DGAT

Dependency Graph as a Tool

Point it at a codebase. Get a fully-described, LLM-annotated dependency graph — instantly.

---

What is this?

DGAT scans any codebase, uses a locally-hosted LLM to write natural-language descriptions for every file and every import relationship, then serves it all through an interactive three-panel UI. Think of it as a self-generating architectural map — no config files, no annotations, no manual work.

---

Architecture

flowchart TD
    subgraph CLI ["dgat (C++17 backend)"]
        A[Walk directory tree] --> B[Fingerprint files\nXXH3-128]
        B --> C[Parse imports\ntree-sitter + regex fallback]
        C --> D[Describe files & edges\nvLLM HTTP API]
        D --> E[Build dependency graph\nDepNode + DepEdge]
        E --> F[Synthesise blueprint\ndgat_blueprint.md]
        F --> G[Persist state\nfile_tree.json · dep_graph.json]
    end

    subgraph Server ["--backend mode"]
        G --> H[HTTP server :8090\ncpp-httplib]
        H --> |GET /api/tree| I[File tree JSON]
        H --> |GET /api/dep-graph| J[Dep graph JSON]
        H --> |GET /api/blueprint| K[Blueprint markdown]
    end

    subgraph UI ["Next.js frontend"]
        I --> L[Explorer panel\nfile tree]
        K --> M[Blueprint tab\nrendered markdown]
        J --> N[Graph tab\nSigma.js WebGL]
        L & M & N --> O[Inspector panel\nnode · edge · dep details]
    end

    subgraph LLM ["Local vLLM"]
        D <--> P[Qwen/Qwen3.5-2B\nor any OpenAI-compat endpoint]
    end

Loading

---

Demo

1. Start the vLLM server

Before running DGAT, bring up a local vLLM instance. DGAT uses it to generate descriptions for every file and dependency edge.

---

2. Run the scan

Point DGAT at your project. It walks the file tree, fingerprints every file, sends them to vLLM, and builds the dependency graph.

Dependency extraction runs in parallel — here's the tail end where import relationships are resolved and edges are formed:

---

3. Open the UI

Start the backend server and the frontend. Three panels: file explorer on the left, blueprint/graph in the middle, inspector on the right.

Blueprint tab — a synthesised architectural overview of the whole project, generated bottom-up from individual file descriptions:

File inspector — click any file in the explorer to see its description, dependencies, and metadata:

Select a file like dep_graph.json to see its role in the project explained inline:

---

4. Explore the dependency graph

Switch to the Graph tab for an interactive WebGL view of all import relationships. Node size reflects connectivity.

Single node selected — click any node to see a full LLM-generated analysis of that file, plus its outgoing/incoming edges at the bottom:

Two nodes selected — click a second node to inspect the direct edge between them: the import statement, and a plain-English explanation of why one depends on the other:

---

Features

• Multi-language import extraction — TypeScript, JavaScript, Python, C/C++, Go, Java, Rust, C#, Ruby, PHP, Bash, and more. Tree-sitter grammars for precision, regex fallback for everything else.
• LLM-annotated graph — every file node and every dependency edge gets a concise description generated by a local model. No cloud, no API keys.
• Project blueprint — a synthesised dgat_blueprint.md built bottom-up from all file descriptions.
• Incremental updates — dgat update re-describes only files whose XXH3 fingerprint changed.
• Static export — embed the entire graph into a single self-contained HTML file. Share with anyone, no server required.
• Live UI — auto-refreshes every 30 s. Three-panel layout with file explorer, blueprint/graph tabs, and an inspector.

---

Getting started

Prerequisites

• C++17 compiler (GCC 11+ / Clang 14+)
• CMake 3.16+
• A running vLLM instance (or any OpenAI-compatible endpoint) on localhost
• Node.js 18+ / Bun (for the frontend)

Build & install

git clone https://github.com/yourname/dgat
cd dgat
cmake -B build && cmake --build build -j$(nproc)

# or use the install script
bash install.sh

Run

# full scan — builds tree, descriptions, dep graph, blueprint
./build/dgat /path/to/your/project

# start the API server (no re-scan)
./build/dgat --backend

# start the frontend
cd frontend && bun dev
# open http://localhost:3000

---

CLI reference

Command	Description
dgat [path]	Full scan of the target directory
dgat --backend	Load saved state, start API server
dgat update	Incremental re-scan (changed files only)
--port <n>	Override API server port (default: 8090)

---

Tech stack

Layer	Tech
Backend	C++17 · cpp-httplib · nlohmann/json · inja · xxHash
Parsing	tree-sitter (C, C++, Python, TS, Go, Java, Rust, …)
LLM	vLLM · Qwen3.5-2B (any OpenAI-compat endpoint)
Frontend	Next.js 14 · TypeScript · Tailwind CSS · Sigma.js · shadcn/ui

---

gonna scavenge some popular repos, and see how my thing behaves with it, and use my populated context for something useful.