Skip to content

Getting Started

  • Node.js >= 20
  • Ollama running locally with a model pulled
Terminal window
# Install Ollama from https://ollama.ai
ollama pull qwen2.5:14b

The qwen2.5:14b model is proven for extraction and review. Smaller models (7b) work but produce less nuanced canon. Larger models (32b+) improve quality but require more VRAM.

Terminal window
npm install -g @mcptoolshop/taste-engine

Verify the installation:

Terminal window
taste --version # 1.0.0
taste doctor # Checks Ollama connectivity and project state
Terminal window
cd /path/to/your/project
taste init my-project --name "My Project"

This creates a .taste/ directory with:

  • taste.db — SQLite database for working state
  • config.json — Project configuration

And a canon/ directory for exported JSON canon files.

Point Taste Engine at your doctrine:

Terminal window
# Single files
taste ingest README.md docs/architecture.md
# Directories (scans recursively for .md files)
taste ingest docs/

The engine detects artifact types automatically: README, architecture note, design doc, feature brief, naming proposal, etc.

Run multi-pass extraction:

Terminal window
taste extract run

This runs 8 specialized passes covering core thesis, design rules, anti-patterns, voice conventions, scope boundaries, quality bars, process rules, and integration contracts.

Terminal window
taste extract status # Run summary
taste extract candidates # List extracted candidates
taste extract contradictions # Show detected contradictions

Extraction produces candidates that need human review:

Terminal window
taste curate queue # Show pending candidates
taste curate inspect <id> # View details with evidence
taste curate accept <id> # Accept into canon
taste curate reject <id> # Reject with reason
taste curate edit <id> # Accept with modifications
taste curate merge <id1> <id2> # Merge duplicates

When done curating, freeze a canon version:

Terminal window
taste curate freeze --tag v1

Freezing creates an immutable snapshot. Reviews always run against frozen canon.

Terminal window
taste review run path/to/artifact.md

The review scores 4 dimensions (0-10 each) and synthesizes a deterministic verdict: aligned, mostly_aligned, salvageable_drift, hard_drift, or contradiction.

Terminal window
taste review list # List all reviews
taste review show <id> # Full review details