enzyme

Enzyme reads a knowledge base — markdown files, Obsidian vaults, Readwise exports, any text corpus — and compiles it into a concept graph. The graph captures the cross-cutting themes in your material as catalysts: pre-computed questions that an agent can search through instead of grepping through your files.

No conversation history needed. No runtime reasoning. The expensive work happens once at init. After that, queries run locally in ~8ms on an on-device embedding model.

Install

curl -fsSL https://raw.githubusercontent.com/jshph/enzyme/main/install.sh | bash

On macOS you can also use Homebrew:

brew install jshph/enzyme/enzyme-cli

Then add the Claude Code plugin:

claude plugin marketplace add jshph/enzyme
claude plugin install enzyme

MCP server

If you prefer MCP over the plugin, Enzyme ships a stdio MCP server that works with any MCP-compatible client (Claude Desktop, Cursor, etc):

claude mcp add enzyme -- enzyme mcp

The MCP server exposes init, petri, catalyze, and status tools — you can initialize and explore your vault entirely from the client without running CLI commands separately.

Quick start

cd /path/to/your/vault    # any folder of markdown files
enzyme init                # compiles concept graph — under 20s for 1k docs

Inside Claude Code, invoke /enzyme to explore your vault by concept.

What it does

Enzyme reads the structure of your knowledge base — tags, links, folders, timestamps — and builds semantic clusters with temporal weight on every entity. From those clusters it generates catalysts: thematic questions that cut across your content and surface connections keyword search can't reach.

A search for "why we keep rewriting the auth layer" finds the ADR from six months ago, a retro note about scope creep, and a Readwise highlight on accidental complexity — even if none of those share keywords with the query.

Core concepts

• Entities — the tags, links, and folders in your content. Each one becomes a semantic cluster.
• Catalysts — pre-computed themes Enzyme discovers across your material. Searching through catalysts connects content that keyword and vector search miss.
• Petri — the compiled index: what's trending, what entities exist, and what catalysts are anchored to each.
• Apply — project your concept graph onto an unfamiliar corpus. enzyme apply /path/to/other/repo maps your catalysts onto new content. See apply docs.

Example: petri output

enzyme petri | jq '.entities[:2]'

[
  {
    "name": "system-design",
    "type": "tag",
    "activity_trend": "active",
    "frequency_12m": 84,
    "catalysts": [
      {
        "text": "What does the commitment to simplicity cost when the pressure to ship keeps winning?",
        "context": "velocity vs craft in infrastructure",
        "era": "2024-Q3"
      },
      {
        "text": "Where does the analysis of user needs gather information that delays rather than clarifies the core value?",
        "context": "research as avoidance",
        "era": "2025-Q1"
      }
    ]
  },
  {
    "name": "working-with-others",
    "type": "tag",
    "activity_trend": "rising",
    "frequency_12m": 47,
    "catalysts": [
      {
        "text": "What assumptions about leadership are held by those who are good at building things?",
        "context": "craft vs delegation",
        "era": "2024-Q4"
      },
      {
        "text": "How does the goal of not depending on others shape the approach to collaboration?",
        "context": "independence vs team trust",
        "era": "2025-Q2"
      }
    ]
  }
]

Each entity has catalysts spanning different eras — questions that cut across months of writing. These are what the agent searches through, not your raw text.

Example: catalyze query

enzyme catalyze "why we keep rewriting the auth layer"