grainulator

Research sprint orchestrator for Claude Code.

Ask a question, get a decision-ready brief. Every finding is tracked as a typed claim, adversarially challenged, confidence-graded, and compiled into self-contained output. Zero third-party dependencies.

Try it

grainulator.app — interactive demo with in-browser AI. No install needed.

Install

Step 1 — Add the marketplace

claude plugin marketplace add https://github.com/grainulation/grainulator/blob/main/.claude-plugin/marketplace.json

Step 2 — Install the plugin

claude plugin install grainulator

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

If you see [email protected]: Permission denied (publickey), the plugin installer is trying SSH but you don't have SSH keys set up with GitHub. Fix with one command:

git config --global url."https://github.com/".insteadOf "[email protected]:"

Then retry claude plugin install grainulator. This tells git to use HTTPS instead of SSH for all GitHub repos.

Alternatively, clone manually:

git clone https://github.com/grainulation/grainulator.git ~/.claude/plugins/grainulator
claude plugin add ~/.claude/plugins/grainulator

For team-wide deployment, commit to your project's .claude/settings.json:

{
  "enabledPlugins": ["grainulator@grainulation-marketplace"]
}

Quick start

Once installed, just talk to Claude:

• "research how our auth system works" — runs a multi-pass research sprint
• "challenge r003" — adversarial testing of a specific claim
• "what are we missing?" — blind spot analysis
• "write it up" — generates a compiled brief

No slash syntax required. The intent router detects what you want and runs the right workflow.

How it works

Claims are the unit of knowledge. Every finding from research, challenges, witnesses, and prototypes is stored as a typed claim in claims.json.

Evidence tiers grade confidence: stated → web → documented → tested → production.

The compiler runs 7 passes over your claims — type coverage, evidence strength, conflict detection, bias scan — and produces a confidence score. If there are unresolved conflicts, it blocks output until you resolve them.

Skills

Autonomous agent

The grainulator subagent (agents/grainulator.md) runs multi-pass research sprints autonomously. It reads the compiler output to decide what command to run next — research, challenge, witness, blind-spot — until the sprint reaches decision-ready confidence.

Launch it from Claude Code with: "research X using grainulator".

Demo site

grainulator.app is a PWA that demos the research sprint concept in-browser:

• Mobile-first chat interface — no shader on mobile, dark background
• In-browser AI via WebLLM (SmolLM2-360M, ~200MB download)
• 50 pre-generated demo topics with fuzzy matching (instant Track A)
• Live local inference when model downloads (Track B)
• Conversational responses with progressive claim disclosure
• Compile flow with 7-pass confidence scoring

Architecture

• Plugin manifest: .claude-plugin/plugin.json
• MCP servers: wheat (claims engine), mill (format conversion), silo (knowledge store), DeepWiki (codebase research)
• Skills: skills/<name>/SKILL.md — 13 prompt-engineered workflows
• Agent: agents/grainulator.md — autonomous sprint subagent
• Hooks: Auto-compile on claim mutation, write-guard on claims.json and compilation.json
• Orchard: Multi-sprint orchestration via orchard.json dependency graphs

The ecosystem

Part of the grainulation ecosystem. Eight tools, each does one thing.

You don't need all eight. Start with claude plugin install grainulator. That's it.

Enterprise deployment

Three levels:

1. Team lead: Commit .claude/settings.json with enabledPlugins to your repo
2. IT admin: Deploy managed settings via MDM with pre-approved permissions
3. Air-gapped: Use CLAUDE_CODE_PLUGIN_SEED_DIR with the plugin baked into container images

Zero dependencies

Every grainulation tool runs on Node built-ins only. No npm install waterfall. No left-pad. No supply chain anxiety. The MCP servers download on first use via npx — no global install required.

License

MIT