bryancostanich/cercano-deep-research

Cercano Deep Research

Multi-source research tool that takes a topic and intent, identifies authoritative sources, systematically searches each one, and compiles a ranked, annotated encyclopedia of findings.

MCP Tool

Tool name: cercano_deep_research

Parameters

| Parameter | Type | Required | Description | |-----------|------|----------|-------------| | topic | string | Yes | The research topic to investigate. | | intent | string | Yes | What you need this research for — drives relevance scoring and source selection. | | depth | string | No | "survey" (quick scan, ~2 min), "standard" (balanced, ~5-8 min), or "deep" (exhaustive, ~15+ min). Default: "standard". | | date_range | string | No | Filter results by date (e.g. "2024-2026", "last 2 years"). | | sources | string[] | No | Override auto-detected sources. If omitted, sources are chosen based on topic domain. | | output_dir | string | No | Write report to this directory as multiple files. Recommended for standard/deep research and required for incremental deepening. | | project_dir | string | No | Project root directory. | | phase | string | No | Run a specific phase: "plan", "search", "analyze", "synthesize". Omit to run all phases. | | use_model | string | No | Override the default model for this research run. |

Depth Tiers

| | Survey | Standard (default) | Deep | |---|---|---|---| | Sources | 2-3 | 3-4 | 4-5 | | Results/query | 3 | 4 | 6 | | Reference chasing | None | 1-hop, max 15 | 1-hop, max 50 | | Analysis | 3-pass (facts, relevance, quality gate) | 3-pass | 3-pass | | Target time | ~2 min | ~5-8 min | ~15+ min |

How It Works

1. Source Planning — Local model analyzes topic + intent and identifies relevant sources from 25+ options across academic, industry, news, reference, and regulatory categories
2. Systematic Search — Searches each source using tailored queries (free APIs for PubMed, arXiv; site-scoped DuckDuckGo for others)
3. Content Extraction — Fetches and extracts readable content from top results
4. Analysis & Annotation — 3-pass pipeline: fact extraction, relevance scoring (1-5), quality gate with retry
5. Reference Chasing (standard/deep only) — Identifies cited works relevant to intent, searches and analyzes them
6. Synthesis — Executive summary, narrative synthesis, contradiction detection, gap analysis, reading order, follow-up suggestions

Incremental Deepening

Run a survey first to get a quick landscape, then deepen to standard or deep without re-doing prior work:

1. Run survey with an output_dir
2. Review the results
3. Run again with the same output_dir and a deeper depth — existing findings are preserved, new sources are added, and middle-scored findings (2-4) are re-evaluated with richer context

The tool stores state in research_state.json inside the output directory. Each deepening pass expands the plan with complementary sources and enriches the analysis.

Output

Structured markdown report with:

• Executive Summary (TL;DR)
• Source Plan (which sources were searched and why)
• Ranked Findings (sorted by relevance, with annotations)
• Discovered References (works found via citation chasing)
• Synthesis (narrative connecting the findings)
• Contradictions & Open Debates
• Gap Analysis (what the research didn't find)
• Recommended Reading Order
• Suggested Follow-Up Research
• Next Steps (suggested deeper research command)

Examples

Quick survey:

{"topic": "quantum computing error correction", "intent": "preparing a conference talk", "depth": "survey", "output_dir": "/tmp/qec-research"}

Standard research (default):

{"topic": "CRISPR gene therapy for sickle cell disease", "intent": "writing a grant proposal for a novel delivery mechanism", "output_dir": "/tmp/crispr-research"}

Deep research:

{"topic": "transformer architecture improvements", "intent": "literature review for PhD thesis", "depth": "deep", "date_range": "2024-2026", "output_dir": "/tmp/transformer-research"}

Incremental deepening (survey → standard):

{"topic": "quantum computing error correction", "intent": "preparing a conference talk", "depth": "standard", "output_dir": "/tmp/qec-research"}