ToolStorePy

ToolStorePy is an automatic MCP (Model Context Protocol) server builder.

Describe the tools you need in plain English. ToolStorePy finds the best matching implementations from a curated vector index, clones the repositories, audits them for security, and generates a ready-to-run MCP server — in one command.

---

📚 Documentation

Full architecture details, examples, and advanced usage:

https://tool-store-py-docs.vercel.app/

---

📦 Install

pip install toolstorepy

PyPI: https://pypi.org/project/toolstorepy/

---

✨ What It Does

Given a queries.json file:

[
  { "tool_description": "evaluate a mathematical arithmetic expression securely" },
  { "tool_description": "convert between different units of measurement" },
  { "tool_description": "calculate cryptographic hash of a file" }
]

Run:

toolstorepy build --queries queries.json --index core-tools

ToolStorePy will:

1. Download a vector index of curated tool repositories
2. Run semantic retrieval + cross-encoder reranking
3. Clone matching repositories via a local bare-repo cache
4. Run a static AST security scan on every repo
5. Optionally run an LLM-based security review (autonomous, no human prompt)
6. Merge .env.example files and validate required secrets
7. Extract @tool functions via AST
8. Generate a unified MCP server
9. Optionally launch the server immediately

Output:

toolstorepy_workspace/
├── mcp_unified_server.py
├── security_report.txt
├── .env.example
└── .venv/

---

🚀 Quick Start

# Using the built-in index
toolstorepy build \
  --queries queries.json \
  --index core-tools

# Using a remote index URL
toolstorepy build \
  --queries queries.json \
  --index-url https://your-index-url.zip

# Custom port
toolstorepy build \
  --queries queries.json \
  --index core-tools \
  --port 9090

# LLM-based security scan (autonomous, no human prompt)
toolstorepy build \
  --queries queries.json \
  --index core-tools \
  --llm-scan

---

⚙️ CLI Reference

build

toolstorepy build --queries <path> [options]

Flag	Default	Description
--queries	required	Path to queries.json
--index	—	Built-in index name
--index-url	—	Remote index archive URL
--workspace	toolstorepy_workspace	Workspace directory
--install-requirements	off	Install repo requirements.txt files
--host	0.0.0.0	Host the MCP server binds on
--port	8000	Port the MCP server listens on
--llm-scan	off	Enable LLM-based autonomous security review
--llm-model	claude-sonnet-4-6	Model for LLM scanning (any LangChain-supported string)
--force-refresh	off	Re-download cached index
--verbose	off	Verbose logging + full tracebacks

cache

# Pre-warm cache from a resolved queries file (items must have git_link)
toolstorepy cache populate --queries resolved.json

# Pre-warm cache from explicit URLs
toolstorepy cache populate --url https://github.com/org/repo1.git \
                           --url https://github.com/org/repo2.git

# List cached repos
toolstorepy cache list

# Clear cache
toolstorepy cache clear

---

🔐 Security Scanning

Every repository is scanned before inclusion. Two modes are available:

AST scan (default)

Static analysis using Python's ast module. Fast and deterministic.

Severity	What triggers it
HIGH	eval/exec, os.system, subprocess with shell=True, pickle.loads, yaml.load without Loader=
MEDIUM	Capability imports (subprocess, pickle, raw sockets, unsafe XML parsers), dynamic reflection
LOW	HTTP clients, crypto primitives, deprecated modules

Repos with HIGH findings prompt you to include or skip before the build proceeds.

LLM scan (--llm-scan)

An LLM reviews the full source of each repo and makes an autonomous INCLUDE/SKIP decision. The human prompt is bypassed entirely — the build proceeds automatically based on the LLM verdict.

Both scanners can be used together. When --llm-scan is active, AST findings are merged into the same security report and the LLM decision is final.

Model-agnostic via LangChain

The LLM scanner uses LangChain's init_chat_model so any supported provider works:

# Claude (default)
export ANTHROPIC_API_KEY=sk-...
toolstorepy build --queries q.json --index core-tools --llm-scan

# GPT-4o
export OPENAI_API_KEY=sk-...
toolstorepy build ... --llm-scan --llm-model gpt-4o

# Gemini
export GOOGLE_API_KEY=...
toolstorepy build ... --llm-scan --llm-model gemini-2.0-flash

Install the integration package for your chosen provider:

pip install langchain-anthropic    # Claude  → ANTHROPIC_API_KEY
pip install langchain-openai       # GPT     → OPENAI_API_KEY
pip install langchain-google-genai # Gemini  → GOOGLE_API_KEY

The security report (workspace/security_report.txt) is always written regardless of which scan mode is used. LLM findings are tagged [LLM] in the report so they're visually distinct from AST findings.

---

🔑 Secret Management

If any cloned tool includes a .env.example, ToolStorePy automatically:

• Merges all .env.example files across repos
• Resolves conflicts interactively
• Validates completeness against an existing .env
• Documents required keys at the top of the generated server file

Output: workspace/.env.example

Steps:

1. Copy .env.example → .env in your workspace
2. Fill in the required values
3. Re-run the server

---

🌐 Server Configuration

The generated server runs on streamable-http transport. Host and port are baked in at build time:

# Default: 0.0.0.0:8000
toolstorepy build --queries q.json --index core-tools

# Custom host/port
toolstorepy build --queries q.json --index core-tools --host 127.0.0.1 --port 9090

The generated mcp_unified_server.py contains:

if __name__ == "__main__":
    mcp.run(transport='streamable-http', host='127.0.0.1', port=9090)

---

🏗️ Pipeline Overview

queries.json
      │
      ▼
vector index retrieval
      │
      ▼
semantic search + reranking
      │
      ▼
repository cloning (bare-repo cache)
      │
      ▼
AST security scan
      │
      ▼
LLM security scan (optional, --llm-scan)
      │
      ▼
.env merge + validation
      │
      ▼
@tool extraction via AST
      │
      ▼
MCP server synthesis

---

⚡ Repository Cache

Repositories are cached as bare git clones at:

~/.repo_cache/

Subsequent builds reuse the cache, skipping remote clones entirely. Pre-warm it manually with:

toolstorepy cache populate --url https://github.com/org/repo.git

---

📁 Project Structure

toolstorepy/
├── cli.py
├── orchestrator.py
├── config.py
├── builder/
│   ├── mcp_builder.py
│   └── parser.py
├── index/
│   ├── downloader.py
│   └── registry.py
├── loader/
│   ├── cache.py
│   └── repo.py
├── search/
│   ├── rerank.py
│   └── semantic.py
└── utils/
    ├── env_merger.py
    ├── llm_scanner.py
    └── security_scanner.py

---

🧩 Extending ToolStorePy

Goal	Location
Add a built-in index	index/registry.py
Change embedding model	orchestrator.py
Add AST security rules	utils/security_scanner.py
Tune LLM scan prompt	utils/llm_scanner.py
Modify MCP server output	builder/mcp_builder.py
Adjust @tool detection	builder/parser.py

---

🗺️ Roadmap

• toolstore.yaml manifest support
• Public tool submission portal
• Versioned index publication
• Dry-run preview mode
• Build manifest export (cross-build pollution fix)
• Async tool support
• Hardcoded secret detection in AST scanner
• Aliased-import tracking in security scanner

---

📜 License

MIT License — Copyright (c) 2025 Sujal Maheshwari

See LICENSE for full terms. Attribution required in public-facing derivative works.

---

🤝 Contributing

Contributions welcome. Open issues or submit pull requests following the existing module structure.

https://github.com/sujal-maheshwari2004/toolstorepy