{ "tool": "image-mcp", "arguments": { "model": "fal-ai/nano-banana-pro", "prompt": "Professional enterprise architecture diagram titled \"Enterprise AI Architecture with LangGraph Runtime - Banking Use Case\"\n\nVertical layered diagram with 5 distinct horizontal layers:\n\nLAYER 1 (top, light blue band): User personas in boxes - Customer Service Agents, Loan Officers, Compliance Team, IT Operations\n\nLAYER 2 (orange band): Wide box \"Banking AI Assistant\" with subtitle \"Cursor IDE / Custom UI\"\n\nLAYER 3 (white): Model layer showing GPT-4/Claude box (blue) on left, diamond \"Model-Agnostic\" in center, LangGraph/CrewAI box (purple) on right. Arrows labeled \"Tool Calls\" and \"API Calls\"\n\nLAYER 4 (purple gradient): \"LangGraph Runtime\" header\nTop row of 5 component boxes: StateGraph, Node Registry, Checkpointing, Secret Management, Session Manager with small labels \"Real-time execution\", \"User-specific permissions\", \"Secure credentials\", \"Audit logs\"\nBottom row of 6 agent boxes (darker purple): Salesforce MCP, Email MCP Gmail, Slack MCP, Database MCP, Document MCP, Custom Banking MCP\n\nLAYER 5 (bottom, dark navy): \"Bank's Existing Infrastructure\" - Six system boxes: Core Banking System, CRM Salesforce, Compliance Database, Document Repository, Communication Systems, Legacy APIs. Downward arrows from Layer 4 labeled \"Authenticated Requests\"\n\nStyle: Clean layered enterprise architecture, professional blue-purple-orange color scheme, clear hierarchy, numbered layer labels on left side, consistent arrow direction top-to-bottom, corporate technical documentation aesthetic", "aspect_ratio": "16:9" } }

{ "tool": "image-mcp", "arguments": { "model": "fal-ai/nano-banana-pro", "prompt": "Clean technical architecture diagram titled \"LangGraph Architecture - Stateful Agent Orchestration Framework\"\n\nWhite background, 2x3 grid layout with soft pastel colored section headers\n\nTOP ROW:\n[Core Components] soft blue header - bullet list: State (shared data TypedDict), Nodes (functions), Edges (routing). Below: flow diagram State to Node to Updated State to diamond Edge Decision to Next Node\n\n[StateGraph Class] soft purple header - bullet list: Define State schema, Add nodes, Add edges, Compile to Pregel. Below: linear flow Define Build Compile Execute\n\nMIDDLE ROW full width:\n[Pregel-Inspired Execution] soft orange header - Left: bullets for Super-steps, Message passing, Parallel within step, Sequential across steps. Right: three connected boxes showing Super-step 1 with Node A Node B, arrow to Super-step 2 with Node C, arrow to Super-step 3 with Node D Node E\n\nBOTTOM ROW:\n[Checkpointing System] soft yellow header - BaseCheckpointSaver interface, checkpoint-postgres, checkpoint-sqlite, State snapshots, Time-travel support. Small flow State to Checkpoint to Storage\n\n[Monorepo Structure] soft gray header - package list: langgraph core, prebuilt APIs, checkpoint interfaces, cli, sdk-py, sdk-js\n\n[Capabilities] soft green header - checkmark list: Durable execution, Human-in-loop, Memory, Streaming, Multi-agent, Sub-graphs, Conditional routing\n\nStyle: Clean minimalist, soft pastel headers, thin borders, generous whitespace, easy to read, professional C4 style documentation diagram", "aspect_ratio": "16:9" } }

{ "tool": "image-mcp", "arguments": { "model": "fal-ai/nano-banana-pro", "prompt": "Modern tech blog hero image for \"Architecture Diagram Showdown\"\n\nSplit screen design showing:\nLeft side: Simple code/repository icon with arrows pointing to a basic diagram outline (representing automated analysis)\nRight side: Brain/AI icon combined with magnifying glass pointing to a detailed polished diagram (representing expert analysis)\n\nCenter: VS symbol or lightning bolt connecting both sides\n\nBottom: Subtle visualization of LangGraph-style node connections\n\nStyle: Clean modern tech illustration, dark navy blue background with vibrant accent colors (coral orange, teal, purple gradients), minimalist icons, slight glow effects, professional blog thumbnail aesthetic, no text needed\n\nMood: Technical comparison, innovation, AI-powered tools", "aspect_ratio": "16:9" } }

Same Model, Two Approaches

Blueprint MCP made waves by generating architecture diagrams from codebases in ~60 seconds. It uses Nano Banana Pro for image generation with built-in automated code analysis.

image-mcp also uses Nano Banana Pro - but as a generalist image tool without code analysis built in. The agent has to do that work.

	Blueprint MCP	image-mcp + Agent
Image Model	Nano Banana Pro	Nano Banana Pro
Code Analysis	Built into the MCP	Agent does the work
Speed	~60 seconds	~10 minutes
Approach	Specialized automation	Generalist + agent effort

The question: When both tools use the same underlying model, can agent-driven deep analysis beat specialized automation?

The Agent-Driven Approach

Without built-in code analysis, we used 3 parallel sub-agents to explore the LangGraph source code:

Agent 1: Core architecture (StateGraph, Pregel, Channels)
Agent 2: Execution model (super-steps, task scheduling, version tracking)
Agent 3: Persistence & capabilities (checkpointing, interrupts, streaming)

What Deep Analysis Found

Reading actual source files uncovered implementation details:

BaseCheckpointSaver (Abstract)
├── InMemorySaver (testing)
├── PostgresSaver (production)
└── SqliteSaver (local dev)
 
Channel Types:
├── LastValue - stores single latest value
├── BinaryOperatorAggregate - reduces with operator
├── Topic - pub-sub for Send dispatch
└── EphemeralValue - not persisted
 
Durability Modes: sync | async | exit
Streaming Modes: values | updates | checkpoints | tasks | debug | messages | custom

Expert Prompt Engineering

With real architectural knowledge, we crafted prompts using:

2025 design trends (bold minimalism)
Microsoft's architecture diagram best practices (directional arrows, clear labels, visual hierarchy)
C4 model principles (layered abstraction, consistent notation)

Results

Banking Use Case Diagram

The enterprise architecture diagram shows clear advantages:

Professional iconography throughout
Cleaner layer separation
Proper MCP server representation
Better color hierarchy

LangGraph Core Architecture

After refinement, a cleaner diagram matching Blueprint MCP's scannability while including deeper technical accuracy from source code analysis.

Verdict

Same model, different strengths:

Blueprint MCP (Specialized + Automated)

✅ Fast (~60 seconds) ✅ Zero prep required ✅ Great for quick exploration ❌ Limited customization ❌ Analysis depth depends on automation

image-mcp + Agent (Generalist + Manual)

✅ Deeper analysis possible ✅ Full control over prompts and style ✅ Can incorporate domain expertise ❌ Slower (~10 minutes) ❌ Requires agent effort

Key Insight

The agent is the differentiator, not the MCP.

When both tools use the same underlying model, quality comes from:

Depth of analysis - automated scanning vs agent reading source files
Prompt engineering - generic vs expert-crafted
Design knowledge - defaults vs researched best practices

A generalist tool + capable agent can beat specialized automation when the agent invests in understanding the domain.

When to Use Each

Blueprint MCP: Quick results, exploring unfamiliar codebases, "good enough" diagrams

image-mcp + agent: Accuracy matters, specific design requirements, control over emphasis

Power move: Blueprint MCP for exploration, image-mcp for polished final output.

Same Model, Different Approach: Agent-Driven Analysis vs Specialized MCP for Architecture Diagrams

Prompt

Prompt

Prompt

Same Model, Two Approaches

The Agent-Driven Approach

What Deep Analysis Found

Expert Prompt Engineering

Results

Banking Use Case Diagram

LangGraph Core Architecture

Verdict

Blueprint MCP (Specialized + Automated)

image-mcp + Agent (Generalist + Manual)

Key Insight

When to Use Each