The problem

AI coding agents are powerful, but the way most developers use them - so-called vibe coding - creates three concrete problems:

Every session starts from scratch. The agent doesn’t remember what it did yesterday. You spend 30-50% of tokens re-explaining context, re-reading files, re-describing the architecture.
PRDs become disconnected tasks. You import a requirements document and tasks land with no connection between them. No dependencies, no calculated priorities, no traceability.
AI tools don’t communicate. Your code assistant doesn’t talk to impact analysis. Documentation search doesn’t feed the task planner. Zero coordination.
No discipline is enforced. The agent generates code without tests, without typing, without following project patterns. It works until it breaks in production.

The solution

mcp-graph is a local-first TypeScript CLI that converts PRDs (Markdown, TXT, PDF, HTML) into persistent execution graphs stored in SQLite.

The system automatically generates a hierarchical task tree with 9 node types - epics, tasks, subtasks, requirements, constraints, milestones, acceptance criteria, risks, and decisions. It infers dependencies, estimates priorities, and creates a navigable graph that any AI agent can query.

Technical differentiators

Token Economy - 70-85% reduction

3-tier compression to deliver only the necessary context to the agent:

Tier	Tokens/node	Usage
Summary	~20	Structure overview
Standard	~150	Context + dependencies + BM25 snippets
Deep	500+	Complete documents + acceptance criteria

Intelligently allocated budget: 60% graph, 30% knowledge, 10% metadata.

Multi-Agent Intelligence Mesh

Coordination of 5 MCPs via reactive EventBus:

mcp-graph → source of truth (graph + knowledge store)
Serena → code analysis, semantic navigation, agent memory
GitNexus → code intelligence, impact analysis, blast radius
Context7 → up-to-date library documentation, local cache
Playwright → browser validation, screenshots, interface testing

100% Local Knowledge Pipeline

Keyword search via SQLite FTS5 with BM25 ranking
Semantic search via TF-IDF (~10 MB vs ~400 MB transformer models)
Hybrid mode with automatic deduplication
5 indexed sources: uploads, Serena memories, code, docs, web captures
SHA-256 deduplication

No API keys. No per-query costs. No data leaving your environment.

Anti-Vibe-Coding methodology

mcp-graph embeds a methodology based on Extreme Programming (XP):

TDD enforced - every feature requires a test BEFORE implementation
Skeleton & Organs - the human defines the architecture, the AI implements with discipline
Anti-one-shot - decomposition into atomic tasks tracked in the graph
8-phase cycle: ANALYZE → DESIGN → PLAN → IMPLEMENT → VALIDATE → REVIEW → HANDOFF → LISTENING

Project numbers (v7.0)

Metric	Value
Version	7.0
Tests	910+ across 105 files (1,337+ test cases)
Indexed symbols	1,017
Graph relationships	2,650
Execution flows	67
MCP tools	26
REST endpoints	44
Multimodal skills	155 (audio, video, CV, autonomous orchestration)
Dashboard tabs	6
TypeScript	strict mode, zero `any`
Determinism	100% (Deterministic-First architecture)

Harness Engineering (v7.0)

v7 introduced the Harnessability Score concept, a composite metric that evaluates how safe the code is for agents to manipulate:

Unified Gate System: 50% less overhead per MCP call, 50% fewer database reads
Knowledge Autoprune: RAG stays fast (under 20ms for 500 documents)
DORA Metrics Elite: 25.4 tasks/day, MTTR 0h, Change Failure Rate 0%
Security Hardened: protection against 20 Path Traversal vectors in a single barrier

Ecosystem

26 MCP tools compatible with Claude, Copilot, Cursor, Windsurf, Zed
44 REST endpoints with full API
155 multimodal skills (audio, video, computer vision, orchestration)
Interactive dashboard built with React 19 + React Flow
npm: @mcp-graph-workflow/mcp-graph
License: MIT

Technologies used

TypeScript / Node.js (strict mode)
SQLite (database + FTS5)
React 19 + React Flow (dashboard)
TF-IDF + BM25 (local hybrid search)
MCP (Model Context Protocol)