Enables harvesting and verification of primary literature from arXiv through the multi-engine crawling pipeline.
Integrates OpenAI's language models as an LLM backend for research synthesis, analysis, and report generation.
Enables harvesting and verification of biomedical literature from PubMed through the multi-engine crawling pipeline.
Click on "Deploy Server".
Wait a few minutes for the server to deploy. Once ready, it will show a "Started" state.
In the chat, type @ followed by the MCP server name and your instructions, e.g., "@deepsearch-research-agent research the efficacy of mRNA vaccines against omicron variants, with confidence scores"
That's it! The server will respond to your query, and you can continue using it as needed.
Here is a step-by-step guide with screenshots.
DeepSearch is a state-of-the-art autonomous deep research agent and consensus synthesis engine. Designed to tackle open-ended, complex scientific, technological, and market intelligence questions, DeepSearch iteratively decomposes research objectives, executes multi-angle boolean search queries, harvests and verifies primary literature, resolves contradictory empirical claims, and outputs publication-grade synthesized reports with mathematical grounding scores and interactive visualizations.
🌳 Recursive Tree-of-Thought Query Decomposition: Breaks down broad questions into multi-stage Directed Acyclic Graphs (DAGs) with automated Boolean operator crafting (site:, filetype:pdf, intitle:).
⚖️ Dialectical Consensus & Contradiction Matrix: Cross-references claims across dozens of sources, computing weighted agreement ratios and isolating contested assumptions.
📐 Calibrated Confidence Scoring: Calculates grounded confidence metrics $C \in [0, 1]$ based on domain authority, recency decay, peer-review weight, and corroboration depth.
🔌 Native Model Context Protocol (MCP) Server: Drop-in MCP integration for Claude Desktop, Cursor, Cline, Zed, and custom LLM runtimes.
🖥️ Deep Research Studio UI: Built-in web studio (design influenced by Material 3) with live query branch visualizer, consensus matrix, interactive SVG metrics charts, and copy-ready client config hub.
📦 Multi-Format Export: Generates peer-reviewed Markdown reports, interactive standalone HTML documents, and machine-readable JSON data bundles.
Related MCP server: Multi-Agent Deep Researcher MCP
+----------------------------------------------------------------------------------------------------+
| DEEPSEARCH SYSTEM ARCHITECTURE |
+----------------------------------------------------------------------------------------------------+
[ Human Query / MCP Client / IDE / Web UI ]
|
v
+----------------------------------------------------------------+
| Stage 1: Autonomous Query Planner |
| - Tree-of-Thought Decomposition (Technical, Market, Risk) |
| - Search Operator Crafting (site:, filetype:pdf, boolean) |
+-------------------------------+--------------------------------+
|
v
+----------------------------------------------------------------+
| Stage 2: Parallel Evidence Harvester |
| - Multi-Engine Crawling (Serper, Tavily, ArXiv, PubMed) |
| - Asynchronous Content Extraction & Boilerplate Scrubbing |
+-------------------------------+--------------------------------+
|
v
+----------------------------------------------------------------+
| Stage 3: Domain Authority & Credibility Engine |
| - Institutional Scoring (0-100), Bias Detection |
| - Recency Decay Factor: R(dt) = exp(-lambda * dt) |
+-------------------------------+--------------------------------+
|
v
+----------------------------------------------------------------+
| Stage 4: Consensus & Contradiction Matrix |
| - Corroborating vs Refuting Claim Partitioning |
| - Weighted Agreement Ratios & Conflict Flagging |
+-------------------------------+--------------------------------+
|
v
+----------------------------------------------------------------+
| Stage 5: Grounded Multi-Perspective Synthesizer |
| - Calibrated Confidence Math: C = sum(w_i * s_i * alpha) |
| - Zero-Unreferenced-Claim Policy & Interactive Visuals |
+-------------------------------+--------------------------------+
|
+-------------------------------------+------------------------------------+
| | |
v v v
[ 📄 Markdown Research Report ] [ 🌐 Material 3 HTML Artifact ] [ 📊 JSON Data Bundle ]# Install via pip
pip install deepsearch-research-agent
# Or run with uvx without installation
uvx deepsearch-research-agent --helpSet your preferred search and LLM API keys:
export SERPER_API_KEY="your_serper_api_key"
export TAVILY_API_KEY="your_tavily_api_key"
export OPENAI_API_KEY="your_openai_api_key"
export ANTHROPIC_API_KEY="your_anthropic_api_key"# Run quick research (1 stage, 5-10 sources)
deepsearch run "CRISPR-Cas12 vs Cas9 off-target editing rates" --depth quick
# Run exhaustive research with custom domain focus and exports
deepsearch run \
"State of Fault-Tolerant Quantum Computing & Logical Qubit Scaling (2026)" \
--depth exhaustive \
--breadth 8 \
--focus-domains "nature.com,science.org,aps.org,ieee.org" \
--output-dir ./output-report \
--format allfrom deepsearch_research_agent import DeepResearchEngine, ResearchConfig
# Configure research depth and constraints
config = ResearchConfig(
depth="exhaustive",
breadth=8,
enable_consensus_matrix=True,
min_source_authority=85,
)
engine = DeepResearchEngine(config=config)
result = engine.run(
query="State of Fault-Tolerant Quantum Computing & Logical Qubit Scaling (2026)"
)
# Access structured data
print(f"Confidence Score: {result.confidence_score:.2%}")
print(f"Harvested Sources: {len(result.sources)}")
print(f"Verified Claims: {len(result.verified_claims)}")
# Export artifacts
result.save_markdown("report.md")
result.save_html("report.html")
result.save_json("data_bundle.json")DeepSearch can be attached as a live context engine in Claude Desktop, Cursor, Cline, and Zed.
# Launch MCP Server in stdio mode
deepsearch mcp
# Launch MCP Server over SSE (HTTP streaming)
deepsearch mcp --transport sse --port 8080claude_desktop_config.json):{
"mcpServers": {
"deepsearch-research-agent": {
"command": "uvx",
"args": ["deepsearch-research-agent", "mcp"],
"env": {
"SERPER_API_KEY": "your_serper_api_key",
"DEEPSEARCH_DEFAULT_DEPTH": "deep"
}
}
}
}For full integration guides for Cursor, Cline, and Zed, see docs/MCP_GUIDE.md and the examples/mcp-clients/ directory.
DeepSearch ships with an interactive, offline-ready web studio (design influenced by Google Material 3 tokens) located at public/index.html.
To launch the studio:
deepsearch studio --port 3000Or open public/index.html directly in any modern browser.
🔍 Deep Research Studio: Live multi-stage query decomposition visualizer, citation stream, and Markdown preview.
📑 Autonomous Query Planner: Interactive Tree-of-Thought graph, custom search operator generator, and boolean string preview.
⚖️ Source & Consensus Matrix: Interactive sources table with authority scoring, claim verification status, and corroboration gauges.
📊 Data Points & SVG Visuals: Native SVG multi-bar, time-series, radar, and donut charts with interactive tooltips.
🤖 AI Agent & MCP Hub: One-click configuration copier for Claude, Cursor, Cline, Zed, and live JSON-RPC tool simulator.
📖 Research Methodology Guide: Built-in interactive architectural documentation and confidence calibration math.
Run the test suite locally:
PYTHONPATH=src pytest tests/ -vThe repository includes a comprehensive 15-job CI matrix in .github/workflows/ci.yml testing across Ubuntu, macOS, and Windows on Python 3.9, 3.10, 3.11, 3.12, and 3.13.
Licensed under the Apache License, Version 2.0.
This server cannot be deployed
MCP-native web evidence and claim verification: cited, source-grounded evidence for AI agents.
Machine-native research commons for agent evidence, discovery, rooms, and bounded research quests.
Shared, peer-validated knowledge archive for AI agents — search, contribute, and validate via MCP
Source-traced evidence research for AI agents. We organise the evidence; you decide.
