On vibe coding

Drugs.

The ‘60s had LSD. The ‘70s, heroin, the ‘80s, cocaine, and the ‘90s, meth. (Don’t forget 2018, when schools removed bathroom doors as kids couldn’t stop Snoop-Dogging on vape sticks.)

I’ve been clean my whole life. Actually, that’s a lie. I was clean until last summer, when I first tried the Cline extension in VS Code. In that moment, I became chronically addicted.

The experience was incredible. It thought, navigated my files like I would, came up with a plan, and then implemented it right before my eyes.

I experienced the vibe coder’s revolution in waves. Today, I’ll break them down, highlight notable companies, and demonstrate what sets each code-gen tool apart. Not all AI coding companies are the same.

Allow me to introduce a scorecard: the “Changed My Engineering Life Score” (CMELS). I'll use this rigorously scientific CMELS to capture just how dramatically each trend or tool reshaped my approach to building. Then, we’ll intuit what differentiates AI coding platforms and make some predictions.

Before we start

My favorite sponsors, including today’s, are RTF fans who hit reply.

It's hard to find a bulletin with more dollar value than in RTF. The 28k audience is made purely of founders and investors, with a 70% open rate (!) and 2% CTR.

Ad spots are backlogged, so if you're interested in sponsoring Return the Fund, please fill out this form instead and I'll get back to you! :)

Today’s edition is brought to you by my new friends at Rime, one of the top emerging voice AI companies I mentioned last month. I reached out to them and they were happy to sponsor today’s edition, so please show them some love by pressing play on their incredibly human-sounding voices.

Last month, when discussing voice AI, I stressed the importance of a model’s relatability—when speaking to it over the phone, it should feel natural. Rime makes your voice calls natural, plain and simple.

They’ve built a fleet of voices nearly indistinguishable from a human’s. The best part? They work directly with VAPI, LiveKit, Voiceflow, and other top platforms, so you can take advantage of A+ voices with a few clicks.

Developers can leverage Rime’s AI voices with just a few lines of code. I tried it—it was ridiculously easy, and shockingly good.

At the very least, if your curiousity is piqued—head to their beautiful landing page and hit the big play button to hear one of the voices in action.

The Copilot Moment

When GitHub Copilot first arrived in VS Code, I was truly mind-blown. It’s safe to say that it left every engineer similarly astonished. (VS Code is to engineers what Excel is to accountants.)

So much of programming is boilerplate repetition. Copilot lives inside your editor, suggesting ways to finish coding every line. The idea was genius, given the predictability of most code—define an input somewhere above, use it somewhere below, and so on.

Importantly, I never trusted it to solve problems for me. Copilot is like power steering in a car. You still have to drive the thing.

CMELS: 9/10

GitHub Copilot Labs

With Microsoft owning VS Code, GitHub, and a considerable portion of OpenAI, they had every infinity stone: a data moat, a technological advantage, an abundance of resources, consolidated talent pools, and distribution amongst engineers. (Not to mention, engineers’ trust.)

They trained an OpenAI model on GitHub-hosted code, humanity’s most extensive database of software. Back then, OpenAI’s was the only intelligence capable of powering a service as deeply integrated as Copilot.

Shortly after releasing Copilot, GitHub introduced Copilot Labs for VS Code. The main addition was a ChatGPT-like interface in VS Code’s sidebar, allowing engineers to ask contextual questions about their code.

Copilot Labs solved two problems: (i) ChatGPT not knowing what your code does until you show it, and (ii) engineers having to frequently switch between VS Code and ChatGPT. You might think the second one is ridiculous, but the smallest annoyances compound over time, I promise you.

Copilot Labs is notable for proving that users value chat interfaces inside their IDEs (interactive development environment—their VS Code). But at the end of the day, that’s all it was.

CMELS: 3/10

Cline

In its earliest days, Cline was “Claude Dev”—an open-source VS Code extension built on Claude 3.5 Sonnet, Anthropic’s greatest leap in improving model intelligence and instruction-following.

Claude Dev was an AI agent. Through tools, it could read, create, and edit files. That was enough. Claude Dev’s philosophy was to assist already-competent engineers, not replace them. It would ask permission before ever reading or modifying a file.

To use Claude Dev, you had to supply an Anthropic API key. There was no monetization strategy. Now, after experiencing tremendous growth, Claude Dev is known as Cline. It’s no longer dependent on Anthropic and has a hosted, self-serve onboarding option.

Cline was the first AI coding agent I trusted to make changes to my code, because, I had to approve everything. I’d find myself suggesting modifications and adjusting diffs myself before approving them. As powerful as Cline was, seeing how many requests required oversight taught me to trust AI cautiously. For now, at least. :)

What I love most about Cline is their shipping velocity. They support newly released models within hours, including model architecture innovations like reasoning, extended thinking, prompt caching, and predicted outputs. Cline can use your browser to test web apps and research documentation. Pretty slick!

I used Cline for over a year, until Augment came for the throne.

CMELS: 6/10

Augment

My deepest frustration with AI coders is their naïveté. When starting a new task, I have to re-explain my software structure, present relevant files, and guard against changes noncompliant with my codebase.

Augment, a unicorn with over $250 million raised, was founded on the principle that contextual awareness leads to performant code generation. They didn’t just build another coding agent; rather, a holistic Context Engine for code repositories. Augment indexes your codebase, then offers Copilot-like completions along with an agentic chat experience.

The code completions are second to none. Because Augment is VS Code-native, it’s vertically integrated and makes onboarding easy. It suggests snippets I’d actually write, because I probably did in the past. It understands how I’ve structured various software components, and uses that understanding to make precise incremental suggestions in real time.

Augment also built a proper coding agent into their VS Code extension, leveraging contextual awareness to complete tasks end-to-end. It’s phenomenally powerful and more accurate than any other coding agent I’ve ever used. Touché, Augment.

Recently, the team announced Remote Agent—an important trend I’ll cover later on. Similar to OpenAI Codex (next up), users can spawn independent AI workers to complete tasks in the cloud, even if you close your laptop. (RIP interns.)

I’m bullish on Augment and a fan of their team. I think their philosophy on AI-assisted development is spot on. Later on, I’ll discuss the genius of their business model; as if the AI stuff isn’t enough.

CMELS: 8/10

OpenAI Codex

On May 16th, OpenAI put out a press release implying a paradigm shift for software engineers. They released Codex, a hands-off coding agent akin to Augment’s Remote Agent.

You connect a GitHub repository, write out a task, and hit “go.” Codex handles the rest. Relative to Augment’s agent, I’ve found Codex to operate like a brute. It uses an incredibly fast model (likely GPT 4.1 nano) and reads files incessantly using terminal commands, which isn’t efficient. Augment (and Cline) have agents with file operation tools, reading and writing in targeted chunks.

With that said, I value Codex living in ChatGPT. Because Codex supports iOS, I can trigger, monitor, and end coding tasks on the go from my phone. This, alone, keeps me returning to Codex. I’m sure other platforms will catch up, but OpenAI is certainly flexing their ChatGPT distribution advantage with Codex.

Simple but annoying tasks are perfect for Codex, especially when on the go.

CMELS: 5/10

Lovable

I, along with my team, spend way too much time in Lovable. At one point, we were paying for 1,000 credits per month. Lovable is built for web applications, and boy, does it do those well.

What makes Lovable so prolific? What specifically helped them reach $10M ARR in 60 days, raise $15M, and fly through $50M ARR in 6 months?

I don’t think it’s their AI. Lovable builds web apps very well; but, it’s built on Claude, like nearly every other code-gen platform. To Lovable’s credit, they’ve masterfully orchestrated Claude through AI tooling, error detection, self-reflection, and reasoning. But Lovable’s AI stack doesn’t explain their explosive product-led growth.

My theory: it’s just so damn easy to use. Anyone can use it. Cursor is an IDE, Augment is for engineers, Codex works on GitHub repositories, and Cline is a VS Code extension. You don’t need to know any of those words to use Lovable.

I use it incessantly to build landing pages and internal tools. Designers use it to bring their mockups to life, and founders use it to generate interactive prototypes.

Lovable isn’t directly comparable to Augment or Cursor; rather, to bolt.new and same.new. It exists in a class of something-from-nothing, English-only vibe coding tools designed to be usable by literally anyone.

It’s beautiful, and incredibly empowering to the masses. I don’t think Lovable threatens engineers. At least, not good ones. It helps them build quicker, and forces them to level up. If you care about engineering and have a growth mindset, Lovable is a blessing.

Lovable is no exception to the adage “ask and you shall receive.” But asking is uncharacteristically hard. The best Lovable apps are born of careful attention to detail and specific iterative guidance. I’ll often go as far as giving it the full source code for web components, pulled from UI libraries.

I give Lovable an 8/10 because of its accessibility and how often I use it.

CMELS: 8/10

(If, and only if you want to try Lovable yourself, use this link to save me some $$ on those monthly credits)

Cursor

The golden child of this space, Cursor rocketed its way to $500M ARR and just raised a whopping $900M. They were very early to build a brand around empowering engineers with AI. Born as a fork of VS Code, Cursor built Tab, an autocomplete engine, and a coding agent that helps engineers ship faster.

To be frank, I don’t have much to say about Cursor. I tried it, and didn’t love it, but that’s just me. I know too many engineers who can’t live without it, and that’s enough.

I’ve found a groove with Augment and Cline in VS Code. Not only is this setup incredibly cost effective (Augment starts free, Cline uses your own API key, and VS Code is open source), it’s given me the best success. I’ll discuss why momentarily.

I have nothing bad to say about Cursor as a product or as a company. I do think it’s overly romanticized especially by non-engineers, but that goes for any product entering its market-propelled hype cycle. I’m concerned about their unit economics and staying power, and wouldn’t invest myself.

With that said, they’ve recently expanded beyond the AI IDE label, releasing background agents and a code reviewer. It’s a TAM + moat play, telling the world “we’re not just a fork whose greatest enemy is time.”

In principle, I hope they continue empowering the propagation of useful software.

CMELS: 0/10

CodeRabbit

Harjot and Gur started CodeRabbit in 2023 to leverage AI in a different layer of the software deployment stack.

Code reviews are an incredibly important function of any software team. Without them, engineering leaders lack the confidence that their software is stable, well-designed, compliant with design principles, and resilient to harmful edge cases.

Unfortunately, in terms of engineering hours, code reviews are expensive. And, as with any human-led process, catastrophic issues disguised as small bugs frequently fly under the radar.

The beauty of CodeRabbit is its ease of use. Connect it to a GitHub organization, and it’ll automatically review any pull requests (significant code changes requiring approval). CodeRabbit’s most notable competitor is Graphite, who recently raised a $52M Series B.

CMELS: 5/10

Notable Mentions

Windsurf

On May 6, OpenAI announced their acquisition of off-white Cursor for $3B. Windsurf and Cursor are both forks of VS Code, the IDE engineers use and love most. These forked IDEs are founded on the belief that VS Code itself will not fulfill engineers’ desire for AI tools. Thus, they leverage VS Code’s open-source nature and say, “I’ll do it for you.”

We got trigger happy with forks. Continue Inc. forked VS Code to build continue.dev, an IDE with customizable AI. They were then backed by YC. Along came PearAI, who forked continue.dev, and themselves were backed by YC. Understandably, PearAI received considerable backlash. Their response was to separate their Continue fork into a “submodule” and open source their code.

The whole “AI IDE” industry is a mess, and not one I want to participate in.

So that’s Windsurf. Microsoft owns VS Code and Github. Microsoft partners with OpenAI to train an early model on GitHub data. Microsoft purchases rights to 49% of OpenAI’s profits. Windsurf forks Microsoft’s VS Code, which OpenAI then acquires for $3B.

Until… Microsoft cracked down on VS Code forks and started shipping agentic coding capabilities directly in VS Code. 😵

Devin

They said it’s the first fully-autonomous software engineer. I tried it, many times, and thought it was significantly worse than OpenAI Codex, Augment’s Remote Agent, and even Cline running on a local VS Code folder. Uninterested.

Reflections

Not all vibe coding tools are the same. Not all of the companies mentioned today are in competition.

Not only are there different types of AI code assistance, but spectrums of vibiness. Hear me out.

Augment is somewhat vibey. It’s built to “augment” human engineers, empowering them with AI. It’s a voice in your ear, a copilot, a mentor giving you suggestions.

Lovable, on the other hand, is extremely vibey. It’s all vibes. Just tell it what you want to build, in plain English, and it never shows you a line of code.

There’s no correlation between vibiness and product efficacy. Just different target customers. My two most-used AI tools are both Lovable and Augment—each shines in different situations. Lovable is limited to web development, and struggles with complex logic. It conforms to typical website styles, and loves to introduce technical debt all over the place. Great for landing pages, demo websites, dashboards, etc. but horrible for data, processing, custom UI components, and unpopular frameworks.

Augment is helpful when I’m fixing bugs or writing features in an existing codebase. Because there are thousands of foundational commits, it refrains from reinventing the wheel; rather, understands my intent and assists as I go about my business. The approach is entirely different.

Lovable and Augment have different philosophies, different business models, different customers, different GTM magnets, and different KPIs.

Augment’s winning model

Instead of forking VS Code, Augment focused on empowering engineers where they already are: VS Code itself. I appreciate that they’re helping engineers level up, not strike out (it’s in the name). They built intelligence, not UI. The Augment Context Engine is pretty incredible, powering both inline completions and the Augment Agent.

Sophisticated codebases are like spiderwebs, with long threads individually weaving together to make maintainable, reliable, extendable software. Thus, when working with components in a new file, it’s critical to understand where they came from and why they were built as they were. This isn’t just RAG… It’s much more complicated than semantic relevance.

Code assistants’ processing + UX speed is of paramount importance. I’ve been told I type like I’m being chased by a hyena. Augment is quick enough to keep up, making suggestions I can accept with a keypress without losing flow.

Unlike Cursor, Augment is free. They targeted the enterprise, with an enhanced developer plan of $50/mo for 12x the messages limit, team management, and shared context between team members. Augment’s Context Engine is designed to span across many repositories, so software team members can build with the synergy of each other’s work.

I’m a big fan.

Cline’s alternative approach

It’s worth mentioning that contextual awareness is not the only approach to performant code generation. Cline has no inherent codebase context when agentically completing tasks. Instead, it has a local directory tree (list of all files) and selectively bounces from file to file—like a scavenger hunt. Their philosophy is that humans onboard this way, starting at an entrypoint and following threads through the codebase.

I love Cline and respect this approach, but I don’t think it’s right for real-time inference. Cline gets away with it because they don’t play the inline completions game, as Augment does. They can afford to sacrifice latency.

A lucrative opportunity

We’ve seen the history of AI-assisted coding tools, how revolutionary each one was, the intricate dynamics between big tech and unicorns, the key differentiators between platforms, winning business models, and more.

To wrap, let’s look at a potentially massive market empowered by advancements in code-generation.

You’ve probably heard of Retool, the $3B+ internal apps unicorn.

Retool is an incredible product. It’s beautiful, a joy to use, and built with an attention to detail that unassumingly shines through in convoluted use cases.

Companies love Retool not only for its ease of building simple internal apps, but for its strong access management controls, making it safely deployable within large enterprises. This company was built to be mission-critical for startups and Fortune 500s alike.

Retool organizations can share databases, internal APIs, documentation, data secured by fine-grained authentication, and shared services. In this way, it’s easy for a company to onboard once and build later, without stressing about provisioning API keys or managing external data access. Retool wraps it all.

What I’m missing with a naïve system like Lovable is the ability to prompt, “make a dashboard for each sales team members to track their performance in real time.” Lovable would appear to start without a care; but, with boilerplate/hardcoded data… Eventually, I’ll have to build an API authenticated into Stripe and a CRM. Lovable will suggest connecting to Supabase, which seems pleasant if it weren’t for your internal data securely warehoused in an authenticated internal bucket—where it should be.

Retool, however, already has these resources configured, shared, authenticated, and locked down. Retool and Lovable need to make a baby. There’s too much synergy for them not to.

I’ve seen some new companies attempting this: Clark, Odapt (YC), and Vybe.

With Retool venturing into AI agents and workflow automation, I’d love to see them get their hands dirty with the vibiest of code-gen tools, built upon their strong foundation of data, security, and enterprise trust.

Anyway. I hope you enjoyed this edition of “Thoughts on Drugs,” written by an addict.

PS: If you want to chat professionally about technical diligence or prospecting in this space, hit reply.

Now, let’s talk about investing in vibe coding companies. (If you’re not a Startup Navigator, see ya next time.)

Subscribe to Startup Navigator to read the rest.