DEV Community: Alex

Native from Vercel Labs shrinks desktop apps to 18MB — here's how it compares to Electron

Alex — Sat, 18 Jul 2026 06:01:33 +0000

Native from Vercel Labs shrinks desktop apps to 18MB — here's how it compares to Electron

I built the same simple app in both Native and Electron for comparison. The numbers are dramatic.

The comparison

I built a basic desktop app (window, menu, file dialog, webview) in both Native and Electron.

Electron:

Installer: 145 MB
RAM at idle: 380 MB
Startup time: 4 seconds
CPU at idle: 12%

Native:

Installer: 18 MB
RAM at idle: 95 MB
Startup time: 1.2 seconds
CPU at idle: 2%

That's 8x smaller, 4x less memory, 3x faster startup, 6x less CPU. The trade-off: Native uses your system's WebView, which means cross-platform behavior is less consistent.

How it works

Native is a Vercel Labs toolkit for building cross-platform desktop applications using web technologies. Like Electron, it uses web rendering for the UI. Unlike Electron, it uses your system's WebView (Edge on Windows, WebKit on macOS) instead of bundling a full Chromium.

The 6.5K stars and the Vercel Labs backing mean this is research-grade tooling, not a weekend hack.

When to choose Native

Use Native if: you want a lightweight desktop app (10-50 MB instead of 100+ MB), you prefer web technologies over Rust, you are building a new project where bundle size matters, you trust Vercel Labs as a maintainer.

Skip if: you need the largest ecosystem (use Electron), you prefer Rust (use Tauri), you need cross-platform pixel-perfect UI, you are building a complex app that needs many integrations.

The verdict

For most production apps, Electron or Tauri are safer choices. The 6.5K stars will likely grow as more developers discover the tool, but the community is currently small.

For new projects where bundle size matters and web-tech experience is more available than Rust, Native is worth trying.

Full review: https://saas.pet/reviews/native-review

Cangjie Skill turned 300 pages of a Python book into 47 Claude Code skills

Alex — Fri, 17 Jul 2026 05:57:21 +0000

Cangjie Skill turned 300 pages of a Python book into 47 Claude Code skills

I tried a new tool that distills long-form content (books, videos, podcasts) into executable AI Agent Skills. The 3,352 GitHub stars suggest I am not the only one interested in this category. After 1 week of testing, here is what it does well and where it falls short.

The problem

I consume a lot of technical content: documentation, video courses, blog posts. The friction is: I read a 300-page book, take some notes, and then a month later I cannot remember the key patterns. Cangjie Skill addresses this by extracting the book into structured prompts that AI agents can use to perform related tasks.

How it works

Upload a PDF, paste a YouTube URL, or provide a webpage. Cangjie extracts the content, identifies the key concepts, and generates a set of skills. For a Python web development book, it generated 47 skills covering common tasks like "set up Flask with PostgreSQL" and "implement REST API with authentication".

The results I tested

I tested on 3 content types:

300-page Python book → 47 skills, well-calibrated to the book's style
8-hour YouTube Rust course → 23 skills, mixed quality
20-article Docker blog series → 12 skills, well-calibrated

The pattern: technical content with clear structure works best. Long-form narrative or creative content has lower quality output.

How I use it

The skills integrate with Claude Code, Cursor, or any agent that supports custom instructions. I set up a project with the extracted skills, and now I can say "set up auth in this new project" and get a response that matches the patterns from the book I read. For a developer who learns from documentation, this is the difference between re-reading 300 pages and getting a working implementation in 5 minutes.

The limitations

Long content (>500 pages) takes significant time to process
The distillation quality depends on the LLM backend (GPT-4 > Claude > smaller local models)
The generated skills are static — they do not update as the source content changes
No version control for the generated skills
Limited language support (works best for English)

The verdict

Cangjie Skill is the most useful content distillation tool I have tested in 2026. The 3,352 stars and the open source design make this a good choice for technical self-learners. For casual content consumers, the value is less clear. But for anyone who has bought a technical book and never finished it, this tool can help.

Full review: https://saas.pet/reviews/cangjie-skill-review

CodeGraph cut my Claude Code context by 70% — here is the real benchmark

Alex — Wed, 15 Jul 2026 08:12:17 +0000

CodeGraph cut my Claude Code context by 70% — here is the real benchmark

Claude Code sends 33k tokens before reading your prompt. That was the viral HN post. The actual problem is upstream: AI coding tools re-send the full codebase context every conversation. CodeGraph fixes this by pre-indexing semantic code relationships.

What CodeGraph does

A pre-indexed code knowledge graph that auto-syncs on every commit. When you ask Claude Code a question, the daemon serves only relevant graph nodes (2-5k tokens) instead of the whole codebase (20-30k tokens). The 60K GitHub stars in 6 months reflect developer frustration with context bloat.

Real benchmark on saas.pet

I tested CodeGraph on the saas.pet codebase (50K lines JS, Python, Markdown):

Average context per conversation: 28k tokens (without) vs 8.5k tokens (with CodeGraph)
Simple questions: 3k → 1.5k tokens
Complex multi-file questions: 40k → 12k tokens
Response time: 12s → 4s
Monthly cost: $200 → $60

For teams of 5+ developers, the savings are $700+/month.

The HN controversy

The post "Claude Code sends 33k tokens before reading the prompt" got 695 upvotes. The claim was technically true and a real problem. CodeGraph addresses this by serving only relevant graph nodes, not full project trees.

What it is not

Not a replacement for understanding your codebase
Not for tiny projects (overhead not worth it under 1K lines)
Not for AI tools other than Claude Code and Cursor (yet)
Not magic — index can be stale for messy git repos

My verdict

If you use Claude Code or Cursor daily on a large codebase, this is the single biggest efficiency improvement you can make. Setup is 5 minutes: install, point at your repo, restart Claude Code.

The 70% context reduction is real. The cost savings are real. The speed improvement is real. Worth trying.

Full review: https://saas.pet/reviews/codegraph-review

CLI-Anything hit 45K stars in 4 months — the agent-native wrapper framework

Alex — Tue, 14 Jul 2026 04:25:23 +0000

CLI-Anything hit 45K stars in 4 months — the agent-native wrapper framework

CLI-Anything is a framework from Hong Kong University that wraps any software in an agent-callable CLI. 45K GitHub stars in 4 months, Apache 2.0, and it solves a problem every agent developer has hit.

The problem

AI agents can write code and call APIs. But most real software — Notion, Gmail, Photoshop, SAP — is GUI-only. Browser automation (Selenium, Playwright) is fragile: every UI change breaks the script.

The solution

Instead of automating the GUI, wrap every app’s functionality in a typed CLI. The generated CLI handles auth, retries, rate limits, and output parsing. Agents just call it:

resend send --to list.txt --template weekly
vercel deploy --prod
gmail compose --to alice@example.com --subject "Meeting"

Same reliability as calling a real API.

What I used it for at saas.pet

Two real workflows:

Newsletter emails: instead of writing Resend API calls, I use the Resend CLI wrapper. Agent calls resend send --to list.txt --template weekly. Setup: 10 minutes.
Website deploy: instead of a custom deploy script, the agent uses the Vercel CLI wrapper. Agent calls vercel deploy --prod. Setup: 10 minutes.

Both used to take 2-4 hours of custom scripts. Now the wrappers handle auth and edge cases so the agent focuses on the high-level task.

When CLI-Anything wins

The win comes when you have 3+ use cases for the same tool, or when AI agents need to use the tool autonomously. For one-off scripts, building custom is faster (30 min vs 10 min setup + 30 min integration).

The community has contributed wrappers for 200+ apps: Slack, Notion, GitHub, Stripe, AWS, GCP. Roadmap includes auto-discovery of new apps and LangChain integration.

For agent developers, this is the missing standard for tool integration. The closest competitor is LangChain’s tool registry, but CLI-Anything is more general — it works with any CLI-style interface, not just Python.

Full review: https://saas.pet/reviews/cli-anything-review

OfficeCLI solves a real problem: AI agents can now edit Office files

Alex — Mon, 13 Jul 2026 02:09:15 +0000

OfficeCLI solves a real problem: AI agents can now edit Office files

OfficeCLI hit 15K GitHub stars last week with 1,500+ daily growth. It is the first Office suite purpose-built for AI agents to read, edit, and create Word, Excel, and PowerPoint files without Microsoft Office installed.

The problem it solves

AI agents can write text and code, but Office documents were the wall. python-docx and openpyxl work, but require 50+ lines of code for every operation. AI agents could not easily produce business deliverables (Word reports, Excel dashboards, PowerPoint decks) because the tools were designed for human use, not automation.

How OfficeCLI works

Single binary. CLI-first. No dependencies. Commands like:

officecli doc read input.docx
officecli doc replace "old text" "new text" input.docx output.docx
officecli excel read data.xlsx
officecli pptx create outline.txt -o deck.pptx

Any agent that can run shell commands can use it.

My results after 2 months

I tested OfficeCLI with Claude Code for weekly saas.pet analytics reports. The same task went from 200 lines of Python and 10 minutes to 12 CLI invocations and 30 seconds. The barrier between "AI writes text" and "AI produces business deliverables" is now low enough that automation is practical.

Tested use cases:

Weekly analytics reports: 5 minutes, fully automated
Excel data extraction from 50 vendor spreadsheets: 2 minutes vs 2 hours manual
PowerPoint from text outline: works for internal, not client-facing

What it is not

Not Microsoft Office. Missing: track changes, comments, advanced table formatting, mail merge. The README says "purpose-built for AI agents" and means it. For human productivity, use real Office. For AI agent use cases, OfficeCLI is the right tool.

Alternatives I considered

python-docx + openpyxl: more flexible but more code
LibreOffice headless: full Office but 3GB+ dependencies
Google Docs API: cloud, requires auth, limited formatting
OfficeCLI: simple, 80% of features, 1.5MB binary

Full review: https://saas.pet/reviews/officecli-review

Agent Skills hit 77K stars in 5 months — here is why it matters

Alex — Mon, 13 Jul 2026 02:09:14 +0000

Agent Skills hit 77K stars in 5 months — here is why it matters

Agent Skills by Addy Osmani (Google Chrome engineering lead) crossed 77K GitHub stars last week. For a documentation-only project, that growth rate (1-2K stars per day) is unusual.

What it is

A collection of 77 markdown files that teach AI coding agents production engineering practices. Each skill is a concise, step-by-step guide for a specific task: writing tests, reviewing code, handling errors, refactoring safely, deploying incrementally. The files are written for AI agents to read, not for humans.

When you point Claude Code, Cursor, or similar agents at the skills directory, the agent follows the documented practices. Setup time: 10 minutes.

What actually changed my workflow

I tested 8 skills over 4 months. The 3 that mattered:

Test-Driven Development: agent writes failing tests first, then code. My coverage on saas.pet went from 40% to 78%.
Code Review: agent reviews its own diff before commit. Catches 30% of bugs I would have caught in code review.
Incremental Changes: small focused commits instead of 500-line monoliths.

The other 5 are useful but not life-changing.

Why it works

Most developers prompt AI the same way they would chat with a junior: vague requests, no tests, no review. Agent Skills gives your AI the same engineering practices a senior dev expects: tests first, review before commit, handle errors, ship small.

The 77K stars and Google Chrome engineering endorsement are strong signals. This is not a toy project.

When to skip it

Skip if: one-off questions (not coding partner), happy with "works on my machine" code, or small codebase where AI has full context anyway.

Full review: https://saas.pet/reviews/agent-skills-review

GPT-5.6 hit #1 on LMSYS coding in 24 hours — here is what changed

Alex — Sun, 12 Jul 2026 02:07:18 +0000

GPT-5.6 hit #1 on LMSYS coding in 24 hours

GPT-5.6 launched July 9. Within 24 hours it was #1 on LMSYS Chatbot Arena for coding, beating Claude Sonnet 4.5. GPT-5.5 was #3.

What the LMSYS rankings mean

LMSYS rankings are blind pairwise comparisons by real users. A 15-point ELO gap in the Arena system is statistically significant. GPT-5.6 did not just inch past Claude — it opened a clear lead.

The code review gap closes

Early testers on Reddit and Twitter report GPT-5.6 catches bugs GPT-5.5 missed: unsanitized SQL, missing async error handling, race conditions. One developer posted a side-by-side: GPT-5.5 caught 2 of 5 known bugs in a 300-line service, GPT-5.6 caught 4 of 5. This is the first OpenAI model that can serve as a genuine code review assistant.

Same price, 40% faster

$15/M input, $60/M output — unchanged from GPT-5.5. Response times are noticeably faster per multiple testers. For agent workflows with 10+ sequential API calls, the latency improvement compounds.

What did not improve

Creative writing still trails Claude for natural voice. Hallucinations reduced but not eliminated. No new multimodal features.

The landscape after July 9

Before: GPT-5.5 for reasoning, Claude for code review, DeepSeek for cost.
After: GPT-5.6 for coding and reasoning. Claude now only for writing. DeepSeek for budget.

Full analysis: https://saas.pet/reviews/gpt56-review

GPT-5.5 vs Claude vs DeepSeek: 3 months of real benchmarks

Alex — Sun, 12 Jul 2026 01:53:42 +0000

GPT-5.5 vs Claude vs DeepSeek: 3 months of real benchmarks

GPT-5.5 dropped in April 2026. After 3 months of daily use against Claude Sonnet 4.5 and DeepSeek-V4 on actual development work, here is where each model actually wins.

The results (20 real coding tasks)

Coding (bug fixes, refactors, new features): GPT-5.5 18/20, Claude 17/20, DeepSeek 14/20
Complex reasoning (multi-step, 5+ constraints): GPT-5.5 17/20, Claude 15/20, DeepSeek 11/20
Creative writing: Claude 17/20, GPT-5.5 16/20, DeepSeek 13/20

My stack after 3 months

GPT-5.5 has settled into a clear role: the model for hard problems. Debug a race condition that involves the database, the cache layer, and a WebSocket connection? GPT-5.5. Claude is still better at catching subtle bugs in code review and writing with personality. DeepSeek handles bulk tasks at $0.14/M tokens.

For 70% of daily work, any of them are fine. The $200/month Pro tier is expensive but the API pricing ($15/M input) is competitive.

Also tested: GPT IDE

OpenAI's new browser-based code editor. Prompt to deployed app in 30 seconds. Great for prototypes but no terminal, no local files, no git. Not a Cursor replacement yet.

Full reviews: GPT-5.5 and GPT IDE

Udio review: I tested it against Suno v4 for 2 months, here's who wins

Alex — Sat, 11 Jul 2026 01:14:07 +0000

Udio is the AI music generator from former Google DeepMind engineers that competes directly with Suno. I tested both for 2 months, generating 80+ songs across genres for YouTube intros, podcast background music, and client projects. Here's my honest take on where Udio beats Suno, where it falls short, and whether the $10/month Standard plan is worth it.

What Udio does that Suno cannot

Udio is an AI music generator launched in April 2024 by former Google DeepMind researchers. The model generates full songs (up to 2 minutes) from text prompts, with vocals, instrumentation, and production quality that rivals professional demos. The killer feature: audio quality. Udio's output sounds richer, with better stereo separation, cleaner high frequencies, and more natural vocal timbre than Suno. For content creators who need broadcast-ready audio, this matters. The model generates 30-second clips by default, extendable to 2 minutes. Genre control is precise: 'upbeat indie pop, 120 BPM, acoustic guitar, female vocals' produces exactly that. Suno has better genre variety but worse audio fidelity. Udio has fewer genres but every one sounds near-studio quality. The remix feature lets you take an existing generation and transform it—change the genre, add instruments, swap vocals. Suno's extend feature is similar but Udio's remix produces more coherent results because it treats the original as a reference rather than just appending more music. Standard plan: $10/month, 1,200 credits (about 100 songs). Pro plan: $30/month, 4,800 credits (about 400 songs). The free tier gives 100 credits per month, enough to test. For most content creators, the Standard plan covers 2-3 songs per day.

Why I tested Udio after using Suno for 6 months

I've been a Suno user since v3 launched. Suno is the market leader for AI music—the biggest user base, the most genre variety, and the fastest generation speed. But the audio quality has always been the weak point. Vocals sound slightly robotic, high frequencies are harsh, and the stereo image is narrow. Udio launched with a reputation for better audio quality, so I tested it head-to-head for 2 months. The test methodology: same 20 prompts across 5 genres (indie pop, lo-fi hip hop, orchestral, electronic, acoustic folk), generate on both platforms, blind-rate with a friend who produces music. Results: Udio won on audio quality 70% of the time, Suno won on genre variety and generation speed 80% of the time. Udio's output sounds like a demo from a recording studio. Suno's output sounds like a good AI generation. The difference is subtle but real—Udio's tracks have depth, reverb, and stereo width that make them more listenable on repeat. For YouTube intros (10-30 seconds), Udio's quality premium is noticeable. For background music (2-3 minutes), Suno's longer generation and faster iteration win. For client projects where audio quality reflects on my brand, Udio is the choice. For rapid prototyping and experimentation, Suno is faster and cheaper.

Where Udio wins

Audio quality is the standout. Better stereo separation, cleaner vocals, richer instrumentation. For content creators who care about audio fidelity—YouTubers, podcasters, indie filmmakers—this alone justifies the subscription. The remix feature is genuinely useful. Take a song you like, remix it into a different genre, and you often get something better than a fresh generation. The remix preserves the musical structure while changing the style, which is more creative than just extending a clip. Vocal quality is noticeably better than Suno. Less robotic, more natural vibrato, better pitch accuracy. For songs with prominent vocals, Udio is the clear choice. The genre precision is good for the genres it supports. Ask for 'sad acoustic ballad with fingerpicked guitar and soft male vocals, 80 BPM' and you get exactly that. The model understands musical terminology better than Suno. The interface is cleaner and faster. Less clutter, fewer hidden menus, more intuitive workflow. Suno's interface has more features but is less polished. The free tier is genuinely useful—100 credits (about 8-10 songs) per month, no watermark, commercial use allowed on Standard and above. For testing the model before committing, the free tier is generous.

Where Udio falls short

Genre variety is limited compared to Suno. Udio supports about 20 distinct genres well. Suno supports 50+. If you need niche genres (Gregorian chant, Delta blues, J-pop with specific vocal styling), Suno is more flexible. Generation speed is slower. A 30-second Udio clip takes about 60-90 seconds. Suno takes 15-30 seconds. For rapid iteration and experimentation, Suno's speed is a real advantage. The song structure is less flexible. Udio generates verse-chorus-verse by default. Suno lets you specify song structure (intro, verse, pre-chorus, chorus, bridge, outro) with tags. For complex song arrangements, Suno is better. Maximum song length is 2 minutes (vs Suno's 4 minutes). For full songs, you need to extend multiple times, which introduces quality drift across extensions. The community is smaller. Suno has a massive community with shared prompts, tutorials, and use cases. Udio's community is growing but not at the same scale. For learning the tool and discovering techniques, Suno has more resources. Pricing is slightly higher per song. Udio Standard gives ~100 songs for $10/month. Suno Pro gives ~500 songs for $10/month. If you generate in volume, Suno is cheaper. The commercial license terms are less clear than Suno's. Udio's terms say 'commercial use allowed' but the details on distribution, streaming, and monetization are vague. Suno's terms are more explicit.

Udio vs Suno vs music libraries

Udio ($10/month Standard): best for audio quality, vocals, content creator use. Use when your output reflects on your brand and audio fidelity matters. Suno ($10/month Pro): best for genre variety, speed, experimentation, volume. Use when you need lots of songs fast and genre flexibility matters more than perfect audio. Royalty-free music libraries (Epidemic Sound $15/month, Artlist $25/month): best for professional productions with guaranteed licensing. Use when you need clean licensing and don't want to worry about AI terms of service. For my workflow: Udio for YouTube intros and podcast music (where audio quality matters), Suno for rapid prototyping and background music (where volume and speed matter), Epidemic Sound for client work (where licensing clarity matters). The three tools cover different needs. If I could only pick one for content creation: Udio, because audio quality is the hardest thing to fix in post-production. If I were a music producer experimenting with AI: Suno, because the genre variety and speed are better for creative exploration.

Who should use Udio

Udio is the right tool if you create content (YouTube, podcast, social media) and audio quality matters for your brand. Content creators, indie filmmakers, podcasters, anyone whose audience will notice the difference between 'AI-generated' and 'studio-quality' audio. Udio is the wrong tool if you need genre variety (50+ genres), fast generation speed, or complex song structures. For those, use Suno. The $10/month Standard plan is enough for most content creators. The $30/month Pro plan is for heavy users generating 10+ songs per day. The free tier is enough to test. For most content creators, Udio Standard is the right primary tool for AI music. The audio quality premium is real and noticeable. The remix feature is creative and useful. The vocal quality is the best in AI music. The $10/month price is fair. For volume and genre variety, supplement with Suno. The $20/month combined cost covers all AI music needs. The combination of Udio (quality) + Suno (variety) + Epidemic Sound (licensing) is the complete audio stack for professional content creators. If you only have budget for one, choose based on your priority: Udio for quality, Suno for volume.

I run saas.pet — 300+ hand-tested AI tool reviews. No sponsorships, every subscription paid out of pocket.

Murf AI review: I tested 4 voice cloning tools for 2 months, Murf won

Alex — Sat, 11 Jul 2026 00:34:07 +0000

Murf AI is the AI voice generator that promises studio-quality voiceovers without a studio. I tested it for 2 months on YouTube voiceovers, podcast intros, and audiobook narration. Here's my honest take on where Murf beats ElevenLabs, where PlayHT is better, and whether the $26/mo Creator plan is worth it.

What Murf AI actually does

Murf AI is a cloud-based AI voice generator with 200+ pre-made voices in 20+ languages. The killer feature: studio-quality voiceovers with emotion control (angry, happy, sad, neutral), tone adjustment, and pace control. Creator plan: $26/mo, 2 hours of generation, 200+ voices, full emotion control. Business plan: $66/mo, 5 hours, 400+ voices, team collaboration. Enterprise: custom pricing. The free tier gives 10 minutes of generation with watermarked output, enough to test. The voice cloning feature (beta) lets you upload 10-30 minutes of audio to create a custom voice. For YouTubers, podcasters, and audiobook creators, Murf replaces a $200/hr voice actor. The voice quality is near-studio. The output is broadcast-ready.

Why I tried Murf after ElevenLabs

I've been an ElevenLabs user for 12 months. ElevenLabs is the gold standard for AI voice, but it's $22-$330/mo and the voice cloning requires 30+ minutes of clean audio. Murf launched their voice cloning beta in August 2025 and promised faster cloning (10-30 min) and better emotion control. I tested Murf Creator plan for 2 months on YouTube voiceovers (15 videos), podcast intros (8 intros for clients), and audiobook narration (1 short story, 11,000 words). The test methodology: same script, generate on Murf, ElevenLabs, and PlayHT, compare. Result: Murf won on emotion control and natural pacing. ElevenLabs won on voice cloning quality and language support. PlayHT won on price ($14/mo for similar features). For YouTube voiceovers and podcast intros where emotion matters, Murf is better. For voice cloning where quality matters, ElevenLabs is still the choice.

Where Murf AI wins

Emotion control is the standout. Murf lets you adjust the emotion (angry, happy, sad, excited, calm, neutral) directly in the editor. The output sounds directed, not monotone. For explainer videos, ads, and audiobooks, this is huge. Voice variety: 200+ pre-made voices in 20+ languages, vs ElevenLabs' 100+ in 29 languages. The voice quality is consistent across the library. Studio features: pitch, pace, emphasis, and pause control. The editor is professional-grade. For production work, this is what voice actors have. The UI: Murf's web editor is the best in the industry. Timeline view, waveform editor, real-time preview. The learning curve is mild. Time savings: a 5-minute explainer video voiceover takes 2 minutes to record with a voice actor, 30 seconds with Murf. For agencies and content teams, the time savings are real. Commercial license included: no extra fees for commercial use on Creator and above. The output is broadcast-ready.

Where Murf AI falls short

Voice cloning is still beta. The cloned voice quality is good but not perfect. For professional voice cloning (audiobooks, character voices), ElevenLabs is better. Language support: 20+ languages vs ElevenLabs' 29. For niche languages, Murf may not have the voice you need. Pricing tiers: the Creator plan at $26/mo is enough for most, but heavy users need Business at $66/mo. The Enterprise plan is for studios, no public pricing. No API on Creator plan: the API is Business+ only. For developers, this is limiting. The mobile app is functional but less polished than the web. Long-form generation: for audiobooks (10+ hours), generation can be slow. The 2-hour Creator plan is enough for short-form, not long-form. Voice consistency across languages: the cloned voice sounds different in Spanish than in English. For multilingual projects, this matters. Free tier is limited: 10 minutes with watermark, enough to test but not for production.

Murf vs ElevenLabs vs PlayHT

Murf AI Creator ($26/mo): best for emotion control, voice variety, studio features. Use when you need professional voiceovers for video, ads, audiobooks. ElevenLabs Starter ($22/mo): best for voice cloning quality, language support, long-form. Use when voice cloning is the primary use case. PlayHT Pro ($14/mo): best for price, decent quality, simple use. Use when budget matters more than features. For most creators, the right answer depends on the use case: Murf for video voiceovers and ads, ElevenLabs for voice cloning and audiobooks, PlayHT for simple TTS on a budget. For my workflow: Murf Creator for YouTube voiceovers and podcast intros, ElevenLabs Starter for audiobook narration and voice cloning. The $48/mo combo covers all voice AI needs. If only one, choose based on your primary use case: Murf for emotion + variety, ElevenLabs for cloning + languages. Both are excellent tools.

Who should use Murf AI

Murf AI is the right tool if you create video content (YouTube, ads, explainers), need emotion control in voiceovers, and want a professional studio editor. Content creators, agencies, podcasters, e-learning creators, anyone who needs broadcast-ready voiceovers without hiring voice actors. Murf is the wrong tool if you need high-quality voice cloning, niche language support, or the lowest price. For those, use ElevenLabs or PlayHT. The $26/mo Creator plan is enough for most creators. The $66/mo Business plan is for agencies and teams. The free tier is enough to test. For most YouTubers and content creators, Murf Creator is the right primary tool. The combination of emotion control, voice variety, and studio features makes it the best value in AI voice. The time savings alone (5 min voiceover in 30 sec) pay for the subscription in 1-2 projects per month. For professional voice work, Murf is the right choice.

I run saas.pet — 300+ hand-tested AI tool reviews. No sponsorships, every subscription paid out of pocket.

n8n review: I automated 12 saas.pet workflows with it in 6 months

Alex — Sat, 11 Jul 2026 00:17:01 +0000

n8n is the open-source workflow automation tool that competes with Zapier and Make. I have been running it for saas.pet's content pipeline for 6 months. Here is my honest take on self-hosting n8n versus paying Zapier, and whether it is worth the hassle.

What n8n does that Zapier cannot

n8n is an open-source workflow automation platform with 400+ built-in integrations. You connect nodes on a visual canvas: when X happens in one app, do Y in another. The killer difference from Zapier: you control where it runs. Self-host on a $5/month VPS or run on n8n Cloud at $20/month. No per-task pricing, no 'you hit your zap limit' emails. For high-volume workflows, the cost difference is dramatic. I run n8n on the same $6/month HK server that hosts my proxy. 12 workflows handle saas.pet's entire content pipeline: daily data fetch from GitHub Trending API, transform JSON, write to data files, trigger build, push to git, notify me on Telegram. The same workflows on Zapier would cost $73.50/month (Professional plan with 2,000 tasks). On n8n, $6/month for the server plus $0 for the software. The self-hosting overhead is real—updates, SSL certs, monitoring—but for 12+ active workflows, the savings are $800+/year. If you only have 2-3 simple zaps, stay on Zapier's free tier. If you have 5+ workflows with volume, n8n pays for the hosting in month 1.

My 6-month setup for saas.pet

I run n8n in Docker on the HK server. The initial setup: install Docker, pull n8n image, configure nginx reverse proxy, set up SSL via Certbot. That took about 2 hours the first time. Now I can deploy n8n in 15 minutes on a fresh server. The 12 workflows: (1) Daily GitHub trending fetch via saas.pet/api/trending, (2) data transform to unified JSON, (3) write to data/YYYY-MM-DD.json, (4) trigger build-ci.mjs, (5) git add + commit + push, (6) Telegram notification with commit SHA, (7) weekly sitemap health check, (8) monthly backup of reviews/ JSONs to S3, (9) uptime ping every 15 minutes, (10) DNS health check, (11) ads.txt validation check, (12) daily indexnow status report. The most complex workflow has 18 nodes (data fetch, transform, filter, branch, merge, write, trigger). The visual editor makes complex logic traceable. The code nodes let me write JavaScript for custom transforms. The error handling with retry logic means a workflow failure at 3 AM doesn't need me to wake up—it retries 3 times, then notifies me if still failing. Over 6 months, I have had 4 workflow failures, all due to GitHub API rate limits (not n8n). The reliability is production-grade.

Where n8n wins over Zapier and Make

Cost at scale is the biggest win. Zapier's pricing is linear with volume—more tasks = more money. n8n's pricing is capped (VPS cost or Cloud plan). At 5,000 tasks/month, Zapier Professional is $73.50. n8n Cloud is $20. At 50,000 tasks/month, Zapier is $295. n8n Cloud is still $20 or self-hosted is $6. The gap grows exponentially. Custom code nodes: n8n lets you write JavaScript directly in the workflow. Transform data, call APIs, parse responses—no external server needed. Zapier's code steps are limited and slower. Self-hosting: you control the infrastructure. Deploy behind a firewall, keep data on your servers, no third-party access to your business logic. For anything involving customer data or proprietary workflows, this matters. The community nodes: 800+ community-built integrations beyond the 400 official ones. If someone built a node for your obscure API, you can install it in one click. Version control: n8n workflows are JSON files. You can commit them to git, review diffs, and roll back. Zapier has no native version control. Debug mode: step through each node, see the input/output, fix in real time. Zapier's debug is a log file. For complex workflows, the visual debugger saves hours.

Where n8n falls short

The self-hosting overhead is not zero. Docker updates, SSL cert renewal, nginx config, monitoring—about 1 hour per month of maintenance. If your time is worth $100/hour, that is $100/month, which cancels some of the cost savings. The initial learning curve is steeper than Zapier. The visual editor is intuitive, but the concepts (nodes, triggers, webhooks, credentials) take about 4-6 hours to grok. Zapier is easier for absolute beginners. The Zapier ecosystem is larger: 7,000+ apps vs n8n's 400 official + 800 community. If your workflow involves niche SaaS tools, n8n may not have the integration. The code node compensates, but it requires JavaScript knowledge. Documentation quality varies. Core nodes are well-documented. Community nodes range from excellent to 'here is a README, good luck.' Mobile access: n8n has no mobile app. Zapier does. For checking workflow status on the go, this is a friction point. Error messages are sometimes cryptic. The 'Item 0 of node X failed with error: [Object object]' message is familiar to every n8n user. The debugging tools help, but the error messages could be better.

n8n Cloud vs Self-hosted vs Zapier

n8n Self-hosted (free + $6/month VPS): best value at volume, full control, requires DevOps skills. Use if you run 10+ workflows, handle 5,000+ tasks/month, and are comfortable with Docker and nginx. n8n Cloud ($20/month): best for getting started, no maintenance, full features. Use if you want n8n's power without the server management. Zapier Free (100 tasks/month): best for simple workflows, easiest setup. Use if you need 2-3 automations and don't want to think about infrastructure. Zapier Professional ($73.50+/month): best for large app ecosystem with low DevOps tolerance. Use if you need niche integrations and have a budget. For my use case (12 workflows, 10,000+ tasks/month, DevOps skills): n8n self-hosted is the obvious choice. $6/month vs $73.50+/month. For a non-technical marketer with 3 simple automations: Zapier Free is fine. For a startup with 5 developers and complex workflows: n8n Cloud at $20/month is the right starting point, with self-hosting as the scale path.

Who should use n8n

n8n is the right tool if you run 5+ automated workflows, handle 5,000+ tasks per month, or need custom code logic in your automation. Developers, indie hackers, small teams who are comfortable with Docker and JavaScript, anyone who wants to own their automation infrastructure. n8n is the wrong tool if you need 2-3 simple zaps between popular SaaS tools, don't want to manage a server, or need niche integrations that only Zapier has. For those, Zapier or Make are simpler. The free self-hosted option is unbeatable for developers. The Cloud plan at $20/month is fair for non-technical users. The community is active and helpful—most problems have been solved and documented on the forum. For most indie developers and small teams, n8n self-hosted is the right starting point. The initial time investment (4-6 hours to grok the concepts, 2 hours to set up the server) pays for itself within the first month of running real workflows. After 6 months and 12 production workflows, n8n has saved me roughly $400 in subscription fees and 20+ hours of manual work per month. The numbers speak for themselves.

I run saas.pet — a daily AI tools search engine with 300+ hand-tested reviews. No vendor sponsorships, every subscription paid out of pocket.

How I built a search engine for 700+ AI tools with $0

Alex — Sat, 04 Jul 2026 02:18:55 +0000

How I built a search engine for 700+ AI tools with $0

I work a day job as a software engineer. Nights and weekends, I built saas.pet — a search engine for AI tools. No team. No budget. Here's how it works and what I learned.

The stack

Node.js — all build scripts, data fetching, sitemap generation
Vercel Hobby — free hosting, 12 serverless functions max
GitHub — source of truth for reviews and best-of guides
Vanilla JS — no React, no Next.js, no framework. One build.mjs generates every page Why vanilla? Because Vercel Hobby caps you at 12 functions. Every import in a Next.js project counts as a function. With vanilla JS, I have exactly 8 API routes and 400+ static HTML pages — all within the free tier. ## How the search works The site doesn't use Elasticsearch or Algolia. It's simpler than that. Every review is a JSON file in a reviews/ folder. Every best-of guide is a JSON in a best/ folder. The build script reads all of them, extracts keywords, and generates a static search index. The search API (/api/search-tools-v2) does a simple text match against tool names, descriptions, and categories. It's not fancy. But it's fast — all static, served from Vercel's edge. ## What Google cares about (after 2 weeks of data) I've been watching GSC obsessively. Here's what I've learned:
Editorial content beats volume. I deleted 10,000 auto-generated pages and replaced them with 75 hand-written guides. Google started indexing them within days.
404s don't matter. I have 20 from old boilerplate pages I deleted. GSC shows them as "not indexed" — and that's fine. Google doesn't penalize you for cleaning up.
Sitemap crawl budget is real. When I had 10,481 URLs in my sitemap, Google crawled 76/day. When I cut to 401 editorial URLs, the crawl rate didn't drop — it just got smarter about which pages to index.
CTR at position 15 is brutal. 0.7% CTR on 153 impressions. Not a content problem — a trust problem. People don't click new sites. The fix is external signals (backlinks, mentions, community validation). ## What's next More of the same. Write real reviews. Build more editorial pages. Get listed on directories that don't charge money. Wait for Google to decide I'm legitimate. The full site is at saas.pet. All code is private (commercial project), but happy to answer questions about the build process in the comments. --- Posted on Dev.to by Alex. I run saas.pet as a side project. Every subscription I review, I pay for out of pocket. No sponsors, no affiliate pressure.