The diagram that boosts AEO and LLM citations.

Visual TL;DR is an auto-generated flow diagram for blog posts and news articles. Readers get the takeaway in 5 seconds. Dwell time goes up, bounce rate goes down, scroll-depth climbs.

Color
Font
Paste a public article URL. We'll fetch it, run the diagram pipeline, and render the result inline. No sign-up required for this preview.
Free preview: one diagram per IP per hour. Theme + font choices flow through to the SVG output (API: options.theme / options.font). For unlimited use, grab an API key.
Download WordPress plugin

How it works

1

Send title + content

POST { title, content_html } to /api/v1/visual-tldr/generate with your Bearer key. Or have the WordPress plugin fire it automatically on publish. Or call generate_visual_tldr from Claude over MCP.

2

Gemini extracts the graph

6-10 typed concept nodes + edges with short verb labels, grounded in the actual article text. Essential nodes are flagged so the Quick view is the truthiest summary.

3

Get inject-ready HTML

svg_html with both desktop and mobile SVG variants in a single CSS-toggled wrapper. Drop into your post body, or let the plugin / MCP write it back via the WP REST API.

From live articles

Real Visual TL;DR diagrams on StartupHub

These are copied verbatim from published articles. Key steps and Full picture tabs, expandable nodes with article quotes, and JSON-LD ImageObject in the page head ship on every post.

Triple-Stage AI Attack Chain Hits PyTorch

8 concept nodes

/posts/triple-stage-ai-attack-chain-hits-pytorch
Visual TL;DR
Untrusted PyTorch model fileDriver
an attacker supplied torch checkpoint opened by a victim application
From the article 6 mentionsTwo independent researchers demoed a triple-stage AI attack chain at Black Hat Asia 2026 that turns a PyTorch model file into remote code execution even with weights_only set to True.
Triple-stage exploit chainCore
model file to memory corruption to remote code execution in three linked steps
From the articleTwo independent researchers demoed a triple-stage AI attack chain at Black Hat Asia 2026 that turns a PyTorch model file into remote code execution even with weights_only set to True.
OpenSearch demo compromiseOutcome
researchers demonstrated the full chain against a live OpenSearch deployment on stage
torch.load with weights_only=TrueContext
mode docs called secure but still dispatches to torch.jit.load on TorchScript
From the article 5 mentionsThe researchers found torch.load with weights_only=True still calls torch.jit.load if the file is TorchScript format, and TorchScript exposes high-risk operators like torch.save and torch.from_file.
TorchScript as_strided bugDriver
C++ memory safety underflow in the as_strided operator exposed to loading path
From the articleCalling the API directly in Python reproduced the bug, but saving it via forward into a TorchScript file and reloading with torch.load with weights_only True initially hid it.
Untrusted PyTorch model fileDriver
an attacker supplied torch checkpoint opened by a victim application
From the article 6 mentionsTwo independent researchers demoed a triple-stage AI attack chain at Black Hat Asia 2026 that turns a PyTorch model file into remote code execution even with weights_only set to True.
torch.load with weights_only=TrueContext
mode docs called secure but still dispatches to torch.jit.load on TorchScript
From the article 5 mentionsThe researchers found torch.load with weights_only=True still calls torch.jit.load if the file is TorchScript format, and TorchScript exposes high-risk operators like torch.save and torch.from_file.
TorchScript as_strided bugDriver
C++ memory safety underflow in the as_strided operator exposed to loading path
From the articleCalling the API directly in Python reproduced the bug, but saving it via forward into a TorchScript file and reloading with torch.load with weights_only True initially hid it.
Heap underflow primitiveEffect
attacker controls arbitrary read and write over process memory via crafted strides
From the articleA second check rejects storage_offset less than zero, which is why minus 1 fails but 2^62 minus 1 succeeds and returns 0x41, the glibc heap chunk header, confirming an underflow.
Triple-stage exploit chainCore
model file to memory corruption to remote code execution in three linked steps
From the articleTwo independent researchers demoed a triple-stage AI attack chain at Black Hat Asia 2026 that turns a PyTorch model file into remote code execution even with weights_only set to True.
Initial PoC failsEffect
model serialization changes broke the researchers first weaponized proof of concept
RCE via native gadgetsOutcome
researchers pivoted from pickle RCE to native code execution through the underflow primitive
OpenSearch demo compromiseOutcome
researchers demonstrated the full chain against a live OpenSearch deployment on stage

a16z Launches $1.1B Machine Age Fund for AI

7 concept nodes

/posts/a16z-launches-1-1b-machine-age-fund-for-ai
Visual TL;DR
US AI compute shortageDriver
Power and compute gaps as AI racks hit 1 MW scale
From the article 3 mentionsThe firm frames AI as alchemy that turns sand into thought, and thought at scale now demands a rebuild of American compute infrastructure.
Machine Age Fund launchCore
From the articlea16z raised $1.1 billion for its Machine Age Fund to accelerate the physical buildout of AI.
Full stack backingEffect
Fund targets chips, memory, networking, storage, data centers, robotics, home AI
From the articleThe fund will back the full physical stack on which AI runs.
Infrastructure rebootOutcome
US compute stack reboots to meet AI power density demands
From the article 3 mentionsCasado leads infrastructure, Raghuram is managing partner on Growth and Infrastructure after decades running data center system software, Ulevitch leads American Dynamism, and George leads Growth.
US AI compute shortageDriver
Power and compute gaps as AI racks hit 1 MW scale
From the article 3 mentionsThe firm frames AI as alchemy that turns sand into thought, and thought at scale now demands a rebuild of American compute infrastructure.
Key partners leadContext
Ben Horowitz, Martin Casado, Raghu Raghuram, David Ulevitch, David George
From the articleCasado leads infrastructure, Raghuram is managing partner on Growth and Infrastructure after decades running data center system software, Ulevitch leads American Dynamism, and George leads Growth.
Casado leads infraContext
Martin Casado directs infrastructure investments across data centers and networking
From the articleCasado leads infrastructure, Raghuram is managing partner on Growth and Infrastructure after decades running data center system software, Ulevitch leads American Dynamism, and George leads Growth.
Raghuram growth leadContext
From the articleCasado leads infrastructure, Raghuram is managing partner on Growth and Infrastructure after decades running data center system software, Ulevitch leads American Dynamism, and George leads Growth.
Machine Age Fund launchCore
From the articlea16z raised $1.1 billion for its Machine Age Fund to accelerate the physical buildout of AI.
Full stack backingEffect
Fund targets chips, memory, networking, storage, data centers, robotics, home AI
From the articleThe fund will back the full physical stack on which AI runs.
Infrastructure rebootOutcome
US compute stack reboots to meet AI power density demands
From the article 3 mentionsCasado leads infrastructure, Raghuram is managing partner on Growth and Infrastructure after decades running data center system software, Ulevitch leads American Dynamism, and George leads Growth.

Rocket.Chat E2EE Cracked: 9-Char Passwords Fail

8 concept nodes

/posts/rocket-chat-e2ee-cracked-9-char-passwords-fail
Visual TL;DR
Rocket.Chat E2EE BackupCore
RSA private keys encrypted with password, stored on server for multi-device sync
From the article 6 mentionsRocket.Chat E2EE was completely broken by a predictable backup password, according to BlackHat research presented at Black Hat Asia 2026.
Single PBKDF2 IterationDriver
Only one round of key stretching made offline password guessing trivially fast
Full Key RecoveryOutcome
Attacker could reconstruct the RSA private key and impersonate any user
Math.random GeneratorDriver
Custom code picked 9 characters from 62-char set using non-crypto Mersenne Twister
From the article 2 mentionsCode review showed a custom generator picking 9 characters from a 62 character set of upper, lower and digits using Math.random.
Rocket.Chat E2EE BackupCore
RSA private keys encrypted with password, stored on server for multi-device sync
From the article 6 mentionsRocket.Chat E2EE was completely broken by a predictable backup password, according to BlackHat research presented at Black Hat Asia 2026.
Math.random GeneratorDriver
Custom code picked 9 characters from 62-char set using non-crypto Mersenne Twister
From the article 2 mentionsCode review showed a custom generator picking 9 characters from a 62 character set of upper, lower and digits using Math.random.
toLowerCase CollapseDriver
Silent case fold shrank character set from 62 down to 36 effective symbols
From the articleThen a toLowerCase call silently collapsed the set to 36 characters and doubled the weight of lowercase letters, creating a strong and exploitable bias.
Single PBKDF2 IterationDriver
Only one round of key stretching made offline password guessing trivially fast
Black Hat Asia 2026Context
Hay Kimura and NTT researchers presented full decryption attack at Singapore conference
From the articleRocket.Chat E2EE was completely broken by a predictable backup password, according to BlackHat research presented at Black Hat Asia 2026.
Full Key RecoveryOutcome
Attacker could reconstruct the RSA private key and impersonate any user
AES-CBC No MACEffect
Messages lacked integrity check, letting attackers tamper with chat content silently
From the articleRocket.Chat uses RSA key backups encrypted by a password and stored on the server for multi device sync, RSA-encrypted group session keys, and AES-CBC messages with no integrity check.
150-Country DeploymentOutcome
Open source platform used by governments and enterprises worldwide now exposed
For AI agents + humans

Four formats, one API call

Same article in. JSON diagram, Markdown TL;DR, interactive HTML, and JSON-LD block out. AI agents pick the format they parse fastest, humans get expandable cards, search engines get the JSON-LD.

Example below is from a16z Launches $1.1B Machine Age Fund for AI, a published StartupHub article. No mock data.

  • JSON diagram: typed nodes + verb-labeled edges from the article text
  • Markdown TL;DR, paste intoNotionSubstackClaude
  • HTML diagram: Key steps / Full picture tabs, quotes from the article
  • JSON-LD ImageObject, citation-ready forGoogle AI Overviews
Try it freeSee the API shape
Visual TL;DR
US AI compute shortageDriver
Power and compute gaps as AI racks hit 1 MW scale
From the article 3 mentionsThe firm frames AI as alchemy that turns sand into thought, and thought at scale now demands a rebuild of American compute infrastructure.
Machine Age Fund launchCore
From the articlea16z raised $1.1 billion for its Machine Age Fund to accelerate the physical buildout of AI.
Full stack backingEffect
Fund targets chips, memory, networking, storage, data centers, robotics, home AI
From the articleThe fund will back the full physical stack on which AI runs.
Infrastructure rebootOutcome
US compute stack reboots to meet AI power density demands
From the article 3 mentionsCasado leads infrastructure, Raghuram is managing partner on Growth and Infrastructure after decades running data center system software, Ulevitch leads American Dynamism, and George leads Growth.
US AI compute shortageDriver
Power and compute gaps as AI racks hit 1 MW scale
From the article 3 mentionsThe firm frames AI as alchemy that turns sand into thought, and thought at scale now demands a rebuild of American compute infrastructure.
Key partners leadContext
Ben Horowitz, Martin Casado, Raghu Raghuram, David Ulevitch, David George
From the articleCasado leads infrastructure, Raghuram is managing partner on Growth and Infrastructure after decades running data center system software, Ulevitch leads American Dynamism, and George leads Growth.
Casado leads infraContext
Martin Casado directs infrastructure investments across data centers and networking
From the articleCasado leads infrastructure, Raghuram is managing partner on Growth and Infrastructure after decades running data center system software, Ulevitch leads American Dynamism, and George leads Growth.
Raghuram growth leadContext
From the articleCasado leads infrastructure, Raghuram is managing partner on Growth and Infrastructure after decades running data center system software, Ulevitch leads American Dynamism, and George leads Growth.
Machine Age Fund launchCore
From the articlea16z raised $1.1 billion for its Machine Age Fund to accelerate the physical buildout of AI.
Full stack backingEffect
Fund targets chips, memory, networking, storage, data centers, robotics, home AI
From the articleThe fund will back the full physical stack on which AI runs.
Infrastructure rebootOutcome
US compute stack reboots to meet AI power density demands
From the article 3 mentionsCasado leads infrastructure, Raghuram is managing partner on Growth and Infrastructure after decades running data center system software, Ulevitch leads American Dynamism, and George leads Growth.
{
  "title": "a16z Launches $1.1B Machine Age Fund for AI",
  "nodes": [
    {
      "id": "us-ai-compute-shortage",
      "label": "US AI compute shortage",
      "type": "cause",
      "desc": "Power and compute gaps as AI racks hit 1 MW scale"
    },
    {
      "id": "machine-age-fund-launch",
      "label": "Machine Age Fund launch",
      "type": "key",
      "desc": "a16z raised $1.1B to accelerate physical AI infrastructure buildout nationwide"
    },
    {
      "id": "full-stack-backing",
      "label": "Full stack backing",
      "type": "effect",
      "desc": "Fund targets chips, memory, networking, storage, data centers, robotics, home AI"
    },
    {
      "id": "infrastructure-reboot",
      "label": "Infrastructure reboot",
      "type": "result",
      "desc": "US compute stack reboots to meet AI power density demands"
    },
    {
      "id": "key-partners-lead",
      "label": "Key partners lead",
      "type": "concept",
      "desc": "Ben Horowitz, Martin Casado, Raghu Raghuram, David Ulevitch, David George"
    },
    {
      "id": "casado-leads-infra",
      "label": "Casado leads infra",
      "type": "concept",
      "desc": "Martin Casado directs infrastructure investments across data centers and networking"
    },
    {
      "id": "raghuram-growth-lead",
      "label": "Raghuram growth lead",
      "type": "concept",
      "desc": "Raghu Raghuram manages growth and infrastructure after decades running data center system software"
    }
  ],
  "edges": [
    {
      "from": "us-ai-compute-shortage",
      "to": "machine-age-fund-launch",
      "label": "drives"
    },
    {
      "from": "machine-age-fund-launch",
      "to": "full-stack-backing",
      "label": "enables"
    },
    {
      "from": "full-stack-backing",
      "to": "infrastructure-reboot",
      "label": "creates"
    }
  ]
}
For AI agents + humans

Four formats, one API call

Every diagram is rendered in JSON, Markdown, SVG, and JSON-LD simultaneously. No second call, no format conversion, no parsing the SVG to pull the graph back out.

JSON diagram
Agent-native
{
  "title": "Triple-Stage AI Attack Chain Hits PyTorch",
  "nodes": [
    {
      "id": "untrusted-pytorch-model-file",
      "label": "Untrusted PyTorch model file",
      "type": "cause",
      "desc": "an attacker supplied torch checkpoint opened by a victim application"
    },
    {
      "id": "torch-load-with-weights-only-true",
      "label": "torch.load with weights_only=True",
      "type": "concept",
      "desc": "mode docs called secure but still dispatches to torch.jit.load on TorchScript"
    },
    {
      "id": "torchscript-as-strided-bug",
      "label": "TorchScript as_strided bug",
      "type": "cause",
      "desc": "C++ memory safety underflow in the as_strided operator exposed to loading path"
    },
    {
      "id": "heap-underflow-primitive",
      "label": "Heap underflow primitive",
      "type": "effect",
      "desc": "attacker controls arbitrary read and write over process memory via crafted strides"
    },
    {
      "id": "triple-stage-exploit-chain",
      "label": "Triple-stage exploit chain",
      "type": "key",
      "desc": "model file to memory corruption to remote code execution in three linked steps"
    },
    {
      "id": "initial-poc-fails",
      "label": "Initial PoC fails",
      "type": "effect",
      "desc": "model serialization changes broke the researchers first weaponized proof of concept"
    },
    {
      "id": "rce-via-native-gadgets",
      "label": "RCE via native gadgets",
      "type": "result",
      "desc": "researchers pivoted from pickle RCE to native code execution through the underflow primitive"
    },
    {
      "id": "opensearch-demo-compromise",
      "label": "OpenSearch demo compromise",
      "type": "result",
      "desc": "researchers demonstrated the full chain against a live OpenSearch deployment on stage"
    }
  ],
  "edges": [
    {
      "from": "untrusted-pytorch-model-file",
      "to": "torch-load-with-weights-only-true",
      "label": "loaded by"
    },
    {
      "from": "torch-load-with-weights-only-true",
      "to": "torchscript-as-strided-bug",
      "label": "reaches"
    },
    {
      "from": "torchscript-as-strided-bug",
      "to": "heap-underflow-primitive",
      "label": "yields"
    },
    {
      "from": "heap-underflow-primitive",
      "to": "triple-stage-exploit-chain",
      "label": "enables"
    },
    {
      "from": "triple-stage-exploit-chain",
      "to": "initial-poc-fails",
      "label": "first attempt"
    },
    {
      "from": "initial-poc-fails",
      "to": "rce-via-native-gadgets",
      "label": "forces pivot"
    },
    {
      "from": "rce-via-native-gadgets",
      "to": "opensearch-demo-compromise",
      "label": "proven in"
    }
  ]
}
Typed nodes and verb-labelled edges from the PyTorch attack-chain article above. Same shape every call. The format AI agents prefer for downstream reasoning.

Why agents prefer JSON diagrams over rendered SVG: an SVG forces an LLM to OCR text, infer node connections from coordinates, and lose the semantic typing. The JSON above takes 200 tokens and is fully traversable. That is why every serious agent-to-article pipeline now pulls the JSON, not the picture.

What ships with every diagram

3 cascading views

Each diagram renders as Quick (4-5 essential nodes), Explain (essentials + inline descriptions), and Deeper (full extracted concept graph). CSS-only toggle, no JS framework needed in the host page.

Concept extraction with Gemini

Gemini grounds the diagram in the article text itself: nodes are typed (cause / effect / result / key / concept) and edges carry short verbs ("leads to", "causes", "enables") so the diagram reads as prose.

Responsive without scaling text

Two SVG variants ship in every diagram. Desktop renders the cascade at native 808px so text stays 16px. Mobile (<640px) swaps to a 340px-wide compact layout via CSS @media: readable at 14px, no horizontal scroll.

WordPress plugin

Download our plugin .zip, upload via wp-admin → Plugins → Add New → Upload, activate, set your API key. Diagrams generate automatically on post publish OR on-demand from the editor. Caches per-post so a re-render doesn't re-spend credits. (Not on the WP marketplace, we ship directly so updates land same-day.)

REST API for any platform

POST title + content_html to /api/v1/visual-tldr/generate, get back svg_html ready to inject. Ghost, Substack-via-RSS, Jamstack, anywhere with a publish webhook works.

MCP for AI workflows

Same generator exposed as the generate_visual_tldr MCP tool. Connect once in Claude.ai / Cursor / Windsurf and your agent can drop a diagram into any post via natural language.

Who uses it

Editorial publishers

Auto-add a Visual TL;DR to every news article. AI Overviews and ChatGPT cite the diagram's alt-text + JSON-LD as a structured summary, lifting AI-attribution traffic.

apnews.combusinesswire.comprnewswire.com

B2B blogs + docs

Long-form technical posts get a flow diagram readers can scan in 5 seconds before deciding to commit to the 10-minute read. Dwell time and scroll depth both move.

substack.commedium.com

AI agent pipelines

Drop the MCP server into a content-publishing agent. After draft + review + image, the agent calls generate_visual_tldr and pastes the result inline before publish.

langchain.comcrewai.comn8n.iozapier.com

Direct API call

Identical response shape via REST or MCP.

curl -X POST https://www.startuphub.ai/api/v1/visual-tldr/generate \
  -H "Authorization: Bearer sk_live_..." \
  -H "Content-Type: application/json" \
  -d '{
    "title": "Why Cache Invalidation Is So Hard",
    "content_html": "<p>...your article HTML...</p>",
    "source_url": "https://yourblog.com/post-slug",
    "options": { "include_attribution": false }
  }'

# Returns:
# {
#   "success": true,
#   "svg_html": "<style>...</style><div>...3-view toggle + SVG...</div>",
#   "image_url": "https://cdn.startuphub.ai/...diagram.webp",
#   "markdown": "# Visual TL;DR, ...",
#   "jsonld": { "@type": "ImageObject", ... },
#   "nodes_count": 8,
#   "attribution_included": false
# }

FAQ

Why do AI agents prefer JSON diagrams over SVG or images?

A rendered SVG forces an LLM to OCR text and infer connections from coordinates. A JSON diagram with typed nodes and labeled edges is fully traversable in 200 tokens. Agents pipeline downstream reasoning, citation, and re-rendering from the JSON; the SVG is for the human reader at the end of the chain.

What does the JSON diagram format look like?

Two arrays: nodes (id, label, type one of cause/effect/result/key/concept, optional desc, essential flag) and edges (from, to, optional label verb). Plus a top-level title string. Same shape every call. No schema drift.

Can I use the Markdown output to feed another LLM?

Yes. The Markdown variant is structured as a concept list plus a flow list, which Claude, GPT, Gemini, and Llama all parse cleanly without prompt engineering. It is the format most agent-to-agent handoff pipelines settle on because it is human-readable, LLM-friendly, and lossless against the source JSON.

Is there an MCP server for Claude / Cursor / Windsurf?

Yes. The generate_visual_tldr tool ships in the same MCP server as our other agent tools (Search startups, Discover email, Agent readiness scan). One auth, drop-in for any model client that speaks MCP. Returns all four formats in the tool response.

How is this different from Mermaid or Graphviz?

Mermaid and Graphviz are diagram syntaxes you write by hand. Visual TL;DR is a pipeline: article in, structured diagram out, no manual authoring. The JSON shape we return is intentionally simpler than the underlying Mermaid AST so agents can read and write it without learning a DSL.

Is there a free tier?

Five diagrams per month, all four output formats, every diagram includes a "From startuphub.ai" footer. No credit card. Sign up returns a Bearer key immediately. Most editorial blogs land on Pro ($19/mo, 500 diagrams, optional attribution) once they decide to pin a diagram to every post.

How this differs from a diagram tool or a summariser

Three different jobs. A diagram tool draws what you tell it to draw. A summariser shortens prose. This one reads an article and returns its structure.

Mermaid or GraphvizText summarisersVisual TL;DR
What you give itSyntax you write by handAn articleAn article title and its HTML
What comes backA rendered diagramA block of textJSON, Markdown, SVG and JSON-LD from one call
StructureShapes and positions you choseProse, no graphTyped nodes, verb-labelled edges, and an essential flag on the ones that carry the argument
Small screensOne drawing, scaled downNot applicableTwo renders, 808px and 340px, laid out separately rather than shrunk
For an AI agentIt would have to read the pictureNot applicableA small JSON graph it can traverse directly

Get started

Add a diagram to your next article

Free tier ships in under a minute. Grab an API key or install the plugin and you're generating diagrams before your coffee's done.

Get free API keyDownload WordPress plugin

Other tools