---
name: deep-research
description: "Orchestrate autonomous deep web research, parallel scraping, citation synthesis, fact checking, and persistent memory using the Deep Research MCP server. Use when asked to 'deep research', 'research this topic', 'fact check claim', 'compare sources', or perform comprehensive multi-source investigation with zero API keys."
category: ai-agents
risk: safe
author: Prabhakar Pal
tags: [research, mcp, search, web-scraping, fact-check, agents]
tools: [mcp-deep-research-server, claude, cursor]
---

# Deep Research Agent Skill

This skill guides AI agents on orchestrating multi-step, evidence-backed research using the tools exposed by `mcp-deep-research-server`.

## When to Use

- When the user asks for in-depth, cited research on a technology, company, scientific topic, or market trend.
- When an objective fact-check is requested to verify controversial claims or rumors.
- When comparing 2–5 distinct URLs or source narratives to identify consensus vs contradictions.
- When saving or retrieving structured research memory across long-running tasks.

---

## Tool Reference & Routing

| Operation | Primary Tool | Description |
|---|---|---|
| **Full Investigation** | `deep_research` | Automated pipeline: searches DuckDuckGo, parallel-scrapes 3–8 pages, extracts stats/entities, synthesizes cited report. |
| **Targeted Query** | `search_web` | Clean search (1–10 results) with time filtering (`day`, `week`, `month`, `year`). |
| **Page Deep-Dive** | `scrape_page` | Cleans HTML into Turndown markdown, extracts main content, headings, and links. SSRF-safe. |
| **Cross-Source Analysis** | `compare_sources` | Scrapes 2–5 URLs in parallel, surfaces common entities, shared consensus, and contradictory claims. |
| **Verification & Audit** | `fact_check_claim` | Searches supporting and debunking sources, returning an evidence-backed heuristic verdict. |
| **Extract Key Points** | `extract_insights` | Heuristic scoring of statistics, named entities, key sentences, and answered questions. |
| **Long-Term Memory** | `memory_save` / `memory_search` | Persistent storage in `~/.mcp-deep-research/memory.json` for cross-session recall. |

---

## Autonomous Research Workflow

Follow this 5-stage discipline for every deep investigation:

```mermaid
graph TD
    A[User Request] --> B[1. Check Memory: memory_search]
    B --> C[2. Formulate 2-3 Multi-Angle Queries]
    C --> D[3. Run deep_research or search_web + parallel scrape]
    D --> E[4. Cross-Examine: compare_sources / fact_check]
    E --> F[5. Synthesize & Cite with URL Provenance]
    F --> G[6. Persist Key Findings: memory_save]
```

### Stage 1: Check Prior Knowledge
Before firing new web queries, check if relevant research exists in persistent memory:
```json
// Example call:
memory_search({ "query": "Fine-tuning LLMs", "limit": 3 })
```

### Stage 2: Query Formulation
Formulate search queries with precise intent:
- For technical comparisons: `"<Tech A> vs <Tech B> benchmarks latency architecture 2026"`
- For market analysis: `"<Industry> market size CAGR top players report"`
- For troubleshooting: `"<Exact Error String> <Framework> root cause fix"`

### Stage 3: Orchestrated Deep Scraping
Invoke `deep_research` with appropriate depth:
- `depth: "quick"` (3 sources) → Fast sanity checks.
- `depth: "standard"` (5 sources) → Standard reports and architectural evaluations.
- `depth: "deep"` (8 sources) → Comprehensive market maps and critical fact-checking.

```json
// Example call:
deep_research({
  "topic": "State of Model Context Protocol ecosystem and adoption in 2026",
  "depth": "deep",
  "saveMemory": true
})
```

### Stage 4: Fact-Checking & Contradiction Resolution
If two sources provide opposing data (e.g. performance metrics or release dates):
1. Run `compare_sources` with the URLs to view the entity alignment matrix.
2. Run `fact_check_claim` on the specific disputed assertion.

### Stage 5: Structured Synthesis & Output Format
Deliver research outputs adhering to this consistent, publication-ready format:

```markdown
# [Research Topic Title]

## 📋 Executive Summary
A 2-3 sentence distillation answering the user's core question directly with key conclusions.

## 🔑 Key Findings & Empirical Data
- **Finding 1**: Direct takeaway with [Source Name](URL) link.
- **Finding 2**: Quantitative metric / benchmark comparison.

## 📊 Comparison / Architecture Table (if applicable)
| Dimension | Option A | Option B | Consensus / Verdict |
|---|---|---|---|
| Latency | ... | ... | ... |
| Cost | ... | ... | ... |

## ⚖️ Evidence & Source Provenance
1. [Source 1 Title](URL) — Key contribution / excerpt
2. [Source 2 Title](URL) — Key contribution / excerpt

## ⚠️ Known Gaps & Unverified Claims
- Notes on conflicting dates or unverified benchmarks.
```

---

## Safety & Best Practices

1. **Never make unverified claims**: Always link the primary URL as inline markdown (`[Source](url)`).
2. **Handle SSRF & Private Networks**: The underlying server automatically blocks private IP ranges (`127.x`, `10.x`, `192.168.x`, `169.254.x`). If a URL is rejected, report that the host is on a restricted network.
3. **Respect Rate Limits**: When scraping multiple pages, use `compare_sources` or `deep_research` which run with concurrency throttles rather than firing unbounded individual requests.