code2graph

Turn a source code repository into a queryable knowledge graph — no LLM required.

code2graph statically extracts the structure of a codebase — files, modules, functions, classes, calls, dependencies, schemas, infrastructure — as a typed graph of nodes and edges. Rank the most relevant nodes for any query with Personalized PageRank and pass focused context to any LLM.

Pure Python. No LLM dependency. Bring your own model.

---

Quick start

pip install codes2graph

# Extract full graph from a repo
codes2graph /path/to/repo --graph all --output repo.graph.json

# Python call graph only
codes2graph /path/to/repo --graph call --output calls.graph.json

# With actionable summary
codes2graph /path/to/repo --graph all \
  --output repo.graph.json \
  --summary-output repo.summary.json

from code2graph import build_graph

graph = build_graph("/path/to/repo", graph_type="all")
# graph.nodes — list of Node objects
# graph.edges — list of Edge objects

---

Why graph-based code context?

Approach	What you lose
Dump entire codebase into prompt	Token budget, focus
Embed + search file chunks	Call relationships, module structure, dependency chains
code2graph	Nothing — relationships are explicit labeled edges

The graph knows that auth.login() calls db.query(), which imports connection_pool, which depends on config.DATABASE_URL. Flat file chunks don't.

---

Graph types

Type	What it extracts
folder	Repo, folder, file nodes with contains edges
call	Functions/methods with calls and defines edges (Python, JS, TS)
entity	Classes, functions, constants with defines and imports edges
schema	Database tables, columns, foreign keys (SQL, ORM models)
workflow	CI/CD pipelines, GitHub Actions, Makefile targets
infra	Dockerfiles, docker-compose, Terraform, Kubernetes manifests
security	Hardcoded secrets patterns, dangerous function calls, exposed endpoints
web	React/Vue components, routes, API endpoints
android	Activities, services, permissions from AndroidManifest.xml
decision	ADR-style architecture decisions
all	Merged graph from all applicable extractors

codes2graph /path/to/repo --graph call   --output call.graph.json
codes2graph /path/to/repo --graph schema --output schema.graph.json
codes2graph /path/to/repo --graph infra  --output infra.graph.json
codes2graph /path/to/repo --graph all    --output full.graph.json

---

Installation

pip install codes2graph

No extra dependencies required — all graph types work with the standard install.

---

Python API

Build a graph

from code2graph import build_graph, Graph, Node, Edge

# Full graph
graph: Graph = build_graph("/path/to/repo", graph_type="all")

# Specific type
call_graph = build_graph("/path/to/repo", graph_type="call")
schema_graph = build_graph("/path/to/repo", graph_type="schema")

Inspect results

print(f"{len(graph.nodes)} nodes, {len(graph.edges)} edges")

# Filter by kind
functions = [n for n in graph.nodes if n.attributes.get("kind") == "function"]
calls = [e for e in graph.edges if e.label == "calls"]

Export

import json

# To dict
d = {"nodes": [vars(n) for n in graph.nodes], "edges": [vars(e) for e in graph.edges]}
json.dump(d, open("graph.json", "w"), indent=2)

---

Graph output format

{
  "nodes": [
    {
      "id": "function:auth.login",
      "label": "login",
      "attributes": {
        "kind": "function",
        "module": "auth",
        "file": "src/auth.py",
        "line": 42
      },
      "content": "def login(username, password): ..."
    }
  ],
  "edges": [
    {
      "id": "edge:auth.login:calls:db.query",
      "from": "function:auth.login",
      "to": "function:db.query",
      "label": "calls"
    }
  ],
  "current_node_id": "repo"
}

---

CLI reference

codes2graph <repo> [options]

Arguments:
  repo                    Path to the repository root

Options:
  --graph TYPE            Graph type: folder, call, entity, schema, workflow,
                          infra, security, web, android, decision, all
                          (default: all)
  --output PATH           Write graph JSON to this file (default: stdout)
  --pretty                Pretty-print JSON output
  --summary-output PATH   Write graph summary JSON (entrypoints, fan-in/out nodes)
  --update-existing PATH  Update an existing graph JSON in place
  --update-summary-output PATH
                          Write update diff summary JSON
  -h, --help              Show help

Update mode

Rebuild a graph from the current repository state while preserving stable node IDs and custom attributes added outside code2graph:

codes2graph /path/to/repo --graph all \
  --update-existing repo.graph.json \
  --update-summary-output repo.update.json

Update mode removes stale nodes/edges for deleted or changed code, adds new nodes/edges, and keeps stable IDs for nodes that haven't changed. Custom attributes on existing nodes are preserved.

---

Use cases

• Code review — extract call graph before/after a PR to see what changed structurally
• LLM code assistance — pass ranked subgraph as context instead of dumping whole files
• Dependency analysis — find all callers of a function, all modules depending on a service
• Security audit — detect hardcoded secrets, dangerous API patterns, exposed endpoints
• Architecture docs — extract infra + schema + decision graphs for living documentation
• Onboarding — give a new developer a ranked subgraph of the most important entry points

---

Design principles

• Pure Python — no LLM, no cloud, no database required
• Deterministic — same repository state always produces the same graph
• Static analysis only — no code execution, safe to run on any codebase
• Works with any model — output is plain JSON; pass to GPT-4, Claude, Llama, or any other model
• Companion to docs2graph — same node/edge schema, combine code and documentation graphs

---

Related projects

Package	What it does
docs2graph	Documents → knowledge graph (same node/edge schema)
graph2sql	Graph-based schema analysis for text-to-SQL

---

Contributing

See CONTRIBUTING.md.

git clone https://github.com/jw-open/code2graph
cd code2graph
pip install -e ".[dev]"
pytest tests/ -v

---

License

Apache-2.0 — see LICENSE