PyCodeKG — A Knowledge Graph for Python Codebases

PyCodeKG turns a Python codebase into a deterministic, queryable knowledge graph — and uses it to produce architectural analyses you can act on, with or without an LLM in the loop.

It walks the AST of every module, class, function, and method in your repo, extracts the typed relationships that actually hold the code together (CONTAINS, CALLS, IMPORTS, INHERITS, RESOLVES_TO), and stores the result in SQLite. A sqlite-vec vector index sits alongside the graph so that "authentication flow" and "verify_jwt" both find the right place to start exploring. From there you can rank functions by structural importance, trace fan-in across import aliases, detect circular imports and dead code, render the call graph in 3D, grow the repo as a tree you can put on a holographic display, snapshot metrics for diffing across releases, or hand the whole thing to Claude over MCP.

The original motivation was simple: produce thorough, defensible analyses of Python codebases that don't depend on inference. Every result is computed from the AST and the graph — no model is asked to guess. When an LLM is present, it consumes the same grounded output as a structured context pack, and the hallucinations that plague "embed-the-repo" tools largely disappear.

Everything runs on your laptop. No cloud APIs, no quotas, no source code leaving the machine.

Technical Paper (PDF) · Author: Eric G. Suchanek, PhD — Flux-Frontiers, Liberty TWP, OH

Sister projects

PyCodeKG is one of ten knowledge graphs sharing the same hybrid semantic-plus-structural design — code, prose, journals, filesystems, agent memory, metabolic pathways — federated by KGRAG so a single query can span all of them. The full list is in docs/SISTER_PROJECTS.md.

Two ways to use it

PyCodeKG is designed to be useful at both ends — as a standalone command-line analysis tool, and as a structured context layer for AI agents.

1. Standalone — pycodekg analyze

This single command is the actual reason I wrote this module. I got tired of my LLM doing grep on my src/ in order to get a good analysis of the codebase. This command runs fifteen analysis passes in seconds and the output drops straight into your favorite coding LLM for targeted improvement suggestions. I used this process to iteratively improve PyCodeKG and several of the analysis passes were added as a result.

The analysis/ has files *analysis.md - I ran PyCodeKG against numpy and matplotlib for fun.

pycodekg init                                        # downloads embedder
pycodekg build --repo .                              # one-time index
pycodekg analyze.                                    # the full report

Note The first-time run can take some time to warm up. Be patient.

The 15-phase pipeline runs in sequence and surfaces several important metrics:

Every finding maps to a file and line number — no hallucinated signatures, no probabilistic guesses. The Markdown output is LLM-ready; the JSON snapshot is CI-gate-ready. Reach for analyze before any non-trivial refactor, at every release, and whenever you inherit an unfamiliar codebase. Full reference: docs/Analyze.md.

pycodekg analyze --quiet --json ~/.claude/pycodekg_analysis_latest.json
jq '.docstring_coverage.total' ~/.claude/pycodekg_analysis_latest.json

2. Agentic — MCP server for grounded AI workflows

Run pycodekg mcp and Claude (or any MCP-aware client) gets nineteen tools backed by the same graph: graph_stats, query_codebase, pack_snippets, get_node, list_nodes, callers, explain, centrality, bridge_centrality, framework_nodes, analyze_repo, snapshot_list / show / diff, and more. Setup for Claude Code, Claude Desktop, Kilo Code, Copilot, and Cline is a single line — see docs/MCP.md and docs/INSTALLATION.md.

The agent benefit isn't subtle. Tools like pack_snippets return actual source with line numbers and surrounding context; callers returns the real fan-in resolved across import aliases, not a regex's best guess. The agent stops fabricating function signatures and starts citing them. Multi-step workflows — "find the auth path, list its callers, summarize what changes if I rename it" — collapse from dozens of greps and file reads into a handful of source-grounded calls.

Independent assessments tend to put it the same way:

"PyCodeKG compresses a multi-step workflow — semantic search, graph expansion, caller tracing, snippet retrieval, and architectural summarization — into a small set of tools that are fast to invoke and easy to chain. In practice, it let me move from broad orientation to intent-driven discovery and then to structural validation without dropping down into manual grep or repeated file reads." — GPT-5 (via Cline)

"What sets it apart from 'search the repo with embeddings' tools is the structural layer… Verdict: 4.5/5 — recommend without reservation for any non-trivial Python codebase." — Claude Opus 4.7

"PyCodeKG is dramatically more effective than traditional grep/file-reading workflows. Unique value: hybrid search combining natural-language intent with precise structural relationships." — Claude Haiku 4.5

Full reports in assessments/.

Get started in 60 seconds

Requirements: Python ≥ 3.12, < 3.14

pip install 'pycode-kg[viz,viz3d]'        # base + Streamlit + 3-D viewer

cd /path/to/your/repo
pycodekg init --repo .                    # download model, build graph, install hooks, snapshot
pycodekg analyze .                        # the architectural report

That's the recommended path. Variants (minimal install, MCP-only, contributor setup) are in docs/INSTALLATION.md. Every capability is a subcommand of the single pycodekg entry point — pycodekg analyze, pycodekg build, pycodekg mcp — so that one name is all a Makefile or an MCP config needs.

How retrieval works

Search is hybrid by design. A query like "authentication flow" runs in two phases:

1. Vector phase — the query is embedded with a local sentence-transformer (cached after first download) and sqlite-vec returns the k closest functions, classes, and modules by exact cosine similarity.
2. Graph expansion phase — each seed hit is expanded hop BFS steps along the typed edges (CONTAINS, CALLS, IMPORTS, INHERITS, RESOLVES_TO) so call chains and module relationships surface alongside the names that matched.

Structure is treated as ground truth; the embeddings are strictly an acceleration layer. When the graph and the vector index disagree, the graph wins. This is why fan-in lookups are accurate even for same-named symbols across modules — RESOLVES_TO edges bridge call sites through their import aliases, and callers() does a two-phase reverse traversal that grep simply cannot replicate.

The graph is built around four node kinds (module, class, function, method) and five edge relations. Schema and edge semantics are documented in docs/CHEATSHEET.md.

The repo as a tree

pycodekg viz3d --layout organic grows the graph instead of plotting it. A skeleton is grown toward the code by space colonization, so the shape is the data rather than decoration:

• The trunk is the repository; its radius follows the pipe model, so it reports how much code the tree carries.
• Each limb is a module. Its length is how many definitions the module holds — the biggest module reaches furthest — and its thickness follows from the tips it carries.
• Each leaf is a class, function, or method, tinted by kind.

pycodekg quilt renders that tree for a Looking Glass light-field display: a multi-view quilt the panel fuses into real depth, with --cast to send it straight to the device. Every render prints its disparity budget first, so you know whether the display will fuse the views or ghost them. Details in docs/VIZ3D.md.

The growth engine is shared across the KG fleet (kgmodule-utils) and the light-field output is quiltwright; what lives here is only the mapping from code to wood.

What you can do with it

If you only read one doc after this one, read docs/Analyze.md — that's where most of the day-to-day value lives.

Architecture

src/pycode_kg/
├── visitor.py                       # AST extraction (three-pass: structure, calls, dataflow)
├── graph.py                         # GraphBuilder: file discovery + dispatch
├── store.py                         # SQLite persistence + canonical edges
├── index.py                         # sqlite-vec semantic index
├── pycodekg.py                      # Public façade
├── pycodekg_thorough_analysis.py    # `analyze` engine
├── architecture.py                  # `architecture` description generator
├── ranking/                         # PageRank, bridge centrality, framework nodes
├── snapshots.py                     # Temporal metric snapshots
├── analysis/                        # Coupling, cycles, orphans, hotspots
├── cli/                             # `pycodekg` subcommands
├── mcp_server.py                    # MCP server (nineteen tools)
├── app.py                           # Streamlit web app
├── viz3d.py / layout3d.py           # PyVista/PyQt5 3-D viewer
├── scene3d.py                       # Organic tree: repo → trunk, module → limb
└── viz3d_timeline.py                # Metric history timeline

The MCP server, the CLI, and the Streamlit app are thin wrappers over the same store + index + ranking core — there is exactly one code path for each capability. The latest architectural deep-dive is in docs/analysis_v0.27.1.md, produced (of course) by pycodekg analyze against this very repo.

Documentation map

Citation

If you use PyCodeKG in your research or project, please cite it:

Suchanek, E. G. (2026). PyCodeKG: A Knowledge Graph for Python Codebases (Version 0.27.1) [Software]. Flux-Frontiers. https://doi.org/10.5281/zenodo.19737993

@software{suchanek_pycode_kg,
  author    = {Suchanek, Eric G.},
  title     = {{PyCodeKG}: A Knowledge Graph for Python Codebases},
  version   = {0.27.1},
  year      = {2026},
  publisher = {Flux-Frontiers},
  url       = {https://github.com/Flux-Frontiers/pycode_kg},
  doi       = {10.5281/zenodo.19737993},
}

License

Elastic License 2.0 — free for non-commercial and internal use; commercial redistribution or hosting requires a license from Flux-Frontiers.

Support & acknowledgments

• Issues — GitHub Issues
• Sister projects — the rest of the KGRAG family
• sqlite-vec, sentence-transformers, PyVista, Streamlit, and FastMCP for the foundations

Built for Python developers and AI agents that work alongside them — egs