$0.60 per session, just on orientation. Here is what my AI agent was doing before writing any code.

GitHub: github.com/glincker/stacklit | npm: npx stacklit init

I have been counting.

Every time I open Claude Code on a repo, it reads 8 to 12 files before doing anything useful. read_file src/auth/service.ts. read_file src/db/connection.ts. One after another. Just orientation. Figuring out what goes where.

On a medium repo, that costs about 400,000 tokens. The agent burns through its context window before writing a single line.

Second session? Same files. Same cost. There is no memory. No map. No way for one agent's exploration to benefit the next.

The file nobody made yet#

.gitignore tells git what to skip. package.json tells npm what to install. But nothing tells AI agents how your code is organized.

So I made it.

npx stacklit init

One command scans your codebase and generates stacklit.json:

{
  "modules": {
    "src/auth": {
      "purpose": "Authentication and session management",
      "files": 8,
      "lines": 1200,
      "exports": ["AuthProvider", "useSession()", "loginAction()"],
      "depends_on": ["src/db", "src/config"],
      "activity": "high"
    }
  },
  "hints": {
    "add_feature": "Create handler in src/api/, add route in src/index.ts",
    "test_command": "npm test"
  }
}

Commit it. Now every agent that touches your repo already knows the structure.

Real numbers, not marketing#

I ran Stacklit on four popular open source projects and counted the tokens:

FastAPI. 108,000 lines of Python. The entire project structure fits in 4,142 tokens. That is less than what an agent burns reading one large file.

What you actually get#

Three files from one command:

stacklit.json -- The index. Modules, dependencies, exports, type definitions, git activity, framework detection, hints. Committed to git. Any agent reads it.

DEPENDENCIES.md -- A Mermaid dependency diagram. Renders directly on GitHub. Click the file in your repo and see the graph.

stacklit.html -- An interactive visual map. Four views: graph, tree, table, flow. Opens in a browser, works without a server. Gitignored because you regenerate it locally.

11 languages via tree-sitter#

Stacklit does not use regex to guess at your code. It uses tree-sitter, the same parser that powers syntax highlighting in VS Code, Neovim, and GitHub. Full AST parsing.

Plus generic fallback for anything else (line count and language detection).

It also has an MCP server#

stacklit serve

Seven tools. get_overview, get_module, find_module, list_modules, get_dependencies, get_hot_files, get_hints.

Add three lines to your Claude Desktop or Cursor config:

{
  "mcpServers": {
    "stacklit": {
      "command": "stacklit",
      "args": ["serve"]
    }
  }
}

Now your agent calls get_overview() once instead of reading a dozen files to build context.

How it compares#

Install#

npx stacklit init          # npm
go install github.com/glincker/stacklit/cmd/stacklit@latest  # go

Single binary. Zero dependencies. Runs in under 100ms on most repos.

The whole thing is MIT licensed and open source.

GitHub: github.com/glincker/stacklit

If this is useful, a star helps other developers find it.