DEV Community

gentic news
gentic news

Posted on • Originally published at gentic.news

Grainulator: The MCP-Powered Research Plugin That Forces Claude Code to Prove Its Claims

#ai #machinelearning #research #deeplearning

Grainulator transforms Claude Code into a research engine with typed claims, conflict detection, and confidence scoring—forcing AI to prove its work.

Grainulator: The MCP-Powered Research Plugin That Forces Claude Code to Prove Its Claims

What It Does — Structured Research with Accountability

Grainulator is a Claude Code plugin that replaces vague research prompts with systematic, auditable research sprints. When you ask Claude to "research how our auth system works," Grainulator orchestrates a multi-pass process where every finding becomes a typed claim stored in claims.json. These claims are then adversarially challenged, confidence-graded, and compiled into decision-ready briefs.

The core innovation: evidence tiers that grade confidence from stated → web → documented → tested → production. The compiler runs 7 passes over claims—checking type coverage, evidence strength, conflict detection, and bias—and produces a confidence score. If unresolved conflicts exist, it blocks output until you resolve them.

Setup — Install in 2 Minutes 

# Install from Claude plugin marketplace
claude plugin marketplace add https://github.com/grainulation/grainulator/blob/main/.claude-plugin/marketplace.json
claude plugin install grainulator

![Glitchy — the Grainulator mascot](https://github.com/grainulation/grainulator/raw/main/site/glitchy.png)


# If you get SSH permission errors (common with GitHub)
git config --global url."https://github.com/".insteadOf "git@github.com:"
# Then retry the install

# Or clone manually
git clone https://github.com/grainulation/grainulator.git ~/.claude/plugins/grainulator
claude plugin add ~/.claude/plugins/grainulator
Enter fullscreen mode Exit fullscreen mode

Requirements: Node.js ≥20 (needed for MCP servers that run via npx).

For team deployment, add to your project's .claude/settings.json:

{
  "enabledPlugins": ["grainulator@grainulation-marketplace"]
}
Enter fullscreen mode Exit fullscreen mode

How To Use It — No Slash Commands Needed

Once installed, just talk to Claude normally. The intent router detects what you want:

  • "research how our auth system works" — Runs a full research sprint
  • "challenge r003" — Adversarial testing of claim ID r003
  • "what are we missing?" — Blind spot analysis
  • "write it up" — Generates compiled brief

Launch research sprints with: "research X using grainulator"

The grainulator subagent (agents/grainulator.md) runs autonomously, reading compiler output to decide what command to run next—research, challenge, witness, blind-spot—until reaching decision-ready confidence.

MCP Architecture — Four Specialized Servers

Grainulator uses MCP (Model Context Protocol) servers for specialized tasks:

  1. wheat — Claims engine that manages typed claims and evidence
  2. mill — Format conversion between different claim representations
  3. silo — Knowledge store for persistent claim storage
  4. DeepWiki — Codebase research capabilities

This MCP architecture means each component can be updated independently and potentially used by other Claude Code plugins.

When To Use It — Beyond Basic Code Questions

Use Grainulator when you need:

  1. Architecture decisions — Research multiple approaches with evidence grading
  2. Library comparisons — Get confidence-scored evaluations instead of opinions
  3. Codebase audits — Systematically document findings with conflict detection
  4. Onboarding research — New team members can explore codebases with structured output
  5. Documentation generation — Compiled briefs become living documentation

The adversarial challenge system is particularly valuable for catching AI hallucinations—forcing Claude to provide evidence before you accept claims.

Try the Demo First

Before installing, visit grainulator.app for an interactive demo with in-browser AI (WebLLM with SmolLM2-360M, ~200MB download). The PWA demonstrates the research sprint concept with 50 pre-generated demo topics and live local inference.

The CLAUDE.md Integration

Add this to your project's CLAUDE.md to document your research methodology:

## Research Standards

We use Grainulator for systematic research with evidence tiers:
- stated → web → documented → tested → production

All findings become typed claims in `claims.json`
Run "challenge [claim-id]" to test robustness
Compile with "write it up" when confidence >80%
Enter fullscreen mode Exit fullscreen mode

Originally published on gentic.news

Top comments (0)