The Problem
You understand a codebase by building a mental model — not of files, but of behaviors: what happens when a user signs up, how payments flow, where things can fail. That model lives only in your head. It's rebuilt from scratch every time. It can't be shared, inspected, or queried.
Cartographer externalizes that model.
It gives Claude Code a persistent memory of your codebase's behaviors and structure — grounded in source evidence, navigable as a map, and accessible to engineers and non-engineers alike.
What It Does
You: "explore this codebase"
Claude traces the system's key behaviors end-to-end.
Entities emerge from the flows — not from directory structure.
Boundaries form around concerns that break together.
The map shows what the system DOES, not how files are organized.
Each fact links back to the exact source code that proves it.
Open localhost:3847 to see the map.
Click a boundary to zoom in (semantic zoom).
Select a behavior flow to see the path light up.
Ask Claude questions — answers cite the stored model.
Getting Started
Prerequisites
Install via Plugin Marketplace
/plugin marketplace add miltonian/cartographer
/plugin install cartographer@cartographer
Install from Source
git clone https://github.com/miltonian/cartographer.git
cd cartographer
npm install
npm run build:ui
claude --plugin-dir ./plugin
Use
/cartographer explore # Deep autonomous exploration
/cartographer analyze # Guided analysis (or focused: analyze auth)
/cartographer review-pr 123 # Overlay a PR on the map
/cartographer map # Open browser UI
/cartographer status # Model statistics
/cartographer snapshot # Save a backup
/cartographer restore <file> # Restore from backup
/cartographer reset # Start fresh
How It Works
Three Pieces
┌──────────────────────────┐
│ Claude Code + Plugin │ Agent. Traces behaviors, records findings.
│ │ MCP (stdio) │
├──────────┼───────────────┤
│ Local Service │ Backbone. Stores world-model, serves API.
│ │ HTTP + WS │
├──────────┼───────────────┤
│ Browser UI │ Map. Renders projections, navigable depth.
└──────────────────────────┘
Behavior-First Analysis
Cartographer leads with what happens, not what exists. The agent:
- Identifies the system's key behaviors ("what are the most important things this system does?")
- Traces each behavior end-to-end through the code
- Records entities as they're encountered along the flow
- Lets boundaries emerge from which entities cluster together
Structure comes from behavior, not from directories.
The Ontology
| Entity Kind | What It Represents |
|---|
boundary | A concern — things that break together |
capability | Something the system can do |
actor | Where external intent enters |
entity | State that persists or transforms |
side-effect | Observable consequence outside the boundary |
invariant | A rule that must hold true |
failure-point | Where the system can fail |
Evidence, Not Hallucination
Every fact carries source evidence with confidence levels:
| Confidence | Meaning |
|---|
proven | Directly observed in source code |
high | One inference step from observed facts |
medium | Synthesized from multiple observations |
low | Educated guess |
speculative | Hypothesis, not yet verified |
The map visually distinguishes all levels. You always know what's proven vs. inferred.
Behavior Flows
Flows are the primary unit of understanding — a named path through the system:
Customer checkout:
Cart validation → Payment processing → Order creation → Confirmation email
Select a flow in the UI panel to see it highlighted on the map. Each step is numbered. Non-participating nodes dim.
Perspectives
Multiple angles on the same codebase. A perspective is a named lens — a subset of entities in focus.
- Default — everything (the satellite view)
- Agent-created — focused on a concern ("auth", "data pipeline")
- Boundary-derived — created by semantic zoom (clicking into a boundary)
Each perspective has its own layout. Open multiple browser tabs with different perspectives via URL params.
Semantic Zoom
