lxDIG MCP

Works with: VS Code Copilot · Claude Code · Claude Desktop · Cursor · any MCP-compatible AI assistant

Supported languages: TypeScript · JavaScript · TSX/JSX · Python · Go · Rust · Java Databases: Memgraph (graph) · Qdrant (vector) Transports: stdio (local) · HTTP (remote/fleet)

What is lxDIG MCP?

An open-source Model Context Protocol (MCP) server that adds a persistent code intelligence layer to AI coding assistants — Claude Code, VS Code Copilot, Cursor, and Claude Desktop. Unlike static RAG or batch-oriented GraphRAG, lxDIG MCP is a live, incrementally-updated intelligence graph that turns any repository into a queryable knowledge graph — so agents can answer architectural questions, track decisions across sessions, coordinate safely in multi-agent workflows, and run only the tests that actually changed — without re-reading the entire codebase on every turn.

It is purpose-built for the agentic coding loop: the cycle of understand → plan → implement → verify → remember that AI agents (Claude, Copilot, Cursor) repeat continuously.

The core problem it solves: most AI coding assistants are stateless and architecturally blind. They re-read unchanged files on every session, miss cross-file relationships, forget past decisions, and collide when multiple agents work in parallel. lxDIG MCP is the memory and structure layer that fixes all four.

Related MCP server: Mono Memory MCP

Table of Contents

• Why lxDIG?

• Key capabilities

• How it works

• Visualize your code graph

• Quick start

• 39 MCP tools — at a glance

• Use cases

• Comparison with alternatives

• Performance

• Roadmap

• Contributing

• Support the project

• License

Why Use a Code Graph MCP Server? Problems lxDIG Solves

Most code intelligence tools solve one of these problems. lxDIG solves all of them together:

Key Capabilities: Code Graph, Agent Memory & Multi-Agent Coordination

1. Code graph intelligence

Turn your repository into a queryable property graph of files, functions, classes, imports, and their relationships. Ask questions in plain English or Cypher.

• Natural-language + Cypher graph queries (graph_query)

• Symbol-level explanation with full dependency context (code_explain)

• Pattern detection and architecture rule validation (find_pattern, arch_validate)

• Architecture placement suggestions for new code (arch_suggest)

• Semantic code slicing — targeted line ranges from a natural query (semantic_slice)

• Find duplicate or similar code across the codebase (find_similar_code, code_clusters)

2. Persistent agent memory

Your agent remembers what it decided, what it changed, what broke, and what it observed — even after a VS Code restart or a Claude Desktop session ends.

• Episode memory: observations, decisions, edits, test results, errors, learnings (episode_add, episode_recall)

• Decision log with semantic query (decision_query)

• Reflection synthesis from recent episodes (reflect)

• Temporal graph model: query any past code state with asOf, compare drift with diff_since

3. Multi-agent coordination

Run multiple AI agents in parallel on the same repository without conflicts.

• Claim/release protocol for file, function, or task ownership (agent_claim, agent_release)

• Fleet-wide coordination view — see what every agent is doing (coordination_overview, agent_status)

• Context packs that assemble high-signal task briefings under strict token budgets (context_pack)

• Blocker detection across agents and tasks (blocking_issues)

4. Test and change intelligence

Stop running your full test suite on every change. Know exactly what's affected.

• Change impact analysis — blast radius of modified files (impact_analyze)

• Selective test execution — only the tests that can fail (test_select, test_run)

• Test categorization for parallelization and prioritization (test_categorize, suggest_tests)

5. Documentation as a first-class knowledge source

Your READMEs, ADRs, and changelogs become searchable graph nodes, linked to the code they describe.

• Index all markdown docs in one call (index_docs)

• Full-text BM25 search across headings and content (search_docs?query=...)

• Symbol-linked lookup — every doc that references a class or function (search_docs?symbol=MyClass)

• Incremental re-index: only changed files are re-parsed

6. Architecture governance

Enforce architectural boundaries automatically and get placement guidance for new code.

• Layer/boundary rule validation (arch_validate)

• Graph-topology-aware placement suggestions (arch_suggest)

• Circular dependency and unused-code detection (find_pattern)

7. One-shot project setup

Go from a fresh clone to a fully wired AI assistant in one tool call.

• init_project_setup — sets workspace, rebuilds graph, generates Copilot instructions

• setup_copilot_instructions — generates .github/copilot-instructions.md from your repo's topology

• Works with VS Code Copilot, Claude Code, Claude Desktop, and any MCP-compatible client

How lxDIG MCP Works: Graph + Vector + BM25 Hybrid Retrieval

lxDIG runs as an MCP server over stdio or HTTP and coordinates three data planes behind a single tool interface:

┌─────────────────────────────────────────────────────────────┐
│                     MCP Tool Surface (39 tools)              │
│  stdio transport (local)  │  HTTP transport (remote/fleet)   │
└──────────────┬────────────┴────────────────┬────────────────┘
               │                             │
   ┌───────────▼────────────┐   ┌────────────▼────────────┐
   │   Graph Plane          │   │   Vector Plane           │
   │   Memgraph (Bolt)      │   │   Qdrant                 │
   │   ─────────────────    │   │   ─────────────────────  │
   │   FILE · FUNC · CLASS  │   │   Semantic embeddings    │
   │   IMPORT · CALL edges  │   │   Nearest-neighbor search│
   │   Temporal tx history  │   │   Natural-language code  │
   └────────────────────────┘   └─────────────────────────┘
               │
   ┌───────────▼────────────────────────────────────────────┐
   │   Hybrid Retrieval (RRF fusion)                         │
   │   Graph expansion + Vector similarity + BM25 lexical   │
   └────────────────────────────────────────────────────────┘

When you call graph_query in natural language mode, retrieval runs as hybrid fusion:

1. Vector similarity search (semantic concepts)

2. BM25 lexical search (keyword matches)

3. Graph expansion from seed nodes (structural relationships)

4. Reciprocal Rank Fusion (RRF) merges all three signals into a single ranked result

The result: structurally accurate, semantically relevant answers — not just the closest embedding match.

System diagram

Visualize Your Code Graph — lxDIG Visual

lxDIG Visual is the open-source browser-based visualization layer for lxDIG MCP. It renders your code dependency graph as an interactive, navigable canvas — turning abstract code relationships into a tangible spatial representation you can explore.

Key features:

• Force-directed interactive graph — files, functions, and classes rendered as explorable nodes with physics-based positioning

• Expand-by-depth navigation — double-click any node to progressively reveal its direct relationships

• Architecture layer awareness — color-coded module boundaries and structural compliance indicators

• Multi-agent visualization — real-time view of coordination when multiple AI agents are active via lxDIG MCP

• Live + mock modes — connects to your running Memgraph instance or uses built-in fallback data

Setup (shares the same Memgraph instance as lxDIG MCP — no extra database needed):

git clone https://github.com/lexCoder2/lxDIG-visual.git
cd lxDIG-visual
npm install && cp .env.example .env
npm run dev:all
# Open http://localhost:5173

After indexing with graph_rebuild, changes appear in the visual explorer immediately — no manual refresh required.

→ github.com/lexCoder2/lxDIG-visual

Quick Start

Recommended setup: Memgraph + Qdrant in Docker, MCP server on your host via stdio. Your editor spawns and owns the process — no HTTP ports, no session headers.

Prerequisites

1. Clone and build

git clone https://github.com/lexCoder2/lxDIG-MCP.git
cd lxDIG-MCP
npm install && npm run build

2. Start the databases

docker compose up -d memgraph qdrant
docker compose ps   # wait for "healthy" (~30 s)

3. Wire your editor

VS Code — add to .vscode/mcp.json:

{
  "servers": {
    "lxdig": {
      "type": "stdio",
      "command": "node",
      "args": ["/absolute/path/to/lxDIG-MCP/dist/server.js"],
      "env": {
        "MCP_TRANSPORT": "stdio",
        "MEMGRAPH_HOST": "localhost",
        "MEMGRAPH_PORT": "7687",
        "QDRANT_HOST": "localhost",
        "QDRANT_PORT": "6333"
      }
    }
  }
}

Claude Desktop — add to claude_desktop_config.json:

{
  "mcpServers": {
    "lxdig": {
      "command": "node",
      "args": ["/absolute/path/to/lxDIG-MCP/dist/server.js"],
      "env": {
        "MCP_TRANSPORT": "stdio",
        "MEMGRAPH_HOST": "localhost",
        "MEMGRAPH_PORT": "7687",
        "QDRANT_HOST": "localhost",
        "QDRANT_PORT": "6333"
      }
    }
  }
}

4. Initialize your project (one call)

{
  "name": "init_project_setup",
  "arguments": {
    "workspaceRoot": "/absolute/path/to/your-project",
    "sourceDir": "src",
    "projectId": "my-repo"
  }
}

This single call sets the workspace context, rebuilds the code graph, and generates .github/copilot-instructions.md for your project. Your agent is ready to query.

Total setup time: ~5 minutes. See QUICK_START.md for the full guide including Docker, Claude Desktop, and HTTP transport.

39 MCP Tools — At a Glance

Use Cases: Claude Code, VS Code Copilot, Cursor & CI Pipelines

Individual developer — Claude Code or VS Code Copilot

• Ask "what calls AuthService.login across the whole repo?" and get a graph answer, not a file dump

• Resume a refactoring task after a VS Code restart — your agent remembers every decision

• Run impact_analyze before committing — know exactly which tests to run

• Use arch_validate to catch layer violations before they become bugs

• Explore your dependency graph visually with lxDIG Visual

Engineering team — multi-agent workflows

• Run a planning agent and an implementation agent in parallel without file conflicts

• Use coordination_overview to see what every agent is working on

• context_pack hands off a high-signal task briefing between agents in one call

• Persistent decision memory means the second agent doesn't repeat work the first already did

CI / automation pipeline

• graph_health as a startup readiness gate

• test_select + test_run for impact-scoped CI that's 5–10x faster than full suite

• arch_validate as an automated architecture compliance check on every PR

Repository onboarding

• init_project_setup on a new codebase — graph + copilot instructions in ~30 seconds

• code_explain to understand unfamiliar subsystems with full dependency context

• setup_copilot_instructions generates AI assistant instructions tailored to your repo's topology

lxDIG MCP vs RAG, GraphRAG, GitHub Copilot & LangChain Agents

Performance

Benchmarks run against a synthetic 20-scenario agent task suite (benchmarks/):

Benchmarks are workload-dependent. Run npm run benchmark:check-regression against your own repository for accurate numbers.

What's Already Shipped

Every feature below is production-ready today:

• ✅ Hybrid retrieval for graph_query — vector + BM25 + graph expansion fused with RRF

• ✅ AST-accurate parsers via tree-sitter for TypeScript, TSX, JS/MJS/CJS, JSX, Python, Go, Rust, Java

• ✅ Watcher-driven incremental rebuilds — graph stays fresh without manual intervention (requires LXDIG_ENABLE_WATCHER=true)

• ✅ Temporal code model — asOf queries any past graph state; diff_since shows what changed

• ✅ Indexing-time symbol summaries — compact-profile answers stay useful in tight token budgets

• ✅ Leiden community detection + PageRank PPR with JS fallbacks for non-MAGE environments

• ✅ SCIP IDs on all FILE, FUNCTION, and CLASS nodes for precise cross-tool symbol references

• ✅ Episode memory, agent coordination, context packs, and response budget shaping

• ✅ Docs & ADR indexing — markdown parsed into graph nodes; queried by text or symbol association

• ✅ Interactive graph visualization via lxDIG Visual — force-directed canvas explorer

• ✅ 557 tests across parsers, builders, engines, and tool handlers — all green

Runtime Modes

Useful scripts

npm run start                       # stdio server (recommended)
npm run start:http                  # HTTP supervisor (multi-session)
npm run build                       # compile TypeScript
npm test                            # run all 557 tests
npm run benchmark:check-regression  # check latency/token regressions

Repository Map

Integration Tips

• Start every session with graph_set_workspace → graph_rebuild (or configure init_project_setup to run automatically)

• Prefer graph_query over file reads for discovery — far fewer tokens, cross-file context included

• Use profile: compact in autonomous loops; switch to balanced or debug when you need detail

• Rebuild incrementally after meaningful edits; the file watcher handles this automatically during active sessions

• Run impact_analyze before tests so your agent only executes what's actually affected

• Open lxDIG Visual alongside your editor for a spatial view of the graph while your agent works

Roadmap

lxDIG is open source and self-hosted today. Planned work ahead — see ROADMAP.md for the full prioritized backlog with detail on each item.

• Language server protocol (LSP) integration for deeper symbol resolution

• Go, Rust, Java parser improvements

• MCP resources surface (expose graph nodes as MCP resources)

• Webhook-triggered graph rebuilds for CI environments

• Plugin API for custom tool registration

• Real-time transparent graph sync — continuous file-watching with live graph and vector index updates surfaced as observable events, so agents and users always know when the graph is current without polling graph_health or triggering manual rebuilds

• Domain knowledge layer — attach external knowledge sources (documentation, standards, specs, research articles) directly to code symbols as graph nodes; a calculateBMI function links to CDC/WHO references, a payment function links to PCI-DSS rules, a GDPR-scoped model links to regulation articles — giving agents real-world context alongside structural context

• Multi-user coordination — shared agent memory, task ownership, and conflict detection across multiple developers on the same repository

• lxDIG Cloud — hosted, zero-infrastructure version for individuals and teams

Contributing

Pull requests are welcome. Whether it's a new parser, a tool improvement, a bug fix, or better docs — contributions of all sizes move this project forward.

• Bugs / features — open an issue first to align on scope

• New tools — follow the handler + registration pattern in src/tools/; include tests

• New language parsers — add tree-sitter grammar + tests in src/parsers/

• Docs — typos, clarifications, and examples are always appreciated

→ Open a pull request · → Browse open issues

Support the Project

lxDIG MCP is built and maintained in personal time — researching graph retrieval techniques, designing the tool surface, writing tests, and keeping everything working across MCP protocol updates. If it saves you time or makes your AI-assisted workflows meaningfully better, consider supporting the work:

• GitHub Sponsors → github.com/sponsors/lexCoder2

• Buy Me a Coffee → buymeacoffee.com/hi8g

FAQ

Q: Does lxDIG require a cloud service or API key? No. lxDIG runs entirely on your machine. Memgraph and Qdrant run in Docker containers you control. No data leaves your environment.

Q: Does it work with Cursor? Yes. Any MCP-compatible client works. Add the stdio config to Cursor's MCP settings the same way as VS Code.

Q: How large a codebase can it handle? The graph plane (Memgraph) scales to millions of nodes. For very large monorepos, use sourceDir to scope indexing to the relevant subdirectory. Incremental rebuilds keep the graph fresh without re-indexing everything.

Q: Do I need to run Qdrant? Qdrant is optional but recommended for large codebases. Without it, semantic_search and find_similar_code are unavailable; all other tools continue to work via graph-only or BM25 retrieval.

Q: Can multiple developers on a team share one lxDIG instance? Yes, via HTTP transport. One running instance handles multiple independent sessions. Team-level shared memory is on the lxDIG Cloud roadmap.

Q: Is this production-ready? The core tools are stable and tested (402 tests, all green). Treat it as beta — APIs may change before a 1.0 release. Pin your version and watch the changelog.

Q: Is lxDIG MCP the same as GraphRAG? No. GraphRAG is a batch retrieval technique applied to documents. lxDIG MCP is a live, incrementally-updated code graph with persistent agent memory, multi-agent coordination, and impact-scoped test selection — not just a retrieval improvement.

Q: How do I add persistent memory to Claude Code? Install lxDIG MCP, add the stdio config to .vscode/mcp.json, and call init_project_setup once per repository. From that point, Claude Code can call episode_add / episode_recall and decision_query to read and write memory that persists across sessions.

Q: Can I visualize the code graph? Yes. lxDIG Visual is the companion browser-based graph explorer. It shares the same Memgraph instance — run npm run dev:all in the lxDIG-visual repo and open http://localhost:5173.

License

MIT — free to use, modify, and distribute.