Stacklit

Your codebase, in 1,500 tokens.

One command generates a committed JSON index that any AI agent can read. No server, no setup.

Quick start

npx stacklit init

Without stacklit: Agent reads 8-12 files to build context. ~400,000 tokens. 45 seconds before writing a line.

With stacklit: Agent reads stacklit.json. ~1,500 tokens. Knows every module, dependency, and convention instantly.

What you get

File	What it does	Committed?
`stacklit.json`	Machine-readable codebase index	Yes
`DEPENDENCIES.md`	Mermaid dependency diagram (renders on GitHub)	Yes
`stacklit.html`	Interactive visual map with 4 views	No (gitignored)

Visual map

MCP server

stacklit serve

Add to Claude Desktop or Cursor MCP config:

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

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

CLI

stacklit init                    # scan, generate all outputs, open HTML
stacklit generate                # regenerate from current source
stacklit view                    # regenerate HTML and open in browser
stacklit diff                    # check if index is stale
stacklit serve                   # start MCP server

Language support

Go (AST), TypeScript, JavaScript, Python, Rust, Java (regex), plus generic fallback for any language.

Compared to alternatives

	Stacklit	Repomix	Aider repo-map	Codebase Memory MCP
Output	~1,500 token JSON	500k+ token dump	Ephemeral text	SQLite DB
Committed to repo	Yes	Too large	No	No
Dependency graph	Yes	No	Yes	Yes
Visual output	HTML (4 views)	No	No	No
MCP server	Yes	No	No	Yes
Runtime needed	No	No	Yes (Python)	Yes (C server)

License

MIT

stacklit