cognitx-codegraph 0.1.104

Code knowledge graph for Claude Code & AI coding agents — index TypeScript, Python, NestJS, FastAPI, React into Neo4j and query architecture in Cypher. Note: v0.2.0 is deprecated, use 0.1.x.

codegraph

Index a TypeScript / Python codebase into a Neo4j property graph, then query architecture in Cypher. Designed as always-on context for Claude Code, Cursor, Codex, Aider, and other AI coding agents.

PyPI: cognitx-codegraph — install with pipx install cognitx-codegraph.

Recognises NestJS controllers / injectables / modules, React components and hooks, TypeORM entities, GraphQL operations, FastAPI / Flask / Django routes, SQLAlchemy models, plus generic Python classes / decorators. Loads typed nodes + edges into Neo4j with no LLM in the pipeline — everything is AST + deterministic resolution.

Quick start

pipx install cognitx-codegraph        # core; or `pipx install "cognitx-codegraph[python,mcp,watch,analyze]"` for all extras
cd /path/to/your/repo
codegraph init                         # scaffolds Neo4j docker-compose, slash commands, CI gate, .arch-policies.toml; runs first index
codegraph query "MATCH (c:Class) RETURN c.name LIMIT 5"
codegraph arch-check                   # exit 1 on architecture violations

codegraph init is the supported onboarding path — it asks 4-5 questions, writes everything you need, starts Neo4j, and runs the first index. See docs/init.md for the full flow and flags.

CLI

Command	Purpose
codegraph init	Scaffold codegraph into a repo (interactive). Auto-detects + reuses the shared codegraph-neo4j container (starts it if stopped, creates it if missing). Pre-flights Docker presence + version with OS-aware install / upgrade suggestions. --yes for non-interactive, --bolt-port / --http-port for custom Neo4j ports, --skip-docker, --skip-index.
codegraph index <repo>	Walk the source, parse with tree-sitter, write the graph. Flags: -p/--package (repeatable), --update (SHA256 incremental), --since <git-ref> (diff-based), --no-wipe, --no-export, --no-benchmark, --no-analyze, --skip-ownership, --max-files, --ignore-file, --json.
codegraph query <cypher>	Run a Cypher query. -n/--limit, --json.
codegraph arch-check	Run architecture-conformance policies. Auto-scopes to configured packages; --scope / --no-scope overrides. Exits 1 on violations, 2 on config errors.
codegraph validate	Sanity-check the loaded graph (counts, orphans, schema).
codegraph audit	Agent-driven extraction self-check. Launches claude / codex / gemini / aider / opencode / droid (or writes a cursor rules file as fallback) in headless + permission-bypass mode against the live graph; the agent flags places where codegraph claims to extract X but missed it on this repo. Output is codegraph-out/audit-report.md plus an optional gh issue create. Prompt templates protected by CODEOWNERS, a CI workflow, and a runtime SHA-256 lock. Flags: --agent, --list-agents, --print-prompt-only, --gh-issue, --bypass/--no-bypass, --unsafe, --timeout, --recompute-lock, --yes, --json. Full reference: docs/cli.md#audit.
codegraph wipe	MATCH (n) DETACH DELETE n.
codegraph stats	Quick node / edge counts.
codegraph export	Produce graph.html (interactive vis-network), graph.json, optional graph.graphml and graph.cypher. Runs after index unless --no-export.
codegraph benchmark	Token-reduction benchmark vs. raw source. Optional --min-reduction for CI gating.
codegraph report	Generate GRAPH_REPORT.md from Leiden community detection. Runs after index unless --no-analyze.
codegraph watch	File-system watcher; rebuilds on save. --debounce, -p. Requires the [watch] extra.
codegraph hook install / uninstall / status	Manage post-commit + post-checkout git hooks that re-index automatically.
codegraph install <platform>	Wire codegraph into an AI agent platform (writes rules file, registers MCP server, etc.). 14 platforms supported — see below. --all installs for every detected platform.
codegraph uninstall <platform>	Remove integration. Preserves shared rules sections (e.g. AGENTS.md) when other platforms still need them.
codegraph repl	Interactive Cypher REPL. Same as running codegraph with no args.

MCP server

Optional extra: pipx install "cognitx-codegraph[mcp]". Console script codegraph-mcp runs a stdio FastMCP server with 16 tools. Read-only by default; the two write tools (wipe_graph, reindex_file) require --allow-write.

Tool	Purpose
query_graph(cypher, limit=20)	Read-only Cypher (limit pushed into the query).
describe_schema()	Node labels, edge types, property keys.
list_packages()	All :Package nodes with file counts.
find_class(name_pattern, limit=50)	Substring search over :Class.name.
find_function(name_pattern, limit=50)	Substring search over :Function/:Method. Returns containing class for methods.
describe_function(name, file=None, limit=50)	Signature, file, callers/callees summary.
callers_of_class(class_name, file=None, limit=50)	Who instantiates / extends / injects a class.
calls_from(name, max_depth=2)	Outgoing CALLS graph from a method.
callers_of(name, max_depth=2)	Reverse CALLS graph.
endpoints_for_controller(controller_name)	NestJS / FastAPI / Flask / Django routes.
files_in_package(name, limit=50)	All files in a package.
hook_usage(hook_name, limit=50)	React components calling a hook.
gql_operation_callers(operation_name, …)	Who triggers a GraphQL query / mutation.
most_injected_services(limit=20)	Top NestJS DI providers.
describe_group(name_or_id, kind=None, limit=50)	Inspect :EdgeGroup hyperedges (protocol implementers, communities).
wipe_graph(confirm=False)	Destructive — requires --allow-write and confirm=True.
reindex_file(path, package=None)	Re-parse and upsert a single file. Requires --allow-write.

Plus 29 prompt templates auto-registered from queries.md (one per Cypher example block).

Schema

Nodes: Package, File, Class, Method, Function, Interface, Endpoint, Column, GraphQLOperation, Event, Atom, EnvVar, Route, External, EdgeGroup.

Edges (selected): IMPORTS, IMPORTS_SYMBOL, IMPORTS_EXTERNAL, DEFINES_CLASS, DEFINES_FUNC, DEFINES_IFACE, HAS_METHOD, HAS_COLUMN, EXPOSES, INJECTS, PROVIDES, EXPORTS_PROVIDER, EXTENDS, IMPLEMENTS, RENDERS, USES_HOOK, DECORATED_BY, CALLS, CALLS_ENDPOINT, HANDLES, HANDLES_EVENT, EMITS_EVENT, READS_ATOM, WRITES_ATOM, READS_ENV, RESOLVES, BELONGS_TO, MEMBER_OF, OWNED_BY, LAST_MODIFIED_BY, CONTRIBUTED_BY, TESTS, TESTS_CLASS.

Every relationship carries confidence (EXTRACTED / INFERRED / AMBIGUOUS) and confidence_score (0.0-1.0). Filter low-confidence edges out of strict checks. See docs/confidence.md.

:EdgeGroup nodes model N-ary "hyperedges" — used for protocol implementers (auto-emitted during indexing) and Leiden communities (codegraph report). Members link via MEMBER_OF. See docs/hyperedges.md.

Architecture conformance

codegraph arch-check runs five built-in policies plus any [[policies.custom]] Cypher you add to .arch-policies.toml:

1. import_cycles — file-level cycles of length 2-6.
2. cross_package — forbidden directional imports between packages.
3. layer_bypass — Controllers reaching Repositories without a Service in between.
4. coupling_ceiling — files importing more than N other files (default 20).
5. orphan_detection — functions / classes / atoms / endpoints with no inbound references. Configurable exclude_prefixes / exclude_names for test framework conventions.

Suppress individual violations with [[suppress]] entries (each requires a reason). Stale suppressions are flagged. CI integration: .github/workflows/arch-check.yml (scaffolded by codegraph init). Full reference in docs/arch-policies.md.

AI platform integrations

codegraph install <platform> writes the appropriate rules file (e.g. CLAUDE.md, AGENTS.md, GEMINI.md, .cursor/rules/codegraph.mdc, .github/copilot-instructions.md), registers the MCP server, and tracks the install in .codegraph/platforms.json so partial uninstalls preserve shared sections.

Platform	Subcommand	Rules file
Claude Code	claude	CLAUDE.md
Codex	codex	AGENTS.md
OpenCode	opencode	AGENTS.md
Cursor	cursor	.cursor/rules/codegraph.mdc
Gemini CLI	gemini	GEMINI.md
GitHub Copilot CLI	copilot	(MCP only)
VS Code Copilot Chat	vscode	.github/copilot-instructions.md
Aider	aider	AGENTS.md
OpenClaw	claw	AGENTS.md
Factory Droid	droid	AGENTS.md
Trae	trae	AGENTS.md
Kiro IDE	kiro	AGENTS.md
Google Antigravity	antigravity	AGENTS.md
Hermes	hermes	AGENTS.md

Templates resolve $NEO4J_BOLT_PORT, $NEO4J_HTTP_PORT, $PACKAGE_PATHS_FLAGS, $CONTAINER_NAME, $DEFAULT_PACKAGE_PREFIX, $CROSS_PAIRS_TOML, $PIPX_VERSION to per-repo values.

Incremental indexing

Two modes:

• --update — SHA256 content-addressed cache in .codegraph-cache/ (auto-added to .gitignore by init). Files whose hash is unchanged are skipped. Stale entries are pruned on save. Cache is invalidated automatically on a package version bump.
• --since <git-ref> — re-index only files changed since the ref (HEAD~1, a tag, a SHA). Diffs the working tree, removes stale subgraphs for deleted/changed files, upserts the rest. Implies --no-wipe. Non-code files are filtered out before processing.

codegraph watch (debounced file-system watcher) and codegraph hook install (post-commit + post-checkout) both layer on top of --update.

File extras

Extra	Adds
[python]	tree-sitter-python — enables Python frontend.
[mcp]	mcp — enables codegraph-mcp stdio server.
[watch]	watchdog — enables codegraph watch.
[analyze]	networkx, graspologic — Leiden community detection for codegraph report.
[benchmark]	tiktoken — exact BPE counting for codegraph benchmark (otherwise chars/4 approximation).
[repl]	prompt_toolkit — history + tab-completion in REPL.
[all]	All of the above.

Configuration

codegraph.toml (or [tool.codegraph] in pyproject.toml) at the repo root:

[tool.codegraph]
packages       = ["packages/twenty-server", "packages/twenty-front"]
ignore_file    = ".codegraphignore"     # optional; auto-detected at repo root
neo4j_uri      = "bolt://localhost:7688"
neo4j_user     = "neo4j"
neo4j_password = "codegraph123"

.codegraphignore uses gitignore syntax with two extensions: @route:<glob> and @component:<glob> for path-based framework-feature suppression.

.arch-policies.toml configures arch-check — full schema in docs/arch-policies.md.

Documentation

• docs/cli.md — full per-command reference: every flag, --json shape, exit codes, examples.
• docs/mcp.md — full per-tool reference for the MCP server (17 tools).
• docs/schema.md — complete graph schema: 15 node types, 33 edge types, properties, indexing phases.
• docs/incremental.md — --update, --since, codegraph watch, codegraph hook install.
• docs/platforms.md — the 14 AI agent platform integrations.
• docs/init.md — codegraph init flow, port flags, prompts, troubleshooting.
• docs/arch-policies.md — policy reference, suppression syntax, custom-policy authoring.
• docs/confidence.md — edge-level confidence labels and scores.
• docs/hyperedges.md — :EdgeGroup model and the describe_group MCP tool.
• queries.md — canonical Cypher query catalogue (also auto-loaded as MCP prompt templates).
• ../CHANGELOG.md — version-by-version shipped-feature log.
• ../ROADMAP.md — session-to-session engineering handoff.

Development

git clone https://github.com/cognitx-leyton/codegraph
cd codegraph/codegraph
python3 -m venv .venv
.venv/bin/pip install -e ".[python,mcp,watch,analyze,benchmark,repl,test]"
docker compose up -d
.venv/bin/python -m pytest tests/ -q   # 800+ tests, ~15s

Tests target 100% passing, zero warnings. Conventional commits (feat(scope), fix(scope), docs(scope), test(scope)). Work lands on dev; main / release / hotfix are protected.

License: Apache-2.0.