Scarlet

To find what is hidden, hold a séance. To bind the names you find, give them to Scarlet.

Scarlet now ships inside khimaira as a workspace member at packages/scarlet/. Its tools are re-registered on khimaira's MCP server under source-prefixed names (mcp__khimaira__scarlet_analyze_project, etc.). The standalone scarlet serve path below still works for backward compatibility.

Scarlet is a codebase cartographer — an MCP server that walks any project, extracts structural metadata via tree-sitter, and generates the documentation scaffolding AI assistants need to actually understand a feature-organized codebase: per-feature CLAUDE.md files, barrel exports (index.js/index.ts), Mermaid dependency graphs, and symbol manifests.

Built for Claude Code, but works with any MCP-compatible client.

The dark cousin of Séance: Séance retrieves knowledge from the dead code via vector search; Scarlet inscribes the names of every entity into a permanent record. Both now live inside the khimaira workspace.

Why "Scarlet"

Named for the Scarlet Woman of Revelation 17 — the figure who sits on a beast "full of names of blasphemy" with a name written on her forehead and a cup full of mysteries in her hand. She is defined by the inscribed names she bears. A catalog walking around in human form.

That maps directly onto what the tool does:

• Catalogs named entities (components, hooks, slices, types, interfaces) from a codebase
• Inscribes them into permanent records — per-feature CLAUDE.md files, barrel exports, symbol manifests
• Surfaces what was hidden — gotchas, invariants, TODOs, deep imports that bypass barrels

Other scarlet threads in scripture reinforce the metaphor:

• Isaiah 1:18 — "though your sins be as scarlet…". Scarlet was the dye that couldn't be washed out; the stain that remained visible. Documentation fixes structure in the same way — makes the shape of a feature indelible so refactors can't silently erase it.
• Rahab's scarlet cord (Joshua 2:18) — a visible marker placed on the house that should be preserved when everything around it is destroyed. Every generated CLAUDE.md is a scarlet cord: a feature tied to an inscription so it survives refactors that would otherwise erase its shape.

There's also Hawthorne's Scarlet Letter — a public mark that fixes identity, making a person's defining characteristic impossible to ignore. Scarlet does the same to features: once their symbols are inscribed, their identity is legible to anyone (or anything) reading the record.

And it pairs with its sibling tool in the ritual:

To find what is hidden, hold a séance. To bind the names you find, give them to Scarlet.

Séance is the summoning. Scarlet is the scribe at the ritual — the one who takes down what comes through.

What it does

flowchart LR
    A[Any Project] -->|claude mcp add| B[Scarlet MCP]
    B --> C[Tools]
    C --> D[analyze_project]
    C --> E[scan_features]
    C --> F[extract_feature_metadata]
    C --> G[generate_barrel]
    C --> H[build_claude_md]
    C --> I[generate_dep_graph]
    C --> J[lint_claude_md]

    D & E & F & G & H & I & J --> L[Project gets:<br/>CLAUDE.md per feature<br/>Barrel exports<br/>Mermaid dep graph<br/>Lint reports]

The core principle

Deterministic parts run in Python. Judgment parts run in Claude.

Generated by Scarlet (Python + tree-sitter)	Generated by Claude (synthesis)
Public API list	Vocabulary (domain terms)
Key files index	Conventions and invariants
Cross-feature consumers	Common task recipes
Dependency graph	Known issues and gotchas
Barrel exports	One-paragraph feature description
Symbol manifests	The why behind every "don't do X"

The tools generate structure. Claude generates meaning. On every CLAUDE.md update, the auto-derivable sections refresh and the human/Claude-written sections are preserved using <!-- BEGIN MANUAL --> ... <!-- END MANUAL --> markers.

Indexing pipeline

flowchart TD
    A[Project root] --> B[analyze_project]
    B --> C{Detect framework, state mgmt,<br/>features root, package manager}
    C --> D[scan_features]
    D --> E[For each feature]
    E --> F[Walk source files]
    F --> G[Tree-sitter parse]
    G --> H[Extract exports:<br/>components, hooks, slices,<br/>interfaces, types]
    H --> I[Find consumers via import graph]
    I --> J[Build skeleton from template]
    J --> K{Existing CLAUDE.md?}
    K -->|Yes| L[Preserve manual sections,<br/>refresh auto sections]
    K -->|No| M[Write new skeleton with<br/>placeholder manual sections]
    L --> N[Write CLAUDE.md]
    M --> N

The CLAUDE.md template

Every generated CLAUDE.md follows the same shape — auto sections refresh on every update, manual sections persist:

• Description (manual) — one-paragraph summary
• Vocabulary (manual) — domain terms specific to this feature
• Public API (auto) — exported components, hooks, slices, types
• Key files (auto) — every source file with the symbols it defines
• Conventions and patterns (manual) — invariants, idioms, the why behind every rule
• Consumers (auto) — features that import from this one
• Common tasks (manual) — step-by-step recipes
• Known issues and gotchas (manual) — landmines, deprecated patterns
• See also (auto) — cross-references to study guides and rules

Installation

Prerequisites

• Python 3.12+
• uv package manager

Setup

git clone git@github.com:fsocietydisobey/scarlet.git
cd scarlet
uv sync

No API keys required — Scarlet is fully local and stateless.

CLI usage

# Scan a project — overview + feature table
uv run scarlet scan ~/work/jeevy_portal

# Generate or refresh one feature's CLAUDE.md
uv run scarlet describe ~/work/jeevy_portal drawings
uv run scarlet describe ~/work/jeevy_portal drawings --dry-run   # preview only

# Refresh all feature CLAUDE.md files
uv run scarlet sync ~/work/jeevy_portal

# Generate a barrel export for a feature
uv run scarlet barrel ~/work/jeevy_portal drawings --ext js

# Generate a Mermaid dependency graph
uv run scarlet graph ~/work/jeevy_portal -o shared-docs/feature-deps.mmd

# Lint all CLAUDE.md files for staleness and missing sections
uv run scarlet lint ~/work/jeevy_portal

# Start the MCP server
uv run scarlet serve

MCP integration

Claude Code

Register globally so it's available in every project:

claude mcp add -s user scarlet -- uv --directory /path/to/scarlet run scarlet serve

Once registered, Claude gets these tools:

Tool	What it does
analyze_project	Detect framework, state mgmt, package manager, feature count
scan_features	List all features with state (CLAUDE.md, barrel, counts)
extract_feature_metadata	Pull all exports from one feature
generate_barrel	Create a barrel export file for one feature
build_claude_md	Generate or refresh a feature's CLAUDE.md
generate_dep_graph	Output Mermaid or JSON dependency graph
lint_claude_md	Validate a feature's CLAUDE.md for staleness
list_consumers	Find all features that import from a given feature

Other MCP clients

{
  "mcpServers": {
    "scarlet": {
      "command": "uv",
      "args": ["--directory", "/path/to/scarlet", "run", "scarlet", "serve"]
    }
  }
}

Configuration

Create a .scarlet.yml at the root of any project to declare conventions:

project_type: nextjs
state_management: redux-toolkit
test_framework: jest
features_root: frontend/src/features
study_guides_path: frontend/study-guides
barrel_export_strategy: re_export_default

This makes Scarlet reusable across projects with different shapes. Without .scarlet.yml, Scarlet uses sensible defaults and detects most things heuristically.

Architecture

graph TB
    subgraph "src/scarlet/"
        CLI[cli.py<br/>CLI entry point]
        SRV[server.py<br/>MCP server]
        CFG[config.py<br/>.scarlet.yml loader]

        subgraph "analyzer/"
            PR[project.py<br/>framework detection]
            FT[features.py<br/>feature scanning]
            MD[metadata.py<br/>tree-sitter export extraction]
            IM[imports.py<br/>import graph walker]
        end

        subgraph "generator/"
            BR[barrel.py<br/>index.js generator]
            CM[claude_md.py<br/>CLAUDE.md generator]
            DG[dep_graph.py<br/>Mermaid renderer]
            TPL[template.py<br/>skeleton template]
        end

        subgraph "validator/"
            LN[linter.py<br/>staleness checks]
        end
    end

    CLI --> PR & FT & MD & BR & CM & DG & LN
    SRV --> PR & FT & MD & BR & CM & DG & LN & IM
    CM --> MD & IM & TPL
    BR --> MD
    DG --> IM

Composition with Séance and Serena

Scarlet is the writer. It produces structure for the read tools to consume:

flowchart BT
    A[Source Code<br/>ground truth]
    A --> B[Serena<br/>LSP - exact symbolic queries]
    A --> C[Séance<br/>vector search - semantic queries]
    B & C --> D[Scarlet<br/>analysis + writing]
    D --> E[Generated Docs<br/>CLAUDE.md, barrels, dep graphs]
    E --> F[Architecture MCP<br/>serves curated docs]

Tool	Question it answers	Mode
Serena	"Where is symbol X defined? Who calls it?"	Read (LSP-precise)
Séance	"Find code related to concept X"	Read (semantic-fuzzy)
Scarlet	"Generate or refresh the structural docs"	Write
Architecture MCP	"What does the curated map say about X?"	Read (curated prose)

The four together form a complete loop: Scarlet writes, the others read.

Tech stack

• tree-sitter — AST parsing for Python, TypeScript, JavaScript
• MCP Python SDK — Model Context Protocol server
• Click — CLI framework
• PyYAML — Config file parsing
• uv — Package management

No vector database, no API keys, no embeddings. Scarlet is fully local, deterministic, and stateless.

License

MIT