google-research-mcp
Implements Anthropic's multi-agent research architecture with subagent spawning, adaptive stopping, and citation processing for automated research.
README
Google Research MCP Server v2.0.0 - Multi-Agent Architecture
An MCP server that implements Anthropic's Multi-Agent Research Architecture with true subagent spawning, adaptive stopping, and citation processing.
Architecture Overview
This implementation is fully compliant with Anthropic's multi-agent research system:
┌─────────────────────────────────────────────────────────────────┐
│ Multi-Agent Research System │
├─────────────────────────────────────────────────────────────────┤
│ │
│ ┌──────────────────────────────────────────────────────────┐ │
│ │ LEAD RESEARCHER (Orchestrator) │ │
│ │ │ │
│ │ • think(plan approach) - Decompose into aspects │ │
│ │ • create subagents - Spawn parallel workers │ │
│ │ • think(synthesize) - Combine findings │ │
│ │ • evaluate coverage - "More research needed?" │ │
│ │ • complete_task - Return final report │ │
│ └──────────────────────────────────────────────────────────┘ │
│ │ │
│ ┌───────────────┼───────────────┐ │
│ ▼ ▼ ▼ │
│ ┌────────────────┐ ┌────────────────┐ ┌────────────────┐ │
│ │ SUBAGENT 1 │ │ SUBAGENT 2 │ │ SUBAGENT N │ │
│ │ (Aspect A) │ │ (Aspect B) │ │ (Aspect N) │ │
│ │ │ │ │ │ │ │
│ │ • web_search │ │ • web_search │ │ • web_search │ │
│ │ • think(eval) │ │ • think(eval) │ │ • think(eval) │ │
│ │ • complete │ │ • complete │ │ • complete │ │
│ └────────────────┘ └────────────────┘ └────────────────┘ │
│ │ │ │ │
│ └───────────────┼───────────────┘ │
│ ▼ │
│ ┌──────────────────────────────────────────────────────────┐ │
│ │ CITATION AGENT │ │
│ │ • Process documents │ │
│ │ • Identify citation locations │ │
│ │ • Insert inline citations [1], [2], etc. │ │
│ │ • Generate references section │ │
│ └──────────────────────────────────────────────────────────┘ │
│ │ │
│ ┌──────────────────────────────────────────────────────────┐ │
│ │ MEMORY MODULE │ │
│ │ • save plan │ │
│ │ • retrieve context │ │
│ │ • persist findings │ │
│ │ • track gaps │ │
│ └──────────────────────────────────────────────────────────┘ │
│ │
└─────────────────────────────────────────────────────────────────┘
Process Flow
Based on Anthropic's sequence diagram:
User LeadResearcher Subagent1 Subagent2 Memory CitationAgent
│ │ │ │ │ │
│──send user query────▶│ │ │ │ │
│ │ │ │ │ │
│ │◀─────────────────────────────────────────────────────│ │
│ │ think(plan approach) │ │
│ │ │ │ │ │
│ │──save plan────────────────────────────────────────▶│ │
│ │ │ │ │ │
│ │──retrieve context──────────────────────────────────▶│ │
│ │ │ │ │ │
│ │ │ │ │ │
│ │══════════════════════════════════════════════════════│ │
│ │ ITERATIVE RESEARCH LOOP │ │
│ │══════════════════════════════════════════════════════│ │
│ │ │ │ │ │
│ │──create subagent──▶│ │ │ │
│ │──create subagent────────────────────▶│ │ │
│ │ │ │ │ │
│ │ │──web_search────▶│ │ │
│ │ │◀───results──────│ │ │
│ │ │ │ │ │
│ │ │ think(evaluate)│ │ │
│ │ │ │ │ │
│ │◀──complete_task────│ │ │ │
│ │ │ │ │ │
│ │ │ │──web_search───▶│ │
│ │ │ │◀──results──────│ │
│ │ │ │ │ │
│ │ │ │ think(evaluate)│ │
│ │ │ │ │ │
│ │◀─────────────────────complete_task───│ │ │
│ │ │ │ │ │
│ │ think(synthesize results) │ │ │
│ │ │ │ │ │
│ │ ┌─────────────────────┐ │ │ │
│ │ │ More research needed?│ │ │ │
│ │ └─────────────────────┘ │ │ │
│ │ │ │ │ │ │
│ │ [Continue] [Exit Loop] │ │ │
│ │ │ │ │ │ │
│ │══════════════════════════════════════════════════════│ │
│ │ │ │ │ │
│ │──complete_task (research result)────────────────────▶│ │
│ │ │ │ │ │
│ │ │ │ │──────────────────▶│
│ │ │ │ │ Process docs + │
│ │ │ │ │ insert citations │
│ │◀───────────────────────────────────────────────────────────────────────│
│ │ │ │ │ Report with │
│ │ │ │ │ citations │
│ │──persist results──────────────────────────────────▶│ │
│ │ │ │ │ │
│◀──return research─────│ │ │ │ │
│ results with │ │ │ │ │
│ citations │ │ │ │ │
Key Features
1. True Subagent Spawning
Each aspect gets its own subagent that runs independently:
- Generates aspect-specific queries
- Executes web searches
- Fetches full page content
- Evaluates findings
- Reports back to Lead Researcher
2. Think/Evaluate Phases
Explicit reasoning phases between iterations:
think(plan approach)- Decompose topic into aspectsthink(evaluate)- Each subagent evaluates its findingsthink(synthesize)- Lead Researcher combines all findings
3. Adaptive Stopping
Dynamic "More research needed?" decision:
- Coverage score calculation (0-100%)
- Configurable thresholds per depth level
- Gap identification and filling
- Exits early when coverage is sufficient
4. Aspect-Based Decomposition
Topics are broken into researchable aspects:
- Basic: 2 aspects (overview, mechanism)
- Moderate: 5 aspects (+use cases, benefits, challenges)
- Comprehensive: 11 aspects (+history, comparisons, implementation, future, research, case studies)
5. Memory Module
Persistent context across iterations:
- Research plan storage
- Findings per aspect
- Gap tracking
- Iteration history
6. Citation Agent
Dedicated citation processing:
- Assigns citation IDs by quality
- Inserts inline citations [1], [2]
- Generates references section
- Groups by quality tier
Tools
| Tool | Description |
|---|---|
google_research |
Full multi-agent research with all components |
deep_search |
Search + fetch full content (single iteration) |
deep_search_news |
News-specific deep search |
fetch_page |
Fetch single page content |
google_search |
Simple search (snippets only) |
web_search |
Search with quality scoring |
research_session |
Manual session management |
run_subagent |
Manually spawn a subagent |
evaluate_coverage |
Check coverage and gaps |
add_source |
Add source to session |
get_citations |
Format citations |
Installation
{
"mcpServers": {
"google-research": {
"command": "npx",
"args": ["google-research-mcp"],
"env": {
"GOOGLE_API_KEY": "your-api-key",
"GOOGLE_CX": "your-search-engine-id"
}
}
}
}
Prerequisites
1. Google API Key
- Go to Google Cloud Console
- Enable "Custom Search API"
- Create an API Key
2. Search Engine ID (CX)
- Go to Programmable Search Engine
- Create engine with "Search the entire web"
- Copy the Search Engine ID
Usage Examples
Full Multi-Agent Research
"Research quantum computing with comprehensive depth"
This triggers the full architecture:
- Lead Researcher plans 11 aspects
- Spawns 3-4 subagents per iteration
- Each subagent researches in parallel
- Evaluates coverage after each iteration
- Continues until 90% coverage or max iterations
- Citation Agent processes final report
Manual Subagent Control
// Create session
research_session({ action: "create", topic: "AI safety", depth: "moderate" })
// Spawn specific subagents
run_subagent({ sessionId: "rs_xxx", aspect: "AI alignment techniques" })
run_subagent({ sessionId: "rs_xxx", aspect: "AI safety research organizations" })
// Check coverage
evaluate_coverage({ sessionId: "rs_xxx" })
// Generate final report
research_session({ action: "complete", sessionId: "rs_xxx" })
Depth Levels
| Depth | Iterations | Aspects | Coverage Threshold | Min Sources/Aspect |
|---|---|---|---|---|
| basic | 2 | 2 | 60% | 2 |
| moderate | 3 | 5 | 75% | 3 |
| comprehensive | 4 | 11 | 90% | 5 |
Source Quality Scoring
Based on Anthropic's source quality heuristics:
| Score | Tier | Examples |
|---|---|---|
| 10 | Primary | .gov, .edu, arxiv, nature.com, PubMed, official docs |
| 8-9 | Authoritative | Wikipedia, Reuters, BBC, NYT, WSJ |
| 7 | Quality | Stack Overflow, TechCrunch, Wired |
| 5-6 | General | Medium, Dev.to, Substack |
| 1-4 | Low | Pinterest, Facebook, Twitter (deprioritized) |
Changelog
v2.0.0 - Multi-Agent Architecture (Anthropic Compliant)
- NEW: True subagent spawning - Parallel workers for different aspects
- NEW: Think/Evaluate phases - Explicit reasoning between iterations
- NEW: Adaptive stopping - Dynamic "More research needed?" decision
- NEW: Aspect-based decomposition - Topics broken into researchable aspects
- NEW: Memory module - Persistent context across iterations
- NEW: Citation Agent - Dedicated citation processing with inline insertion
- NEW:
run_subagenttool - Manual subagent control - NEW:
evaluate_coveragetool - Check coverage and gaps - NEW:
deep_search_newstool - News-specific deep search - Improved report generation with subagent reports
- Full iteration history tracking
v1.2.0 - Deep Research Edition
- Full page content fetching
- Readability-style extraction
- Source quality scoring
v1.0.0
- Initial release
License
MIT
Recommended Servers
playwright-mcp
A Model Context Protocol server that enables LLMs to interact with web pages through structured accessibility snapshots without requiring vision models or screenshots.
Magic Component Platform (MCP)
An AI-powered tool that generates modern UI components from natural language descriptions, integrating with popular IDEs to streamline UI development workflow.
Audiense Insights MCP Server
Enables interaction with Audiense Insights accounts via the Model Context Protocol, facilitating the extraction and analysis of marketing insights and audience data including demographics, behavior, and influencer engagement.
VeyraX MCP
Single MCP tool to connect all your favorite tools: Gmail, Calendar and 40 more.
graphlit-mcp-server
The Model Context Protocol (MCP) Server enables integration between MCP clients and the Graphlit service. Ingest anything from Slack to Gmail to podcast feeds, in addition to web crawling, into a Graphlit project - and then retrieve relevant contents from the MCP client.
Kagi MCP Server
An MCP server that integrates Kagi search capabilities with Claude AI, enabling Claude to perform real-time web searches when answering questions that require up-to-date information.
E2B
Using MCP to run code via e2b.
Neon Database
MCP server for interacting with Neon Management API and databases
Exa Search
A Model Context Protocol (MCP) server lets AI assistants like Claude use the Exa AI Search API for web searches. This setup allows AI models to get real-time web information in a safe and controlled way.
Qdrant Server
This repository is an example of how to create a MCP server for Qdrant, a vector search engine.