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.

Available as a Python package — install with pip install dgat and you're ready to go.

---

Quick start

1. Install DGAT

pip install dgat
# requires Python 3.11+

2. Configure your LLM provider

# Interactive setup
dgat config init

# Or configure manually (vLLM example)
dgat config set providers.vllm.endpoint http://localhost:8000
dgat config set providers.vllm.model Qwen/Qwen3.5-2B

Supported providers: vLLM, Ollama, OpenAI, Anthropic, OpenRouter (any OpenAI-compatible endpoint)

3. Start your LLM server

# vLLM example
vllm serve Qwen/Qwen3.5-2B --port 8000

4. Run on your project

dgat scan /path/to/your/project
# outputs: file_tree.json, dep_graph.json, dgat_blueprint.md

---

CLI commands

Command	Description
dgat scan [path]	Full codebase scan — builds tree, descriptions, dep graph, blueprint
dgat update [path]	Incremental re-scan (changed files only)
dgat search <query>	Search files by name or description
dgat describe <rel_path>	Get LLM-generated description for a specific file
dgat deps <rel_path>	Show files that the given file depends on
dgat dependents <rel_path>	Show files that depend on the given file
dgat blueprint	Get the architectural blueprint (dgat_blueprint.md)
dgat mcp	Start MCP server (stdio mode)
dgat mcp --http	Start MCP server (HTTP mode)
dgat backend	Start API backend server
dgat config show	Show current configuration
dgat config set <key> <value>	Set a configuration value
dgat config test	Test if the provider API is working

---

Python API

Import DGAT directly in your Python code:

from dgat import run_scan, run_update
from dgat import FileTree, DepGraph

# or from dgat.scanner import search_files

---

MCP server

Use DGAT as a tool in AI coding agents via the Model Context Protocol:

dgat mcp              # stdio mode (for local agents)
dgat mcp --http       # HTTP mode (for remote agents)

Available tools: scan, update, describe_file, get_dependencies, get_dependents, get_blueprint, search_files, get_file_tree, get_dep_graph

---

Architecture

flowchart TD
    subgraph CLI ["dgat CLI (Python wrapper)"]
        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 MCP ["MCP Server"]
        H[AI Coding Agent] --> I[DGAT MCP tools]
        I --> J[scan, update, search,\ndescribe, deps, etc.]
    end

    subgraph UI ["Renderwise Forge UI"]
        G --> K[HTTP server :8090]
        K --> |GET /api/tree| L[File tree JSON]
        K --> |GET /api/dep-graph| M[Dep graph JSON]
        K --> |GET /api/blueprint| N[Blueprint markdown]
        L --> O[Explorer panel\nfile tree]
        N --> P[Blueprint tab\nrendered markdown]
        M --> Q[Graph tab\nSigma.js WebGL]
        O & P & Q --> R[Inspector panel\nnode · edge · dep details]
    end

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

Loading

---

Features

• Multi-language import extraction — TypeScript, JavaScript, Python, C/C++, Go, Java, Rust, C#, Ruby, PHP, CUDA, 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.
• MCP integration — use DGAT as a tool in AI coding agents via the Model Context Protocol.
• 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.

---

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:

---

Tech stack

Layer	Tech
Package	Python 3.11+ · pip-installable
CLI	Click · Rich (terminal output)
Parsing	tree-sitter (C, C++, Python, TS, Go, Java, Rust, …)
LLM	vLLM · Qwen3.5-2B (any OpenAI-compat endpoint)
MCP	JSON-RPC 2.0 (stdio + HTTP)
Frontend	React · TypeScript · Tailwind CSS · Sigma.js · shadcn/ui

---

Example output

The examples/dgat-self/ folder contains the output DGAT produced when pointed at its own source tree — a blueprint, file tree, and dependency graph all generated by Qwen3.5-2B running locally via vLLM. Browse it to get a feel for the output format without running a scan yourself.

---

.dgatignore

DGAT respects a .dgatignore file in the root of the scanned project. It works like .gitignore — one glob pattern per line — and tells DGAT which files and directories to skip during description and graph-building passes (files that match are still shown in the tree but their LLM descriptions and dependency edges are suppressed).

# .dgatignore example
node_modules/
*.lock
vendor/

Files already covered by .gitignore are automatically excluded from LLM processing regardless.

---

Docs

• Overview — how DGAT works end to end
• File Tree — TreeNode structure, fields, and why each one exists
• Dependency Graph — DepNode, DepEdge, DepGraph, and how the graph is built
• Incremental Updates — how the diff mode works with XXH3 fingerprints
• Import Extraction — tree-sitter parsing, regex fallbacks, and path resolution
• Python Package — installing and using DGAT as a pip package

---

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