The flight recorder for AI coding — keeps your docs in sync with your code.
A Claude Code skill that automatically maintains documentation as AI agents modify your codebase. Zero dependencies. Fully local.
A single Claude Code session can modify 40+ files across your project. When the session ends, your README, CHANGELOG, and architecture docs remain frozen in the past. Future-you opens the project a week later and has no idea what happened. Future AI sessions hallucinate against stale docs. The gap between code reality and written documentation grows with every session.
Codetape captures the semantic meaning of code changes — not just diffs, but why things changed — and keeps your documentation synchronized. It turns Claude into a code historian that records traces, detects drift, and regenerates docs on command. Everything stays local. No cloud, no accounts, no runtime dependencies.
npx codetape init # Set up tracking in your project# ... make some code changes with Claude Code ...
/trace # Record what changed and why
/trace-sync # Update your docs automaticallyThat's it. Your README, CHANGELOG, and ARCHITECTURE.md stay in sync.
| Command | Description |
|---|---|
/trace |
Record a semantic trace of recent changes |
/trace-sync |
Sync documentation from traces |
/trace-review |
Detect documentation drift |
/trace-map |
Generate architecture docs with dependency graphs |
/trace-log [component] |
Query change history |
/trace-commit |
Generate conventional commit message |
/trace-init |
Initialize tracking (in-session alternative to CLI) |
serve |
Start live dashboard with auto-refresh (CLI: npx codetape serve) |
Codetape has two layers: a CLI that scaffolds the project, and a Skill that provides AI intelligence inside Claude Code sessions.
The CLI (npx codetape init) creates a .codetape/ directory and installs the skill into .claude/skills/codetape/. The skill teaches Claude how to record traces, detect drift, and regenerate docs — all through natural language commands.
All data lives locally in .codetape/. Source code is never modified — only documentation files are touched. The skill uses progressive disclosure: SKILL.md is the entry point, with references/ and templates/ loaded on demand.
.codetape/
├── traces/ # Semantic change records (timestamped JSON)
├── config.json # Sync targets, strategies, language settings
├── references/ # Deep-dive docs loaded on demand
└── templates/ # Output templates for each doc type
Zero runtime dependencies. No background processes. No watchers. Runs only when you invoke a command.
Recommended — per-project setup:
npx codetape initGlobal — install for all projects:
npx codetape install -gManual — copy the skill directly:
cp -r skill/ .claude/skills/codetape/| Strategy | Behaviour | Used For |
|---|---|---|
append-managed |
Only modifies <!-- codetape --> markers |
README, CLAUDE.md |
prepend |
Adds new entries at top | CHANGELOG |
overwrite |
Regenerates entirely from traces | ARCHITECTURE.md |
merge |
Preserves human-written sections | Mixed docs |
TypeScript/JavaScript, Python, Rust, Go, Swift, GDScript (Godot).
- Phase 1: Core skill + CLI
- Phase 1.5: Session briefing, drift detection, CLAUDE.md auto-sync
- Phase 2: Decision memory, trace archival, quality self-check
- Phase 3: MCP server, cross-platform (Cursor, Codex, Windsurf), web dashboard
Issues and pull requests are welcome. Please open an issue first to discuss larger changes.
git clone https://github.com/888wing/codetape.git
cd codetape
node bin/codetape.js init # dogfood it