Codebase Graph — AST Intelligence Across 158 Languages

Kodelyth ECC integrates DeusData/codebase-memory-mcp — a single static binary that indexes any codebase into a tree-sitter AST knowledge graph with Hybrid LSP semantic type resolution.

Structural queries like "who calls X" or "what does the auth flow look like" now cost ~3,400 tokens instead of ~412,000 tokens via file-by-file grep. 99% token reduction.

Their binary, their curl script, their MIT license. ECC installs, wires, and surfaces it. No fork, no code copy, no npm dependency.

What you get

AST-parsed graph — 158 languages via tree-sitter grammars vendored into the binary
Hybrid LSP — semantic type resolution for Python, TypeScript / JavaScript / JSX / TSX, PHP, C#, Go, C, C++, Java, Kotlin, and Rust (parameter binding, return-type inference, generic substitution, JSX component dispatch, JSDoc inference)
Cross-service linking — HTTP routes, gRPC, GraphQL, tRPC, EventEmitter channels
14 MCP tools — search_graph, trace_path, get_architecture, manage_adr, semantic_query, detect_changes, search_code, dead code detection, Cypher queries, and more
Zero infrastructure — SQLite-backed, persists to ~/.cache/codebase-memory-mcp/
Local only — your code never leaves your machine

Auto-install via ECC

Add --codebase-graph to your install:

npm i -g kodelyth-ecc
kodelythecc --target claude-code --codebase-graph

Or after ECC is installed:

kodelythecc codebase install

Both flows:

Detect if codebase-memory-mcp is on your PATH (idempotent — reuses existing install)
If not, install via their official curl script (~/.local/bin/codebase-memory-mcp)
Run their install command which auto-registers MCP entries in every detected AI-coding agent (~/.claude.json, Codex CLI, Gemini CLI, Zed, OpenCode, Antigravity, Aider, KiloCode, VS Code, OpenClaw, Kiro)

First index

Open a project in your AI tool. Say:

Index this project

The MCP tool index_repository builds the graph. Django-scale takes ~6 seconds. Linux kernel (28M LOC, 75K files) takes 3 minutes.

Verify:

kodelythecc codebase status

codebase-memory-mcp: codebase-memory-mcp 0.8.1
  indexed projects: 8
  cache dir:        /Users/you/.cache/codebase-memory-mcp
  next: open a project in your AI tool and say "Index this project"

Query the graph from the CLI

kodelythecc codebase query search_graph '{"name_pattern": ".*Handler.*"}'
kodelythecc codebase query trace_path   '{"function_name": "main", "direction": "outbound"}'
kodelythecc codebase query get_architecture '{}'
kodelythecc codebase query detect_changes '{}'

All queries run locally. No LLM cost. Results are structured JSON your AI tool can consume in a single MCP call.

CLI reference

kodelythecc codebase install                              # install binary + auto-register agents
kodelythecc codebase status [--json]                      # binary version + indexed projects + cache dir
kodelythecc codebase register                             # re-run their auto-configure step for installed agents
kodelythecc codebase query <cli-cmd> [json]               # pass-through to `codebase-memory-mcp cli`
kodelythecc codebase --help                               # focused help

Graph edge types (selected)

CALLS — function-to-function
IMPORTS — module dependency
DEFINES — file defines a symbol
IMPLEMENTS — interface/trait implementation
INHERITS — class inheritance
HTTP_CALLS, ASYNC_CALLS — cross-service
EMITS, LISTENS_ON — pub-sub channels
DATA_FLOWS — arg-to-param mapping with field access chains
SIMILAR_TO — MinHash + LSH near-clone detection
SEMANTICALLY_RELATED — vocabulary-mismatch, same-language, score ≥ 0.80

Common queries (via your AI tool)

Once indexed, ask your AI tool things like:

"Who calls ProcessOrder?"
"What's the impact of changing AuthMiddleware?"
"Show me the architecture of this repo"
"Find dead code — functions with zero callers"
"Which HTTP routes touch the users table?"

The AI translates natural language to MCP calls behind the scenes. You never write Cypher unless you want to.

Dashboard view

kodelythecc dashboard → Codebase tab shows:

Binary version
Indexed project count (real, from list_projects)
Graph nodes / edges
Language distribution
Entry points (top 5)
Project list with per-project node + edge counts

When no active session graph exists, dashboard shows the indexed project list with node/edge counts. When you open a project in your AI tool, its architecture snapshot fills in.

All numbers come from live queries — zero hardcoded values.

Performance

Benchmarked on Apple M3 Pro (from their docs):

Operation	Time
Linux kernel full index	3 min (28M LOC, 75K files → 4.81M nodes, 7.72M edges)
Linux kernel fast index	1m 12s (1.88M nodes)
Django full index	~6s (49K nodes, 196K edges)
Cypher query	<1ms
Name search (regex)	<10ms
Dead code detection	~150ms
Trace call path (depth=5)	<10ms

RAM-first pipeline: all indexing runs in memory with LZ4 compression and in-memory SQLite. Memory is released after indexing completes.

Attribution

License: MIT
Maintainer: DeusData/codebase-memory-mcp
Research paper: Codebase-Memory: Tree-Sitter-Based Knowledge Graphs for LLM Code Exploration via MCP
ECC's wrapper: scripts/codebase/index.js — thin, no fork
Fallback: If upstream disappears, ECC will fork + vendor. MIT permits.

Codebase Graph — AST Code Intelligence Across 158 Languages