Unified MCP Tool Graph is a research-driven project that aggregates and structures tool APIs from diverse Model Context Protocol (MCP) servers into a centralized Neo4j graph database. This graph functions as an intelligent infrastructure layer that enables large language models (LLMs) and agentic AI systems to dynamically retrieve the most relevant tools for any task — without being overwhelmed by redundant or confusing options.
- The system now spins up only the MCP servers required for a given user query. Five popular MCP servers (including the Dynamic Tool Retriever MCP) are kept warm by default; others are started on demand and kept alive for 10 minutes after last use.
- Dynamic Tool Retriever MCP returns not just tool metadata, but also the config needed to run/connect to the MCP server for each tool (fetched from the vendor's GitHub README automatically).
- Automatic MCP Config Extraction: Uses the vendor's GitHub repo to extract the MCP server config (from README) for each tool, so agents can spin up/connect to the right server on the fly.
- Error Handling: If config extraction fails, the system logs a warning and continues, ensuring robust tool retrieval.
- Only the exact tools required for the user query are loaded into the agent's context (not all tools from all MCP servers). This prevents LLM confusion and infinite tool loops.
- End-to-End Flow:
- User query is received.
- Dynamic Tool Retriever MCP queries the Neo4j graph and returns the top relevant tools plus their MCP server configs.
- The agent spins up/connects to only the required MCP servers (using the configs), and loads only the retrieved tools.
- The agent executes the workflow and returns the answer.
- A2A Agent Example: See
Example_Agents/A2A_DynamicToolAgent/for a fully dynamic A2A agent that orchestrates MCP servers and tools per request. - LangGraph Example: See
Example_Agents/Langgraph/for a LangGraph agent using the same dynamic, minimal-tool approach.
🔬 This repository focuses on the creation and evolution of the Unified Tool Graph Database. Chatbot-based integration (e.g., LangChain) is treated as a modular extension of this foundational layer.
📢 Support for Cline, IDE's coming soon..
As LLMs and autonomous agents evolve to interact with external tools and APIs, a critical bottleneck has emerged:
How can models efficiently select the right tool from an ever-expanding universe of APIs — without going into infinite loops or picking the wrong ones?
-
Tool Confusion:
LLMs struggle when many tools offer similar functions (e.g.,create_post,schedule_post,post_to_social), leading to indecision and incorrect tool calls. -
↺ Infinite Chains:
Without a structured understanding of tool differences, LLMs often get stuck in unproductive chains, calling tools repetitively or selecting suboptimal ones. -
Unstructured Access:
Most current implementations dump all available tools into the LLM's context, overwhelming it with options and increasing hallucination risks.
This project proposes a structured, queryable solution: a vendor-agnostic Neo4j graph database of tools/APIs sourced from MCP servers (e.g., LinkedIn, Google, Facebook, Notion, etc.).
-
Centralized Tool Intelligence:
Store API descriptions, metadata, parameters, and inter-tool relationships in a graph format. -
LLM-Friendly Query Layer:
Agents can retrieve only the 3–4 most relevant tools per task using metadata and relationships, minimizing confusion. -
Semantic Differentiation:
Capture similarities and differences between tools using graph relationships (e.g.,overlaps_with,extends,preferred_for_task) to guide decision-making.
While the graph is the core, it enables powerful downstream use cases:
A modular LangChain/Autogen chatbot extension that queries the graph and surfaces a minimal, accurate toolset for any given user intent.
This prevents LLMs from blindly scanning a massive tool library and instead gives them just what they need to complete the job — nothing more, nothing less.
Key Implementation:
- MCP servers are spun up on demand (using configs from the tool retriever MCP and GitHub), and shut down after inactivity.
- Only the 5 most popular MCPs are kept running at all times; others are ephemeral.
- Agents (A2A or LangGraph) only see the tools relevant to the current query, not the full universe of tools.
| Goal | Description |
|---|---|
| Tool Ingestion | Fetch APIs and schemas from public/private MCP servers and normalize them |
| Tool Relationship Mapping | Define graph edges like overlaps_with, requires_auth, preferred_for, belongs_to_vendor |
| LLM-Oriented Queries | Return task-specific tool bundles in real time |
| Scalable Ecosystem | Continuously add vendors and tools without retraining or hardcoding |
| Agent-Aware Structure | Guide LLM reasoning with metadata-rich, searchable tool representations |
-
Reduces Tool Confusion in LLMs
Prevents tool overload by showing only task-relevant options. Avoids infinite call loops and incorrect tool selections. -
Vendor-Agnostic Integration
Unifies APIs from different providers into a single intelligent system. -
Maps Interoperability
Captures how tools relate or depend on each other, useful for chaining APIs in workflows. -
Optimized Agentic Reasoning
Empowers LLMs to reason efficiently with fewer distractions in the context window. -
Scalable & Modular
Can be updated independently of LLM or chatbot infrastructure. Extendable across any agent stack.
-
"I want to schedule a post on LinkedIn and share it in Slack."
→ Graph returns only the relevantcreate_post,schedule_post, andsend_messagetools. -
Custom AI Assistants for Enterprises:
Only expose internal tools from the graph, filtered by access, scope, or function. -
Smart Recommender Agents:
Suggest best-matched tools based on tags, popularity, success rate, or dependencies.
Integrations with LangGraph and A2A are available in the Example_Agents directory for streamlined agent workflows and dynamic tool orchestration.
┌─────────────────────────────────────────────────────────────┐
│ MCP Unified Gateway │
│ (Port 8000) │
└─────────────────────┬───────────────────────────────────────┘
│
┌─────────────────────┴───────────────────────────────────────┐
│ MCP Proxy Server │
│ (Port 9000) │
└─┬─────────────┬─────────────┬─────────────┬─────────────────┘
│ │ │ │
┌─▼──────────┐ ┌▼─────────────▼┐ ┌─▼─────────▼┐ ┌─▼─────────────┐
│ Dynamic │ │ Everything │ │Sequential │ │ Time Server │
│ Tool │ │ Server │ │ Thinking │ │ │
│ Retriever │ │ │ │ Server │ │ │
└────────────┘ └──────────────┘ └────────────┘ └───────────────┘
│
┌─────▼─────┐
│ Neo4j │ (Optional)
│ Database │
└───────────┘
- User submits a query (e.g., "Schedule a LinkedIn post and share it in Slack.")
- Dynamic Tool Retriever MCP queries the Neo4j graph and returns the most relevant tools plus their MCP server configs (fetched from GitHub if needed).
- MCP Server Manager spins up/connects to only the required MCP servers (using the configs), and keeps them alive for 10 minutes after last use.
- Agent (A2A or LangGraph) loads only the retrieved tools and executes the workflow.
- Result is returned to the user, with minimal tool confusion and maximum efficiency.
- Graph Ingestion Scripts
- Schema Blueprint + Cypher Queries
- Tool Visualization Playground
- LangChain DTR Chatbot Plug-in
- How-to Tutorials & Use Cases
git clone https://github.com/your-username/unified-mcp-tool-graph.git
cd unified-mcp-tool-graph
# Coming soon: ingestion pipeline, schema docs, and sample queriesIf you’re passionate about agentic AI, graph databases, or LLM integration — we’d love your help!
- Submit ideas or vendor sources
- Open PRs for schema/design improvements
- Star the repo to support this research
MIT License — free for academic, personal, and commercial use.
Instead of dumping 100+ tools into a model’s prompt and expecting it to choose wisely, the Unified MCP Tool Graph equips your LLM with structure, clarity, and relevance.
It fixes tool confusion, prevents infinite loops, and enables modular, intelligent agent workflows.
Let’s build smarter systems — one tool graph at a time.
Star the repo to follow the journey and make tools truly intelligent, searchable, and modular.
Created a GraphDB with Vendor and tools with embedings for vector search. Need a way to categorize vendors and connect vendors with same category. eg( Web Search, File System, Social Media etc.) No of Tools Supported 11066 No of MCP Servers(Official + Community) 4161
Here are the top 5 tools recommended based on your task of creating a LinkedIn post about AI trends, along with brief explanations to help you choose:
---
### 1. **linkedin_post_generator** (Likely the Most Relevant)
**Description**: Generates optimized LinkedIn posts with engaging headlines, content structure, and hashtags tailored for professional audiences.
**Why Use**: Directly creates polished posts while suggesting multimedia (images, videos) and tone adjustments for maximum engagement.
---
### 2. **ai_search** (Top Search Tool)
**Vendor**: Higress AI-Search
**Key Feature**: Aggregates **real-time data** from Google/Bing/Quark and academic sources like ArXiv for cutting-edge AI research and industry trends.
**How to Use**:
- Query: `Latest AI trends 2024` → Get recent news, research papers, and opinion pieces.
- Useful for ensuring your post cites fresh, credible sources.
---
### 3. **tavily-search** (Comprehensive Web Research)
**Vendor**: Tavily
**Key Feature**: AI-powered search with advanced filters (e.g., `time_range`, `exclude_domains`) to curate content from trusted sources.
**How to Use**:
- Example: Search for "recent AI breakthroughs 2024" → Pull articles from tech blogs, IEEE, or MIT Tech Review to highlight in your post.
---
### 4. **query_repository** (If Curating Open-Source or GitHub Content)
**Vendor**: GitHub Chat
**Use Case**:
- If your post references open-source projects (e.g., "Top AI tools from GitHub 2024"), use this to query repositories for trending AI-related code/repos.
---
### 5. **social_media_analytics** (Post-Creation Optimization)
**Use Case**:
- After drafting your post, analyze engagement metrics (e.g., "Which AI topics are trending this week?") to refine your draft based on what performs well on LinkedIn.
---
### Recommended Workflow:
1. **Research** using **ai_search** or **tavily-search** to gather the latest AI news and data (e.g., LLM progress, ethical AI debates).
2. **Craft the Post** with **linkedin_post_generator** to ensure professional formatting and engagement hooks.
3. **Validate** with quick checks via **query_repository** (if citing open-source tools) or **social_media_analytics** for topic popularity.
Let me know if you'd like step-by-step instructions for any specific tool!
This project uses mcp-proxy to expose stdio-based MCP servers as HTTP endpoints. mcp-proxy acts as a bridge between stdio MCP servers and HTTP clients, supporting both Streamable HTTP and SSE transports.
- MCP servers are launched as subprocesses (stdio) and registered with
mcp-proxy. - Each server is exposed at:
http://localhost:<port>/servers/<name>/(Streamable HTTP, POST)http://localhost:<port>/servers/<name>/sse(SSE, GET)
- The official MCP Python SDK can connect to the
/sseendpoint usingsse_client:
from mcp.client.sse import sse_client
from mcp.client.session import ClientSession
import asyncio
async def main():
mcp_url = "http://localhost:9000/servers/time/sse"
async with sse_client(mcp_url) as (read, write):
async with ClientSession(read, write) as session:
await session.initialize()
# ... interact with the server ...
if __name__ == "__main__":
asyncio.run(main())See the mcp-proxy documentation for more details on configuration and advanced usage.