DEV Community: Ilias Ism

Building Outputmaxxing Remote MCP Servers with a 2-Step Spend Confirmation Gate

Ilias Ism — Sat, 07 Nov 2026 15:55:59 +0000

Most Model Context Protocol (MCP) servers shipped in the last year have a structural flaw: they stop at detection.

They wrap GET endpoints, dump a 20,000-token JSON blob of analytics or search results into your context window, and then force the human to open a browser tab, click through a 6-step wizard, and finish the job manually.

Internally, we call the opposite philosophy Outputmaxxing: every tool surface—especially an MCP server—must end in a concrete output or state transition, not just a diagnostic report.

This week, we shipped and upgraded 4 remote Streamable HTTP MCP servers across our product fleet so developers and founders in Claude Code, Cursor, and Codex can go from discovery to execution in a single chat session:

LinkDR (Awesome Tools listing) — 14 tools for domain authority, Google top-100 SERPs, AI search citations, in-chat publisher site approvals, and ordering backlinks.
AI SEO Tracker (Awesome Tools listing) — 12 tools for ChatGPT, Perplexity, Gemini, Copilot & Google AI Overview visibility tracking and missing competitor citation discovery.
GenPPT (Awesome Tools listing) — 8 tools to turn prompts, URLs, or PDFs into approved slide outlines, rendered 16:9 visuals, and native .pptx / .pdf exports.
Awesome Tools — 7 keyless public tools (+ 16 admin tools) to search software directories, submit a SaaS tool, and verify backlinks.

Here is the engineering breakdown of how we architected remote Streamable HTTP MCP servers that safely handle mutations and real payments—plus the production traps we hit along the way.

1. Designing for "Outputmaxxing": Read Tools vs. Write Loops

When we audited how developers actually used our v1 MCP servers, we noticed that read-only tools (get_domains, aiseo_list_sources) were always followed by the user asking: "Okay, now add these 15 prompts to my project" or "Approve those two DR 50 publishers and order 3 more links."

To close the loop across all four products, we paired every diagnostic read with a state-mutating write tool:

Product	Remote MCP Endpoint	Read / Diagnostic Tools	Output / Action Tools
LinkDR	`https://linkdr.com/mcp`	`get_domains`, `search_domains`, `search_keywords`, `get_keyword_serp`, `get_my_ai_answers`, `list_my_pending_approvals`	`create_my_project`, `update_my_project_strategy`, `decide_my_site_approval`, `order_backlinks`
AI SEO Tracker	`https://aiseotracker.com/mcp`	`aiseo_get_scan`, `aiseo_list_sources`, `aiseo_list_missing_sources`, `aiseo_list_snippets`	`aiseo_create_project`, `aiseo_add_prompts`, `aiseo_activate_plan`
GenPPT	`https://genppt.com/mcp`	`genppt_account_status`, `genppt_list_decks`, `genppt_get_deck`	`genppt_create_deck`, `genppt_generate_slides`, `genppt_export_deck`, `genppt_upgrade_plan`
Awesome Tools	`https://awesome.tools/mcp`	`search_tools`, `get_tool`, `list_categories`, `list_tags`	`submit_tool`, `verify_my_backlink`, `submit_feedback`

For example, in AI SEO Tracker, aiseo_list_missing_sources queries every third-party URL cited by ChatGPT, Perplexity, Gemini, Copilot, and Google AI Overviews where your tracked competitors appear and your domain is absent, sorted by totalPromptsWithoutYourSite. You can take that exact list of high-priority URLs and feed it straight into LinkDR's MCP tools in the same workspace.

2. The 2-Step `confirmPaymentSpend` Safety Pattern

The hardest design problem in agentic commerce is preventing accidental spend without forcing the user out of the terminal.

If a tool immediately charges a saved card on the first call, a hallucinated parameter or eager agent loop can trigger an unwanted $500 invoice. Conversely, if the tool only returns a generic pricing page URL, the agent workflow is broken.

We standardized a 2-step confirmPaymentSpend protocol across order_backlinks (LinkDR), aiseo_activate_plan (AI SEO Tracker), genppt_upgrade_plan (GenPPT), and submit_tool (Awesome Tools):

Step 1: Zero-Side-Effect Quote Preview (`confirmPaymentSpend: false`)

By default, confirmPaymentSpend is false. Calling the tool in preview mode performs zero billing mutations:

It does not create unpaid database drafts that pollute abandoned-cart crons or burn daily IP rate limits.
It calculates the exact itemized price (totalUsd), checks if the authenticated workspace has a usable payment method on file (brand, last4), and builds a pre-filled 1-click checkout fallback URL.

{
  "status": "preview",
  "requiresConfirmation": true,
  "quote": {
    "package": "3x DR 40+ Authority Backlinks",
    "subtotalUsd": 720,
    "discountUsd": 25,
    "totalUsd": 695
  },
  "savedCard": {
    "available": true,
    "brand": "visa",
    "last4": "4242"
  },
  "checkoutUrl": "https://linkdr.com/agent/checkout?project=acme&pack=dr40x3&code=AGENTIC",
  "nextStep": "Show the user the itemized quote ($695.00 USD on Visa ••4242) and re-call order_backlinks with confirmPaymentSpend: true and expectedTotalUsd: 695 after explicit approval."
}

Step 2: Price-Locked Execution (`confirmPaymentSpend: true` + `expectedTotalUsd`)

When the human replies "Looks good, place the order", the agent calls the tool a second time with:

confirmPaymentSpend: true
expectedTotalUsd: 695

On the server, we assert Math.abs(calculatedTotalUsd - expectedTotalUsd) < 0.01 so the price cannot drift between preview and execution.

If the user has a saved Stripe card on file (LinkDR one-time orders), we create and pay a Stripe Invoice off-session, mark the order paid inline (and idempotently via invoice.payment_succeeded), and protect savedCardInvoiceId rows from 24-hour abandoned-checkout cleanup.
If the card requires 3D Secure (authentication_required), or if the user is on a keyless/hosted checkout flow (Polar on AI SEO Tracker, Stripe Checkout on GenPPT and Awesome Tools), the server cleanly hands back the hosted payment URL.

3. Dual-Mode Public + Admin Routing on a Single `/mcp` Endpoint

For Awesome Tools (https://awesome.tools/mcp), we wanted any founder to be able to add https://awesome.tools/mcp to Cursor or Claude Code with zero API key setup to search the directory, submit their product, or verify their backlink—while keeping our 16 operator tools (upsert_tool, review_intake, capture_tool_media) on the same server.

We solved this at the JSON-RPC dispatcher layer:

No Authorization header: The server exposes the 7 public founder tools (search_tools, get_tool, list_categories, list_tags, submit_tool, verify_my_backlink, submit_feedback).
Authorization: Bearer $MCP_API_KEY: The server unlocks all 16 admin tools.
Idempotent Draft Reuse: If a founder previews a paid submission ($49 Featured or $149 Sponsor) with confirmPaymentSpend: false, the server returns the price card without reserving the domain or emailing the submitter. When they confirm, it reuses any existing pending draft for that domain+email pair without burning their 3/day IP submission limit, while blocking cross-submitter domain hijacks.

4. Production Gotchas We Hit (So You Don't Have To)

Gotcha 1: Cloudflare `403 error code: 1010` on Headless MCP Clients

Shortly after deploying https://genppt.com/mcp and https://linkdr.com/mcp, we noticed that curl and Node fetch succeeded with 200 OK, but Python-based MCP clients and registry validators (User-Agent: Python-urllib/3.12) were rejected at the Cloudflare edge with 403 Forbidden (error code: 1010).

Why: Cloudflare's Browser Integrity Check (browser_check: on) blocks default Python urllib and certain headless HTTP user agents before the request ever reaches your origin container.

The Fix: Disable browser_check via the Cloudflare Zone Settings API across every domain hosting a public /mcp endpoint:

curl -X PATCH "https://api.cloudflare.com/client/v4/zones/$ZONE_ID/settings/browser_check" \
  -H "Authorization: Bearer $CF_API_TOKEN" \
  -H "Content-Type: application/json" \
  --data '{"value":"off"}'

Gotcha 2: Key Rotation Race Conditions (`SELECT ... FOR UPDATE`)

In GenPPT (https://genppt.com/mcp), users can generate and rotate gppt_ MCP API keys from /app/account (POST /api/account/mcp-key). Storing raw tokens in Postgres is a non-starter, so we store SHA-256 hashes and wrap key rotation in a SELECT ... FOR UPDATE transaction lock on the user row so concurrent clicks can never leave two active MCP keys or orphan a revocation. We also added automatic gppt_, bzly_, and Bearer token redaction inside every *_submit_feedback tool so agents cannot accidentally leak credentials into logs or Slack alerts.

Gotcha 3: Stateless Streamable HTTP & JSON-RPC Edge Cases

Many MCP clients probe resources/list and prompts/list immediately after initialize and notifications/initialized. Make sure your Streamable HTTP route:

Returns 202 Accepted (with an empty body) on notifications/*.
Returns empty arrays ({ resources: [] }, { prompts: [] }) or a clean JSON-RPC -32601 with HTTP 200 rather than a hard HTTP 500 or 404.
Enforces a strict request body cap (e.g., 64 KiB) on public endpoints.

5. Publishing Remote Servers to the Official MCP Registry (`/.well-known/mcp-registry-auth`)

When publishing remote Streamable HTTP servers (without an npm or PyPI package) to the Official MCP Registry (registry.modelcontextprotocol.io) under reverse-DNS namespaces like com.linkdr/mcp or com.genppt/mcp, you can prove domain ownership over HTTP using an Ed25519 signature:

Generate a 32-byte Ed25519 keypair.
Serve the public key at https://<your-domain>/.well-known/mcp-registry-auth as plain text:

   v=MCPv1; k=ed25519; p=<base64-encoded-32-byte-public-key>

Authenticate with mcp-publisher login http --domain=<your-domain> --private-key=<hex-seed> and publish your server.json pointing to your streamable-http remote URL.

Try the 4 MCP Servers in Claude Code or Cursor

All four remote Streamable HTTP MCP servers are live today:

{
  "mcpServers": {
    "linkdr": {
      "url": "https://linkdr.com/mcp"
    },
    "aiseotracker": {
      "url": "https://aiseotracker.com/mcp"
    },
    "genppt": {
      "url": "https://genppt.com/mcp",
      "headers": {
        "Authorization": "Bearer gppt_your_api_key"
      }
    },
    "awesome-tools": {
      "url": "https://awesome.tools/mcp"
    }
  }
}

You can also browse all of our MCP-enabled tools under the new MCP tag on Awesome Tools.

Building a Zero-Watermark Browser Meme & GIF Generator With Perspective Slots

Ilias Ism — Wed, 04 Nov 2026 12:15:37 +0000

When you open most legacy meme generators on a phone or desktop browser, you get hit with three things: ad networks shifting the layout, low-resolution 500px exports stamped with a mandatory corner watermark, and rigid top/bottom Impact text boxes that don't match how modern multi-panel memes actually work.

While rebuilding MagicMeme (a free browser meme template library and GIF maker with 2,300+ templates), we redesigned the canvas engine around how multi-panel and image-slot templates are actually used in 2026.

1. Multi-Panel & Image-Capable Slots (Example: "They're The Same Picture")

Take a classic two-panel comparison format like They're The Same Picture ("Corporate Wants You To Find The Difference").

In a naive meme generator, you can only type top and bottom Impact text over the whole image. In reality, people use They're The Same Picture in two different ways:

Text labels on the two white paper cards: Typing two things that look different on paper but are identical in practice directly onto the Left Picture and Right Picture panels.
Pasting or dropping two images: Pasting two screenshots or product logos directly into the angled paper polygons on Pam Beesly's desk.

In MagicMeme's canvas editor, each template defines typed panel slots (text or image-capable with perspective polygon bounds). You can click directly on the canvas or type in the sidebar form to label the Left and Right panels, or press ⌘V / drag-and-drop photos straight into the slots, switch between Auto, Impact, Dark Text, and Subtitle typography, or click Try Example to cycle through sample captions.

2. Clean 2x Retina Exports & Direct PNG Clipboard Copy

Three small browser engineering details made the biggest difference in daily use:

Separating Editor UI Chrome from the Export DOM: On-canvas selection rings, drag handles, and slot outlines are tagged with data-editor-ui="true" and filtered out during the html-to-image render pass so your downloaded PNG or JPG never contains editor outlines.
2x Pixel Ratio with Zero Watermark: Every export renders at pixelRatio: 2 in full resolution with zero watermarks.
Writing image/png to navigator.clipboard: Modern browsers reject image/jpeg blobs in ClipboardItem. Rendering the copy pass as image/png lets you click Copy Image and paste the finished meme straight into Slack, Discord, X, or iMessage in one keystroke.

You can browse all 2,300+ blank templates and test the live editor on MagicMeme or jump straight into They're The Same Picture Meme Generator.

What Makes an AI Headshot Look Real vs. Fake in 2026: Focal Length & Skin Texture

Ilias Ism — Wed, 04 Nov 2026 12:13:22 +0000

If you have browsed LinkedIn recently, you can probably spot a bad AI headshot in half a second: plastic skin with zero pore texture, uncanny studio rim-lighting that doesn't match the background, or a suit jacket whose lapels melt into the collar.

While building Headshot Generator AI, we compared the output and requirements of the leading tools in the space (documented in our 2026 AI Headshot Generator Comparison: HeadshotPro vs Aragon vs BetterPic).

Here is what actually separates a natural, camera-authentic portrait from an obvious AI filter, and how to get clean results from just 1 to 3 phone selfies.

1. Why 20-Selfie Batch Fine-Tuning Often Hurts Identity Preservation

The first generation of AI headshot tools (like Secta Labs and BetterPic) required users to upload 12 to 20+ photos and wait 1 to 2 hours while a custom LoRA/DreamBooth model trained in the background, then dumped 100 random photos into a gallery.

That batch approach has two major flaws:

Garbage-in averaging: Most people do not have 20 well-lit, recent photos of themselves on their phone. To hit the 20-photo minimum, users upload old group crops, sunglasses shots, or wide-angle bathroom mirror selfies. Training on mixed focal lengths averages out your facial bone structure.
Zero per-shot art direction: Waiting two hours for 100 random shots means you cannot see one portrait, adjust the background or wardrobe, and iterate in a minute.

On Headshot Generator AI, you only need 1 to 3 clear phone selfies, and each portrait renders in about 60 seconds at up to 4K resolution. New accounts start with 5 free credits (no credit card required) on the Free AI Headshot Generator and Free AI Profile Picture Maker.

2. The 4 Optical Rules That Make a Portrait Pass the "Real Camera" Test

As we cover in our Professional Headshots Guide, a believable headshot follows four physical camera rules:

85mm–105mm Portrait Focal Length: Phone front cameras shoot at ~24mm wide-angle, which stretches your nose and narrows your ears when held at arm's length. A real studio headshot uses an 85mm lens from 6 to 8 feet away so facial proportions stay flat and natural.
Micro-Contrast & Skin Texture: Real studio strobe lighting preserves subtle skin texture, catchlights in the eyes, and natural flyaway hairs rather than airbrushing the face into a wax mannequin.
Crop Survival (4:5 to 40px Circle): A LinkedIn or Slack avatar is displayed at 40×40 pixels next to comments. Your head and shoulders need to fill 60–70% of the frame with a clean headshot background so your face still reads at thumbnail size.
Natural Posing & Wardrobe: Slight 15-degree shoulder angles and solid mid-tone layers always beat stiff passport mugshots (see our guides on Headshot Poses and What to Wear for Headshots).

3. Beyond LinkedIn: Seasonal & Social Portrait Workflows

Because single-style generation takes ~60 seconds from 1 to 3 selfies instead of a 2-hour training job, you can use the exact same selfie set across different contexts:

LinkedIn, Resume & Executive Bios: Clean studio and office light via Professional Headshots and Profile Picture Maker.
Holiday Cards & Festive Portraits: Turn a phone selfie into fireside, snowy market, or vintage holiday card portraits in a minute with the Free AI Christmas Photo Generator.
Dating Profile Audits & Candids: Score your current photos and generate natural candid replacements with Rate My Dating Profile and AI Dating Photos.

How We Built an AI Pitch Deck Generator That Never Hallucinates Startup Metrics

Ilias Ism — Wed, 04 Nov 2026 12:12:35 +0000

Almost every "AI presentation maker" demo on social media uses the exact same five-word prompt:

"Make a 10-slide presentation about content marketing."

Nobody at work or in a startup builds presentations like that.

When a founder is raising a Seed round, or an agency lead is preparing a quarterly business review (QBR), you don't start from a five-word topic. You start from a website URL, a 10-page PDF report, a traction memo, or a financial model with real numbers.

And when you feed real documents into traditional template-based AI slide generators, two things break immediately:

Metric Hallucination: The LLM invents a TAM figure, rounds your $1.32M Q3 revenue into a made-up percentage, or drops your unit economics.
Fixed-Box Text Overflow: Template fillers try to stuff AI-generated bullets into rigid PowerPoint text boxes. As soon as a headline is seven words instead of four, the text overflows into the chart, and you spend two hours nudging boxes around at 11 PM.

Here is how we solved both problems across Pitches.ai (our AI pitch deck generator for startups) and GenPPT (our general AI presentation maker).

1. Grounding the Outline in Source Documents Before Rendering a Single Pixel

In our 2026 AI Presentation Maker Benchmark, we fed the exact same 1-page Q3 financial report PDF ($1,321,000 Q3 revenue, 9,650 active subscribers, 3.1% net churn, 41% gross margin) into 20 different AI slide tools.

The tools that jumped straight from prompt to rendered slides either dropped half the numbers or broke the layout on export.

Instead of one-shot generation, Pitches.ai and GenPPT split deck creation into a three-stage approval pipeline:

Stage 1: Source Extraction & Narrative Sequencing

When you paste a startup URL or attach a PDF/memo on Pitches.ai, the extraction pass pulls your exact figures and maps them into the proven 12-slide investor sequence (modeled on real fundraising decks like the Airbnb Pitch Deck Template, Uber Pitch Deck Template, and Tesla Pitch Deck Template):

Cover & One-Line Value Prop
The Problem (Quantified pain)
The Solution
Product / How It Works
Why Now
Market Size (TAM / SAM / SOM grounded in your source or live research)
Competition & Moat
Business Model & Pricing
Traction & Unit Economics (100% locked to your uploaded figures)
Go-To-Market
Team
The Ask & Use of Funds

Because you review and edit the slide-by-slide outline in plain text before any visuals render, you can verify every metric in 30 seconds.

Stage 2: Visual Concept Sheet Approval

Before generating a 12-slide deck, the studio renders a Visual Concept Sheet so you can pick the typography, color palette, and art direction upfront.

Stage 3: Full-Slide Visual Rendering (Zero Text-Box Overflow)

Instead of stuffing text into fragile PowerPoint shapes that shift when opened on another machine, each slide is rendered as a complete, cohesive visual with charts, diagrams, and hierarchy baked in, and exports cleanly to PowerPoint (.pptx) and PDF. If one slide needs a tweak, you tell the in-studio chat agent to revise just that slide without touching the other eleven.

2. Free Tools for Fast Document-to-Slide Handoffs

Not every workflow needs a full studio session. We also shipped standalone utilities for the three most common jobs:

Pitches.ai Free Pitch Deck Builder: Generating and iterating on your investor pitch deck in chat is 100% free (you only pay if you decide to export the final PowerPoint or PDF).
Famous Startup Pitch Deck Examples: Browse slide-by-slide teardowns and free PDF downloads of the original decks used by Airbnb, Uber, Dropbox, and Tesla.
Free PDF to PowerPoint Converter: Drop any PDF report or lecture handout to extract a structured slide outline and preview slides with no account required.
AI Presentation Maker for Teachers: Turn lesson plans and curriculum PDFs into classroom-ready slide decks.

We Analyzed 10,000+ Direct Publisher Backlink Quotes in 2026: What Real Links Cost

Ilias Ism — Wed, 04 Nov 2026 12:11:32 +0000

Over the last year at LinkDR, we logged and audited 10,000+ direct publisher quotes across B2B SaaS, developer tools, e-commerce, and local service campaigns.

Most discussions about SEO backlinks online fall into one of two extremes: agencies claiming you need a $10,000/month digital PR retainer to get a single link, or Fiverr/reseller spreadsheets selling "DR 50+ backlinks" for $25.

When you look at what real website owners and editors actually quote when you email them directly, and cross-reference those domains against Ahrefs organic traffic and AI search citations (ChatGPT, Perplexity, and Google AI Overviews), the data tells a very clear story.

1. The 43% Zero-Traffic Trap on Reseller Spreadsheets

The single biggest waste of SEO budget in 2026 is buying backlinks based on Domain Rating (DR) or Domain Authority (DA) alone.

When we audited thousands of DR 30+ domains pitched by link brokers and public guest-post spreadsheets:

43% of DR 30+ domains selling links had fewer than 1,000 monthly organic visits from Google in Ahrefs.
Many had inflated their DR using 301 redirect loops or cross-linked blog networks while ranking for zero real keywords.
After Google's recent core and spam updates, links from zero-traffic "general news" guest-post farms pass virtually zero ranking weight.

The rule we enforce on every campaign: Before approving any domain, check that it gets at least 1,000+ monthly organic visits in Ahrefs from keywords that match its actual topic.

2. What Real Publishers Actually Charge in 2026 (By DR Tier)

When you filter out zero-traffic link farms and only look at legitimate niche publications and SaaS blogs with real organic search traffic, almost every editor charges an upfront editorial or link-insertion fee to cover their team's time.

Across 10,000+ direct publisher quotes, here are the real median publisher fees:

Domain Rating (Ahrefs DR)	Organic Traffic Floor	Median Direct Publisher Fee	Typical Agency Retainer Markup	LinkDR Per-Link Price
DR 30–39	1,000+ visits/mo	$85 – $100	$250 – $350	$160 / link
DR 40–49	2,000+ visits/mo	$100 – $135	$350 – $450	$280 / link
DR 50–59	5,000+ visits/mo	$140 – $220	$500 – $750	$400 / link
DR 60+	10,000+ visits/mo	$250 – $395+	$750 – $1,500	$500 / link

If someone offers you a "DR 60 backlink for $30," you are not getting a deal. You are buying a link on a scraped or de-indexed domain. Conversely, if an agency charges you $800 for a standard DR 40 niche edit, you are paying a 4x retainer markup.

3. Why SaaS Link Building Now Controls AI Search (ChatGPT & Google AI Overviews)

For SaaS link building, the goal of a backlink is no longer just passing PageRank to your homepage.

When a buyer asks ChatGPT, Perplexity, or Google AI Overviews "What is the best [category] tool for startups?", the model does not invent recommendations from thin air. It runs a retrieval pass over the top-10 ranking Google comparison listicles and industry roundups and synthesizes the tools named in those articles.

If your competitors are listed in 8 of the top 10 Google comparison articles and your product is in 0, LLMs will rarely recommend you, even if your own documentation is well optimized.

4. The 3-Step Link Building Stack for Technical Founders

If you are launching or scaling a software product in 2026, run your off-page SEO in three stages:

Stage 1: Lock in Foundational Directory Backlinks (Day 1)

Before spending a dollar on outreach, submit your product to curated software and AI directories that carry real Domain Rating and send early-adopter referral traffic. We maintain a free, filterable list of 99+ Best AI Tool & Startup Directories for Backlinks sorted by DR, traffic, and free vs. paid submission links.

Stage 2: Competitor Link-Intersect & Direct Editor Outreach

Export the referring domains linking to 2 or more of your top ranking competitors in Ahrefs (filtering to dofollow, in-content text links on sites with 1,000+ organic visits).

If you want to run the outreach yourself, use our Free Domain Email Finder to look up verified editor and marketing contacts by domain, ranked by role and confidence score, and send a short two-line email asking for their insertion or editorial rate.

Stage 3: Managed Per-Link Placements With Domain Approval

When you would rather spend your engineering hours on product than chasing publisher PayPal invoices and negotiating insertion rates, you can order per-link placements on LinkDR (High-DR Backlinks and Link Building Services):

Automated competitor-backlink and Page-1 SERP outreach
Strict organic traffic floors in Ahrefs
Optional Site Approval so you inspect and approve every domain before placement
60-day delivery refund guarantee and 12-month free link replacement

Why Speech-Only YouTube Transcripts Fail Coding Tutorials (And How Video OCR Fixes It)

Ilias Ism — Wed, 04 Nov 2026 12:11:31 +0000

When we looked at the last 30 days of analytics on Scribiz, our video-to-text tool, 37.4% of all traffic came from ChatGPT (758 visitors from chatgpt.com vs 310 from Google Search).

Even more interesting: 65% of those ChatGPT referrals landed on a single subpage: scribiz.com/video-ocr (495 visitors, 1,301 pageviews). One of our first Pro subscribers clicked a citation inside ChatGPT, landed on /pricing, signed in, and paid within 60 seconds, then immediately ran /video-ocr and /video-to-text.

Why is ChatGPT sending hundreds of developers and researchers to a Video OCR page? Because speech-only transcripts miss half of what happens in a technical video.

The Problem: Speech Transcripts Miss the Code, Slides, and Terminal Output

If you watch a 20-minute coding tutorial, a university lecture, or a product architecture walkthrough on YouTube, the speaker rarely reads every line of code or every slide formula out loud. They say:

"As you can see in this config block here, we set these three flags and run the command..."

If you only fetch YouTube's caption track (or run Whisper on the audio), your LLM prompt or Obsidian note gets the spoken words and zero of the terminal commands, config keys, benchmark tables, or burned-in TikTok/Reel captions.

Even worse, as we documented in Can ChatGPT or Claude Watch a YouTube Video?, neither ChatGPT nor Claude natively watches YouTube video frames when you paste a watch URL into chat. A YouTube watch page HTML only contains the title and description; the spoken words sit in a separate timedtext track, and the visual slides sit inside the video stream.

How Two-Layer Video Extraction Works (Audio + Scene-Deduped Frame OCR)

To give an engineer or an AI agent the complete context of a video, you have to separate "video to text" into two distinct layers and align them on timestamps:

Layer 1: Spoken Words (Captions First, Audio Second)

Public YouTube videos with existing captions: Use the YouTube Transcript Generator (or Download YouTube Transcript for SRT, VTT, TXT, Markdown, or JSON). If captions already exist, reading them takes a fraction of a second.
YouTube videos with no captions (or missing "Show transcript" button): When a creator disables captions or the Show transcript button is missing on YouTube, you have to transcribe the audio track directly via Transcribe YouTube Video.
Short-form video (TikTok & Instagram Reels): Most short-form clips have no downloadable caption track at all. The Free TikTok Transcript Generator extracts both the spoken audio and the text burned onto the video.

Layer 2: Scene-Deduped Video OCR (`/video-ocr`)

You cannot run OCR on every frame of a 30 fps video (a 10-minute video has 18,000 frames, which would produce thousands of duplicate copies of the same slide).

Instead, a clean Video OCR pipeline works in four steps:

Detect scene changes: Sample frames when the visual layout changes (a new slide, a terminal scroll, a code file switch).
Perceptual hash & deduplicate: Drop near-identical frames where only the speaker's face moved in the corner webcam box.
Extract structured on-screen text: Read the slide headings, code blocks, terminal commands, and chart labels from each unique scene.
Align by timestamp: Merge the timestamped scene notes with the spoken transcript so you can see what was said alongside what was shown at 02:14.

Wiring Video Context Into Claude Code, Cursor, and ChatGPT via MCP or CLI

Instead of manually copying and pasting transcripts between browser tabs, you can give your coding agent direct access to video context in two ways:

1. Remote MCP Server (`https://scribiz.com/mcp`)

As detailed on the Scribiz Video MCP page, you can add the hosted MCP server to Cursor, Claude Desktop, or Claude Code with zero local dependencies:

{
  "mcpServers": {
    "scribiz": {
      "url": "https://scribiz.com/mcp"
    }
  }
}

Once connected, your agent gets get_video_context, ask_video, search_video, and get_transcript. With a Scribiz API key, passing watch: true instructs the server to read the video's picture as well, returning timestamped on-screen slides and code next to the spoken transcript.

2. Command-Line Tool (`npx scribiz`)

If you prefer working from the terminal or scripting batch runs in CI:

# Inspect a video URL or local media file from the terminal
npx scribiz context "https://www.youtube.com/watch?v=jNQXAC9IVRw"

# Ask a specific question with timestamp citations
npx scribiz ask "https://www.youtube.com/watch?v=jNQXAC9IVRw" "What commands were shown on screen?"

Which Tool Should You Pick for Your Workflow?

We recently published side-by-side comparisons against the most common tools in this space:

Tactiq Alternative: Pick Tactiq if you want a browser extension sitting inside live Google Meet or Zoom calls. Pick Scribiz when you have a recorded YouTube, TikTok, or video file and need the transcript, on-screen OCR, summary, and chapters.
NoteGPT Alternative: Pick NoteGPT if you want flashcards and mind maps in a study suite. Pick Scribiz when you want clean SRT, VTT, Markdown, or JSON output (plus on-screen slide OCR) for your own notes or AI agents.
HappyScribe Alternative & Eightify Alternative: Pick HappyScribe for human-proofread multi-language subtitles or Eightify for quick in-browser YouTube bullet summaries; pick Scribiz when you need both speech and screen text from a link or file.

You can test the Video OCR tool or YouTube Transcript Generator in the browser with no sign-up required for your first video.

I Tested 20 AI Presentation Makers With the Same PDF Report (2026 Benchmark)

Ilias Ism — Tue, 03 Nov 2026 17:04:19 +0000

Most "best AI presentation maker" roundups test each tool with a 5-word prompt like "make a presentation about content marketing."

Nobody at work builds slides like that. In practice, you start with a dense PDF report, an engineering RFC, a strategy memo, or meeting notes full of real numbers—and if the AI hallucinates one metric or stuffs 80 words into a fixed template box that overflows on .pptx export, you end up rebuilding the deck by hand at 11 PM.

So we ran the exact same 1-page Q3 financial report PDF (Fernhollow Coffee Co. Q3 2026 Review — $1,321,000 Q3 revenue across July $412k, August $438k, September $471k, 9,650 subscribers, 3.1% churn, 41% gross margin, and 3 operational risks) through 20 AI presentation tools and graded each one on:

Numeric & factual fidelity (how many of the 8 specific figures survived without hallucination)
Layout & typography (whether headlines and charts collide when a title is 7 words instead of 4)
PowerPoint (.pptx) & PDF export (whether the exported file actually looks like what you saw in the browser)

Summary of the Top Contenders

Tool	Best For	8-Fact Accuracy	Export Format	Where It Breaks
GenPPT	Turning a PDF report, brief, or outline into a designed PowerPoint or PDF deck	8 / 8	`.pptx` & PDF	Full-slide visual render means you edit via slide revision rather than dragging individual shapes
Gamma	Fast scrollable web read-aheads (`~45s` draft)	7 / 8	Web link, `.pptx`, PDF	Multi-column card layouts often shift or stack awkwardly when exported to `.pptx`
Plus AI	Working inside a locked corporate Google Slides or PowerPoint slide master	8 / 8	Native Google Slides / `.pptx`	Slower generation (`~2 min`) and basic box-and-bullet visual layouts
Canva Magic Design	Quick short visual decks with team brand assets	4 / 8	`.pptx`, PDF, Canva	Truncates long prompts and doesn't parse multi-page data PDFs cleanly
Pitches.ai	Investor & fundraising pitch decks	8 / 8	`.pptx` & PDF	Purpose-built for narrative/investor pitch structure rather than general classroom or internal training decks

Why Template-Box Fillers Struggle With Real Documents

Traditional slide generators take a fixed template (say, a 3-card comparison slide with a 32-character title box) and ask an LLM to fill the slots. As soon as your real data has a 9-word title or a 5-row table, the text overflows the bounding box or collides with the chart legend.

With GenPPT, we took a three-step approach instead:

Outline Lock First: It extracts the figures from your PDF or notes into a slide-by-slide outline you review and edit before a single pixel renders.
Visual Concept Sheet: It generates a visual concept sheet so you pick the typography, palette, and aesthetic direction upfront.
Full-Slide Visual Render: Each slide is rendered as a complete designed visual—with the July/August/September revenue bars, callouts, and typography composed together—and exported straight to PowerPoint or PDF.

Try the Benchmark Yourself

Full 20-Tool Comparison + Downloadable Test PDF + Generated Slides: genppt.com/blog/best-ai-powerpoint-generators
Free PDF-to-PowerPoint Converter (no signup required): genppt.com/tools/pdf-to-powerpoint

Close Your MacBook Lid Without Killing Your AI Agents (Open-Source macOS Keep-Awake)

Ilias Ism — Tue, 03 Nov 2026 17:03:52 +0000

Every time you close your MacBook lid on battery or unplug your display to walk to a meeting, macOS enters clamshell sleep. Your SSH sessions drop, your Docker builds freeze, and your background AI coding agents die mid-run.

caffeinate -dims only prevents idle sleep while the lid is open. To keep a MacBook running with the lid closed, macOS requires toggling pmset -a disablesleep 1 at the kernel power management layer—which is risky if you forget it on and throw your laptop into a backpack.

So I built and open-sourced Awake (il.ly/awake) — a tiny native macOS 26 Control Center widget, Menu Bar timer, and CLI (awake / asleep) that keeps your MacBook running with the lid closed, locks the screen the instant the lid shuts, and automatically re-enables sleep if the battery temp hits 40°C or drops to 10%.

What It Does

Works with the lid closed on battery or AC: Combines pmset -a disablesleep 1 with caffeinate -dims and live IOKit AppleClamshellState monitoring.
Instant screen lock on lid close: The moment the lid closes while Awake is active, it locks your session so an awake MacBook is never an unlocked MacBook.
Backpack thermal & battery circuit breaker: Polls AppleSmartBattery every 30 seconds in-process via IORegistry. If battery temperature reaches >= 40°C (like inside a closed sleeve or backpack) or charge drops to <= 10%, Awake immediately trips the breaker and lets the Mac sleep, then auto-resumes once cooled or plugged in.
macOS 26 Control Center + Menu Bar + CLI: Toggle from Control Center, pick a countdown timer (30m, 1h, 2h, 4h, or indefinite) from the Menu Bar espresso cup, or wrap a long command from the terminal:

# Install via script or Homebrew
curl -fsSL https://il.ly/awake/install.sh | sh
# or: brew install --cask illyism/awake/awake

# Keep awake for 2 hours
awake 2h

# Keep awake until a specific command finishes
awake -- bun run build