLICY │ Classify safety levels, apply execution rules
└──────┬───────────┘
▼
┌──────────────────┐
│ 6. GENERATE │ DeepSeek-V3 LLM → FastMCP server code
└──────┬───────────┘
▼
┌──────────────────┐
│ 7. MCP TESTS │ Validate generated code (syntax, deps, tools)
└──────┬───────────┘
▼
┌──────────────────┐
│ 8. DEPLOY │ Deploy upstream APIs + MCP servers to Blaxel
└──────┬───────────┘
▼
┌──────────────────┐
│ 9. AGENT TEST │ AI agent calls live MCP tools end-to-end
└──────┬───────────┘
▼
✅ Live MCP endpoints + test report
Quick Start
Prerequisites
- Python 3.11+
- Node.js 18+ and npm
- Blaxel CLI:
brew tap blaxel-ai/blaxel && brew install blaxel
1. Install dependencies
# Backend
cd blaxel-swagger-finder
pip install -r requirements.txt
# Frontend
cd ../Columbia-Hackathon-Test-Pilot-Frontend
npm install
2. Configure environment
Create a .env in the project root:
BL_API_KEY=your-blaxel-api-key
BL_WORKSPACE=your-workspace-name
FEATHERLESS_API_KEY=your-featherless-key
GEMINI_API_KEY=your-gemini-key # optional fallback
3. Run
# Terminal 1 — Backend (port 8000)
cd blaxel-swagger-finder
python backend.py
# Terminal 2 — Frontend (port 8080)
cd Columbia-Hackathon-Test-Pilot-Frontend
npm run dev
Open http://localhost:8080, paste GitHub repo URLs, and hit Start.
Standalone CLI
Generate an MCP server from a spec file without the web UI:
python generate.py examples/sample.yaml
python generate.py https://petstore.swagger.io/v2/swagger.json --name petstore
python generate.py path/to/spec.json --no-deploy -v
Generated Output
Each MCP server is written to output/<server-name>/:
output/<server-name>/
├── src/server.py # Complete FastMCP server (LLM-generated)
├── blaxel.toml # Blaxel deployment config + env vars
├── pyproject.toml # Python project config
├── test_server.py # Auto-generated test suite
└── requirements.txt # Dependencies
Tech Stack
- Backend: Python, FastAPI, SSE streaming, Blaxel SDK
- Frontend: React, Vite, TailwindCSS, shadcn/ui, Framer Motion
- LLM: DeepSeek-V3 via Featherless (code generation), Gemini (spec cleanup fallback)
- Infrastructure: Blaxel (sandboxes, serverless functions, MCP hosting)
- Testing: AI agent with LLM-generated test plans, narrative + analytical result summaries
