A Claude Code plugin for multi-agent deep research with evidence graphs, quality gates, and publication-quality report generation.
Turns Claude Code into a deep research system that goes beyond simple summarization:
- Evidence-first: Every claim traces to specific source excerpts. No unsupported assertions.
- Triangulation: Key claims require corroboration from multiple independent sources, or are explicitly marked as single-source.
- Honest uncertainty: Conflicts between sources are surfaced, not hidden. Confidence levels are calibrated.
- Structured artifacts: Research runs produce reproducible, inspectable data (sources, claims, evidence edges, notes) — not just a final essay.
- Quality gates: Hooks enforce that agents can't mark work complete without producing required artifacts.
claude --plugin-dir /path/to/deep-research-pluginOr add to your Claude Code settings to load it automatically.
- Claude Code v2.1.32+
- Python 3.10+
- For full multi-agent parallelism, enable Agent Teams:
export CLAUDE_CODE_EXPERIMENTAL_AGENT_TEAMS=1
/deep-research:run What are the implications of the EU AI Act for US technology companies?
This kicks off the full pipeline: planning → source discovery → deep reading → claim extraction → triangulation → synthesis → adversarial review → report generation.
/deep-research:status
/deep-research:report
/deep-research:init
| Agent | Model | Role |
|---|---|---|
dr-lead |
Opus | Orchestrates the run, decomposes questions, assigns tasks, performs adversarial review |
dr-scout |
Sonnet | Wide-pass discovery: searches, evaluates sources, writes preliminary notes |
dr-analyst |
Opus | Deep reading, atomic claim extraction, triangulation, conflict identification |
dr-writer |
Opus | Synthesis and publication-quality report generation |
- Intake & Planning — Decompose the question into research strands, generate diverse query sets
- Wide-Pass Discovery — Search across core, synonym, contrarian, primary-source, and time-bounded queries
- Deep Analysis — Read top sources, extract atomic claims with citations, build evidence edges
- Triangulation — Cross-reference claims across sources, flag single-source and contested claims
- Synthesis — Write evidence-backed narrative with explicit confidence levels
- Adversarial Review — Attempt to falsify key claims, check for missing perspectives
- Audit & Report — Validate all artifacts, generate the final structured report
Each research run produces a directory under .deep-research/runs/ containing:
run.json — Run metadata and configuration
plan.md — Research plan with strands and query strategy
queries.json — Generated search queries by category
sources.jsonl — Sources with metadata and credibility scoring
notes/ — Structured notes per source
claims.jsonl — Atomic claims with citations
evidence.jsonl — Triangulation edges (supports/contradicts)
conflicts.md — Documented disagreements between sources
synthesis.md — Narrative synthesis
report.md — Final structured report
audit.json — Quality audit results
Hooks enforce artifact requirements:
- TaskCompleted: Blocks task completion if required artifacts are missing or incomplete
- TeammateIdle: Prevents agents from going idle before finishing assigned work
- PreCompact: Snapshots run state before context compaction
The plugin works with built-in WebSearch and WebFetch by default. For enhanced capabilities, configure MCP servers for:
- Search APIs (Brave, Tavily, Google Custom Search)
- Academic metadata (Semantic Scholar, CrossRef)
- Internal tools (Notion, Slack, GitHub)
MIT