llmcc

llmcc brings multi-depth architecture graphs for code understanding and generation.

Our goal is to build a multi-depth, tree-like context / architecture view of a codebase, so a coding agent can walk up (zoom out) for structure and intent, then walk down (zoom in) to the exact crates/modules/files/symbols it needs—getting a highly comprehensive understanding of any codebase (any programming language).

Supported Languages

Why multi-depth graphs?

People (and coding agents) need to understand systems from different dimensions. Sometimes you need the high-level architecture to see boundaries, ownership, and how subsystems connect; other times you need the low-level implementation details to make a safe, precise change. llmcc provides multiple depths so you can choose the right “distance” from the code for the task.

Walkthrough: Codex (midterm size multi-crate rust project)

This repo includes many examples under sample. Download and open them in browser for the best viewing experience.

Depth 1: crate graph

<img src="https://github.com/allenanswerzq/llmcc/raw/v0.2.65/sample/rust/codex-pagerank/depth_1_crate.svg" alt="Codex crate graph (depth 1)" style="max-width: 100%; height: 100%;" />

Depth 2: module graph

<img src="https://github.com/allenanswerzq/llmcc/raw/v0.2.65/sample/rust/codex-pagerank/depth_2_module.svg" alt="Codex module graph (depth 2)" style="max-width: 70%; height: auto;" />

Depth 3: file + symbol graph

<img src="https://github.com/allenanswerzq/llmcc/raw/v0.2.65/sample/rust/codex-pagerank/depth_3_file.svg" alt="Codex file and symbol graph (depth 3)" style="max-width: 100%; height: auto;" />

Here's a small portion of the graph at depth 3, showing the core abstraction layer for prompt handling in Codex. Developers and AI agents can quickly grasp the architecture by examining this view.

<img src="https://github.com/allenanswerzq/llmcc/raw/v0.2.65/doc/codex.jpg" alt="codex core logic" style="max-width: 100%; height: auto;" />

Performance

llmcc is designed to be very fast, and we will try to make it faster.

The repo contains benchmark for many famous project output here: sample/benchmark_results_16.md.

Excerpt (PageRank timing, depth=3, top-200):

Installation

npm / npx (Recommended)

The easiest way to use llmcc is via npm. No build required:

# Run directly without installing
npx llmcc-cli --help

# Or install globally
npm install -g llmcc-cli
llmcc --help

Cargo (Rust)

cargo install llmcc

From Source

git clone https://github.com/allenanswerzq/llmcc.git
cd llmcc
cargo build --release
./target/release/llmcc --help

CLI: generate graphs

Generate a crate-level graph for Codex (DOT to stdout):

llmcc \
    -d sample/repos/codex/codex-rs \
    --graph \
    --depth 1

Generate a PageRank-filtered file+symbol graph (write to a file):

llmcc \
    -d sample/repos/codex/codex-rs \
    --graph \
    --depth 3 \
    --pagerank-top-k 200 \
    -o /tmp/codex_depth3_pagerank.dot

Render DOT to SVG (requires Graphviz):

dot -Tsvg /tmp/codex_depth3_pagerank.dot -o /tmp/codex_depth3_pagerank.svg

For generating sample graphs:

just gen rust