Atlante

Turn code bloat into a refactor queue.

Your coding agent keeps adding code to the same handful of files. They grow past 2,000 lines, then past 4,000, then they stop fitting in context comfortably. Suddenly every refactor prompt fails, every edit takes three tries, and the agent starts inventing APIs that already exist two screens above.

Atlante scans your workspace locally, without any LLM, and gives you a ranked refactor queue: biggest files first, with imports, exports, fan-in, and fan-out beside them. Click Copy Agent Prompt, paste it into your coding agent, and start cleaning up the worst offenders.

Atlante gives your coding agent the map before you ask it to refactor.

Why this exists

AI-assisted coding amplifies volume. Structural judgment doesn't scale at the same speed:

• files grow past 3,000 lines before anyone notices, and then they're too big to load into any agent's context usefully
• the agent writes helpers that will never be called, and you can't spot them from inside the editor
• duplicate imports and cycles appear by accident, not by choice
• silent fan-in turns innocent modules into single points of failure

Linters look at one line. Type checkers look at types. Neither tells you which file to refactor next so your agent can keep working. Atlante does.

Three principles, in order:

1. Deterministic. Same input, same output. Always.
2. Local. Nothing leaves your machine. No accounts, no uploads, no telemetry.
3. Prescriptive. Not "here's a graph to interpret" but "these are the files that matter, ranked".

The workflow

1. You're deep in a Claude Code / Codex / Cursor / Aider session. Things feel slow. Edits start failing.
2. Run Analyze Workspace.
3. Open the table. It is already sorted by lines of code, largest first.
4. Click Copy Agent Prompt to generate a structured triage prompt with your top files and dependency signals.
5. Paste it into your coding agent. It picks the next low-risk extraction, explains why, and tells you what not to touch.
6. Resume shipping.

That's it. No AI in the loop, no cloud, no waiting. The analysis is serialized under .atlante/ as diff-friendly JSON, so you can commit it and watch your repo's structural health over time.

Agent workflow

Atlante gives your coding agent the map before you ask it to refactor.

Click Copy Agent Prompt, paste into your coding agent, ask for the next low-risk, high-value extraction.

What you get today

• Refactor queue. Every file ranked by lines of code, with imports, exports, fan-in, and fan-out. Sortable, searchable, filterable.
• Copy Agent Prompt. One-click structured prompt with your top files and dependency signals, ready to paste into Claude Code, Cursor, or any other coding agent. The prompt asks for the next low-risk extraction with explicit reasoning.
• Quick filters. Largest files, High fan-in, High fan-out. One click, see what stands out.
• File details drawer. Symbols, imports (resolved vs external vs unresolved), dependents, open-file action.
• Dependency Constellation. Interactive graph of internal dependencies, clustered by top-level folder, with focus mode and impact-based node sizing.
• Project library. Analyze multiple projects, switch between them from the sidebar.
• Persistent analysis. Results stored under .atlante/ in your workspace: stable JSON, diff-friendly, versionable.

Supported languages: JavaScript, TypeScript, Python via Tree-sitter AST; Java, C#, Go, Rust, Kotlin, Swift, Ruby, PHP via a generic/fallback parser.

What's next

The next layer is diagnostics: deterministic rules that turn the inventory into actionable refactor flags (god-file, giant-function, hub-file, file-cycle, dead-export, and more). The full plan lives in docs/todo/diagnostics.md.

See docs/vision.md for the longer framing.

Install

Install from the VS Code Marketplace, or search for Atlante in the Extensions panel inside VS Code.

git clone https://github.com/Alex31y/Atlante.git
cd Atlante
npm install
npm run build
npm run package

Use

1. Open any workspace.
2. Command palette -> Atlante (opens the panel).
3. Click Analyze Workspace in the sidebar.
4. Use the ranked table as your refactor queue, then flip to the graph when you need the dependency shape.

Commands

Per-feature docs: docs/reference/.

Settings

Development

npm install
npm run build        # full build
npm run watch        # rebuild on change
npm run lint         # tsc --noEmit
npm test             # vitest
npm run package      # produce .vsix

Repo layout:

src/
+-- extension/   # VS Code host (commands, providers, services, watchers)
+-- webview/     # React UI (inventory + graph)
+-- workers/     # Worker-thread Tree-sitter parsing
+-- shared/      # Types, constants, import resolver

Non-goals

• No chat, no LLM, no API keys, no embeddings.
• No cloud analysis, no telemetry on your code.
• Not a linter, not a type checker, not a runtime profiler.

Read more

• Why I built Atlante. The longer story, the principles, the why now.

License

MIT. See LICENSE.