◐ Shell
clean mode source ↗

GitHub - miltonian/cartographer: Agent-first code understanding system — builds a persistent, evidence-grounded world-model of your codebase

An agent-first code understanding system.

Build a persistent, evidence-grounded world-model of any codebase.
Navigate it as a map. Trace behaviors. Inspect evidence down to source.

Cartographer map view

Getting Started · How It Works · Architecture · Contributing

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:

  1. Identifies the system's key behaviors ("what are the most important things this system does?")
  2. Traces each behavior end-to-end through the code
  3. Records entities as they're encountered along the flow
  4. 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

Click a boundary to enter it. Its children become the world. The breadcrumb shows your path: Overview › Service › WorldModelStore. Click back to return. The agent can create deeper levels on request.

PR Review

/cartographer review-pr 123

Overlays a PR on the map as a changeset: green (added), amber (modified), red (removed), gray (affected). See the spatial impact of a change at a glance.

Snapshots

/cartographer snapshot before-refactor
/cartographer snapshots
/cartographer restore <filename>

Auto-snapshots before destructive operations. Manual snapshots anytime. Keeps last 10.

MCP Tools

Tool Purpose
cartographer_set_project Set project root
cartographer_write_entity Record an entity with evidence
cartographer_write_relationship Record a relationship
cartographer_write_slice Record a behavior flow or PR changeset
cartographer_create_perspective Create a named lens
cartographer_switch_perspective Switch active lens
cartographer_query Search entities/relationships
cartographer_get_entity Full entity details + evidence
cartographer_get_summary Model statistics
cartographer_snapshot Save a backup
cartographer_restore Restore from backup
cartographer_open_map Open browser UI
cartographer_clear Reset the world-model

Architecture

See docs/architecture.md for the full system diagram.

Key constraints:

  • Behavior-first — trace what happens, let structure emerge
  • Agent-first — Claude Code is the control plane
  • Evidence-grounded — every claim traces to source
  • Language-agnostic — the ontology doesn't know about TypeScript or React
  • UI is a projection — React Flow concepts don't leak into the stored model

Roadmap

  • Stale entity detection (flag entities whose source has changed)
  • User correction from browser UI (edit descriptions, move boundaries)
  • Capability packs (optional AST extractors for higher-confidence facts)
  • Flow animation (step-through replay)
  • Change impact analysis
  • Large codebase optimization
  • Multi-project support
  • Shareable models

Philosophy

Most code tools show you what the code says. Cartographer shows you what the code means.

Understanding is behavior prediction. You understand a system when you can predict what happens under change. That requires a model of flows, boundaries, state, invariants, and failure modes — not just a file tree.

The map is not the territory. But a good map makes the territory navigable.

License

MIT

Contributing

See CONTRIBUTING.md. The most valuable contribution right now is trying it on your codebase and telling us what's missing.