CodeExplorer CodeExplorer
Title Problem & audience 1. Scan 2. Explore 3. AI synergy 4. Control & SIG Enterprise & banking Contact
1 / 16 ⬇ HTML
CodeExplorer CodeExplorer (ce)

Architectural intelligence & governance
for engineers and AI agents

Turn polyglot codebases into a local, deterministic 5-layer property graph—powering interactive architecture maps in VS Code, surgical MCP context for AI agents, and automated Semantic Invariant Gates in CI/CD.

1. Scan

⚡ Polyglot AST graph

Tree-sitter + config indexing into a single local SQLite file in seconds.

2. Explore

🪐 Living visuals

Interactive orbital topology, async event flows, and DB lineage inside VS Code.

3. Connect AI

🤖 MCP + CLI context

44× fewer tokens and 7.6× faster analysis for cloud & on-prem LLMs.

4. Control

🛡️ Invariant gates

Deterministic CI/CD architecture gates (< 2s) with Ratchet Baselines.

Viacheslav Mikhailov • code-explorer.app Slide 1 of 16 • Use ← / → arrow keys
● The core problem in modern software engineering

Codebases grow faster than
humans & AI can understand them

As systems scale across microservices, event buses, and databases, both engineers and AI coding agents lose structural visibility—leading to architectural drift.

Problem 1 • For engineers & leads

📉 Blindness & stale diagrams

  • Confluence & C4 diagrams rot the day after they are drawn.
  • IDE "Find Usages" breaks across REST APIs, Kafka/RabbitMQ topics, and SQL tables.
  • Tech Leads turn into human linters reviewing 40–50 Merge Requests a day.
Problem 2 • For AI agents & LLMs

🤖 Context blindness via grep

  • AI agents explore code via blind text search (grep) and reading hundreds of raw files.
  • Burns 40×+ extra tokens and overflows context windows.
  • LLMs hallucinate transitive call chains and break architectural boundaries.
Problem 3 • For governance

🔓 Zero structural control in CI

  • Static linters (Sonar, ESLint, Detekt) only check single-file syntax and style.
  • Nothing stops a PR from bypassing an integration layer or leaking DB entities into APIs.
CodeExplorer — Why standard tools & raw AI struggle at scale Slide 2 of 16
● How we solve it & for whom

One deterministic code graph.
Two purpose-built interfaces.

CodeExplorer parses your repository into a single local SQLite property graph (.codeexplorer/graph.db) and serves two distinct consumers:

For AI agents & models

🤖 MCP Server + CLI

Gives AI agents (Claude Code, Cursor, Windsurf, Copilot, or local on-prem LLMs) surgical architectural intelligence:

  • Local stdio MCP Server: 19 specialized graph tools (get_architecture_map, get_call_chain, analyze_code_impact, inspect_data_lineage).
  • Headless CLI & OpenCypher: Agents query exact multi-hop relationships in milliseconds instead of reading 50 files.
  • Result: 44× fewer tokens, 7.6× faster task completion, 100% recall.
For engineers, architects & tech leads

👨‍💻 VS Code visuals + CLI

Gives humans instant visual comprehension and automated architectural control directly inside their workflow:

  • Interactive VS Code extension: Living orbital architecture maps, domain clustering, event pub/sub flows, and DB lineage with 1-click source navigation.
  • CLI for dev & CI/CD (ce parse, ce gate): Instant local impact checks and automated Merge Request gates in CI.
  • Result: Zero stale diagrams and automated architectural enforcement.
100% local & air-gapped • Single native binary • Zero cloud dependencies Slide 3 of 16
● Stage 1 of 4: Scanning & semantic graph extraction

From raw polyglot code to a 5-layer semantic graph in seconds

The CodeExplorer workflow moves through 4 stages: 1. Scan → 2. Explore architecture → 3. Connect AI → 4. Control via SIG.

12+
Languages (Tree-sitter)

Java, Kotlin (Backend & Android), C#, TypeScript/JS, Python, Go, Rust, SQL + Swift (iOS) track.

130+
Framework recognizers

Spring Boot, ASP.NET, NestJS, Kafka, RabbitMQ, MassTransit, JPA/Hibernate, MapStruct, EF Core, Redis.

5–25s
Full cold scan

Indexes 100k–500k LOC repositories into a 5-layer graph (L1 structure → L4 topology → L5 views).

< 2 sec
Incremental re-scan

SHA-256 content hashing re-parses only modified files in local SQLite (.codeexplorer/graph.db).

# Stage 1: Index any repository locally or on a CI runner with a single command (100% offline)
$ ce parse .
✓ Parsed 1,420 files across 14 services in 6.4s (incremental: 0.8s) → .codeexplorer/graph.db
Stage 1: Scanning — Fast, deterministic, air-gapped AST + config indexing Slide 4 of 16
● Stage 2 of 4: Exploring architecture (VS Code visuals)

Complete system topology extracted automatically from code

No manual drawing. CodeExplorer synthesizes ASTs and configuration files (application.yml, appsettings.json, .env) into concentric architectural orbits.

🪐 Concentric orbits

  • Orbit 0 (Blue): Client apps, gateways & BFFs.
  • Orbit 1 (Amber): Message topics & event queues.
  • Orbit 2 (Red/Purple): Domain microservices & workers.
  • Outer ring (Purple): Databases, Core systems & caches.

🎯 Instant architectural onboarding

Open any unfamiliar repository and immediately see every deployable host, how services communicate, and which databases they touch.

Full Orbital Architecture Map 🔍 Click to expand full screen
Stage 2: Exploring architecture — Living orbital map in VS Code Slide 5 of 16
● Stage 2 of 4: Exploring architecture (Macro abstraction)

Smart architectural grouping: zero "spaghetti" diagrams

Collapse hundreds of microservices and message topics into high-level logical tiers (Apps ×4 → Services ×6 → Topics ×14).

📦 Automatic & custom grouping

  • 1-click tier collapse: Group all apps, services, or message topics into clean orbital clusters.
  • Custom domain groups: Select any set of nodes to form custom bounded contexts or team ownership domains.
  • High-level clarity: Ideal for architects, tech leads, and product managers estimating cross-team scope.
Grouped Orbits View 🔍 Click to expand full screen
Stage 2: Exploring architecture — Macro orbital grouping Slide 6 of 16
● Stage 2 of 4: Exploring architecture (Cluster drill-down)

Click any group to inspect contained entities & call volumes

Clicking Topics (14) reveals message volume badges (12 msgs, 6 msgs, 5 msgs) and lists every member in the inspector.

🔎 Complete group breakdown

  • Grouped entities (14): Lists all 14 integration events and queues inside the cluster with instant search filter.
  • Connected services (7): Shows every service interacting with the group and its exact role (Publisher vs. Subscriber).
  • Aggregated edge badges: Shows exact message and call counts on every arrow.
Topics Group Sidebar Inspector 🔍 Click to expand full screen
Stage 2: Exploring architecture — Interactive cluster inspector Slide 7 of 16
● Stage 2 of 4: Exploring architecture (Async & event lineage)

1-click event lineage: who publishes & who subscribes

Standard IDE "Find Usages" stops at Kafka, RabbitMQ, or EventBus boundaries. CodeExplorer links producers and consumers into one graph.

⚡ Instant async blast radius

  • Selected topic: OrderStatusChangedToAwaitingValidation...
  • Producer: Ordering Api Publisher
  • Consumers: Catalog Api Subscriber & webapp Subscriber
  • Config resolution: Resolves topic names from application.yml and constants automatically.
Single Topic Publisher and Subscriber Inspector 🔍 Click to expand full screen
Stage 2: Exploring architecture — Asynchronous producer/consumer traceability Slide 8 of 16
● Stage 2 of 4: Exploring architecture (Data & Core Banking lineage)

Database, ORM entity, raw SQL & stored procedure lineage

Trace data flow from API controllers and workers down to specific database schemas, tables, and stored procedures—with clickable code provenance.

🗄️ End-to-end persistence traceability

  • ORM & repositories: Tracks JPA/Hibernate (@Entity, @Table, @Query), EF Core, Dapper, Prisma, TypeORM.
  • Raw SQL & stored procedures: Parses embedded SQL in JdbcTemplate, migrations, and Core Banking stored procedure calls (@Procedure, SimpleJdbcCall).
  • Clickable evidence: Every dependency edge points to the exact file & line in source code.
Database Lineage and Evidence Inspector 🔍 Click to expand full screen
Stage 2: Exploring architecture — Database, SQL & stored procedure lineage Slide 9 of 16
● Stage 3 of 4: Connecting AI agents (MCP Server & CLI)

Turning blind LLMs into architecture-aware engineers

Instead of burning tokens on grep and guessing file dependencies, AI agents query CodeExplorer's deterministic graph via local stdio MCP or CLI.

🔌 19 specialized MCP tools & CLI

  • get_architecture_map — Instant system topology, services, and ingress/egress points in one call.
  • get_call_chain & analyze_code_impact — Exact multi-hop callers, callees, and blast radius before refactoring.
  • inspect_data_lineage — Which services read or write to a specific database table or stored procedure.
  • cypher_query — Full OpenCypher graph queries over AST symbols and relationships.

⚡ Verified benchmark impact

Works with cloud agents (Claude Code, Cursor, Copilot) and 100% air-gapped local LLMs (vLLM, Ollama, Qwen, DeepSeek):

44×
Fewer tokens
7.6×
Faster tasks
100%
Graph recall
# Zero open network ports — pure local stdio pipe
$ ce mcp --workspace .
Stage 3: AI integration — Model Context Protocol (MCP) & headless CLI Slide 10 of 16
● Stage 4 of 4: Transition from visibility to automated control

Semantic Invariant Gates (SIG): executable architecture in CI

Visibility alone doesn't stop architectural drift. Once your architecture is a graph, every architectural boundary becomes a deterministic CI gate.

Enterprise capability 1 • Core system isolation

🛡️ Protecting Core Banking / legacy adapters

Block any Merge Request where business logic or controllers invoke Core Banking clients or legacy stored procedures outside the designated integration.* adapter layer:

id: CORE-ARCH-001
name: core-banking-only-via-integration-layer
cypher: >
  MATCH (caller:Type)-[:CALLS|USES_TYPE*1..2]->(target)
  WHERE target.namespace CONTAINS 'corebanking'
    AND NOT caller.namespace CONTAINS 'integration.corebanking'
  RETURN caller.file_path AS file, caller.start_line AS line,
         caller.name AS symbol, target.name AS target
Enterprise capability 2 • API & ORM hygiene

🧬 Enforcing @Entity → MapStruct → DTO

Guarantee persistence entities never leak into REST controllers or public DTOs—explicitly permitting mapping only inside *Mapper classes:

id: ORM-HYGIENE-002
name: jpa-entity-only-in-mapstruct-mappers
cypher: >
  MATCH (c:Type)-[:CALLS|USES_TYPE|DEPENDS_ON]->(e:Type)
  WHERE e.annotations CONTAINS 'Entity'
    AND (c.name ENDS WITH 'Controller' OR c.name ENDS WITH 'Dto')
    AND NOT c.name ENDS WITH 'Mapper'
  RETURN c.file_path AS file, c.start_line AS line,
         c.name AS symbol, e.name AS target
Stage 4: Architectural control — Semantic Invariant Gates (SIG) Slide 11 of 16
● Stage 4 of 4: Adopting SIG in legacy codebases without friction

The "Ratchet" baseline & why linters / LLMs cannot replace SIG

Enable strict architectural gates in a 500k-line legacy codebase today in 5 minutes—without turning existing Pull/Merge Requests red.

Capability Static linters (Sonar / Detekt / ESLint) Raw LLM PR review CodeExplorer SIG (ce gate)
Multi-hop transitive calls ❌ Single-file syntax only ⚠️ Hallucinates paths ✅ 100% deterministic AST call graph
Cross-layer / ORM / event rules ❌ Blind to semantic topology ⚠️ Non-deterministic (cannot block CI) ✅ Strict graph gate (0 false negatives)
Legacy codebase adoption ❌ Floods CI with 1,000 old errors ❌ No baseline state ✅ Ratchet baseline (blocks only +1 new)
Execution speed in CI/CD 🐢 2–5 minutes 🐢 1–3 minutes + GPU cost ⚡ < 2 seconds per Pull / Merge Request 
# 1. Snapshot legacy debt into baseline (5 min)       2. Headless GitLab CI / GitHub Actions gate (< 2s)
$ ce gate --update-baseline                          $ ce parse . && ce gate --baseline .codeexplorer/gate-baseline.json
Stage 4: Architectural control — Zero-friction ratchet baseline in CI/CD Slide 12 of 16
● Stage 4 horizons: The strategic future of SIG & CodeExplorer

Where we are taking Semantic Invariant Gates next

Evolving SIG from single-repository CI checks into an autonomous, enterprise-wide architectural governance platform:

1. Natural language → verified SIG

🪄 AI-synthesized gates

Architects describe invariants in plain English ("Payment services must never query Customer tables directly") and the AI agent synthesizes, tests, and locks the deterministic OpenCypher gate via MCP.

2. Self-healing AI guardrails

🔄 Real-time agent correction

Before an AI coding agent even opens a Pull Request, it runs ce gate via MCP in its inner loop—catching and auto-fixing its own architectural violations before a human ever sees the code.

3. Declarative domain packs (ce pack)

📦 Pluggable enterprise rule packs

Zero-compile YAML archetype & SIG packs for industry standards (Banking Core / DBOS, Clean/Hexagonal architecture, DDD, Mobile MVVM) shared across all teams in seconds.

4. Federated multi-repo governance

🏛️ Organization-wide graph server

Stitching 100+ repositories into a centralized on-premise graph to enforce cross-repo API contracts, event schema compatibility, and organization-wide ratchet baselines.

Stage 4: Strategic horizons — Autonomous architectural governance Slide 13 of 16
● Enterprise & regulated banking readiness (Security & infrastructure)

Built for strict air-gapped perimeters & high-velocity CI/CD

Engineered from day one to satisfy strict banking InfoSec, Central Bank data residency, and on-premise CI/CD requirements:

Zero external surface

🔒 100% air-gapped & zero telemetry

  • 100% offline native binary: Parsing, VS Code visualization, and CI gates run completely offline without internet access.
  • Zero outbound calls: No telemetry, no analytics, no external webhooks. Source code never leaves your perimeter.
  • Local SQLite storage: The entire graph lives in a single file (.codeexplorer/graph.db) inside your repo/runner cache, governed by your OS & GitLab permissions.
Deterministic & self-contained

⚙️ Zero cloud LLM keys • headless CI ready

  • No LLM required for core engine: ce parse and ce gate are 100% deterministic AST/graph engines—zero cloud API keys needed, zero hallucinations.
  • Drop-in GitLab CI / runner ready: Runs headless on any Linux/Windows runner in 1–3 seconds per Merge Request (incremental SHA-256 hashing).
  • Pure stdio MCP: Connects to IDEs and internal on-prem LLMs via standard input/output streams—opens 0 network ports.
Enterprise & banking readiness — 100% on-premise, air-gapped, deterministic Slide 14 of 16
● Enterprise adoption & commercial partnership

From a 2-week single-repository pilot to organization-wide scale

Proven adoption path for high-velocity engineering teams (50+ engineers, 40–50 Merge Requests/day, multiple parallel stack instances):

Phase 1 • 2 weeks (under NDA)

🚀 1-repository pilot

  • Deploy the air-gapped ce binary on 1 target repository inside your internal GitLab CI.
  • Configure 5 production gates (Core adapter isolation, JPA @Entity → MapStruct → DTO, layer boundaries).
  • Lock existing legacy debt via ratchet baseline so CI stays green from day 1.
Phase 2 • Custom stack packs

🧩 Tailored domain packs

  • Custom YAML rule & recognizer packs for your internal platforms (e.g., Core Banking adapters, Digital Banking OS, stored procedures, internal SDKs).
  • Full-stack coverage across backend (Java, Kotlin, Spring, SQL, .NET) and mobile (Android Kotlin & iOS Swift).
Phase 3 • Enterprise SLA

🏛️ Centralized multi-repo graph

  • On-premise centralized graph server uniting all microservices, mobile apps, and parallel stack copies.
  • Commercial enterprise licensing with dedicated SLA, priority parser engineering, and architect training.
Enterprise rollout — Zero-risk pilot under NDA to full production governance Slide 15 of 16
● Questions & discussion

Thank you!

🌐 Website & documentation

Product portal & live demos

code-explorer.app

Interactive architecture showcase, OpenCypher playground, VS Code extension & CLI install guides.

✉️ Direct contact

Viacheslav Mikhailov

[email protected]

Creator of CodeExplorer • Enterprise pilots, custom architecture packs & licensing.

CodeExplorer (ce) — Architectural intelligence for humans & AI agents Slide 16 of 16