I've been thinking about the small stuff lately. Not the big model releases or the enterprise deals — the small, weird, personal tools that developers build because something in their workflow annoyed them enough to fix it. lil agents is one of those tools, and I keep coming back to it because the design choices Ryan Stephen made feel like a position statement, even if that wasn't the intent.
Here's the basic thing: lil agents is a free, open-source macOS app that puts two animated characters — Bruce and Jazz — on your dock. They walk back and forth. You click one, a popover terminal opens, and you're talking to Claude, Gemini, OpenAI Codex, or GitHub Copilot. No browser tab. No switching contexts. No account to log into. The AI is just... there, on your desktop, waiting. When it's thinking, little bubbles appear with phrases like "Pondering the cosmos..." When it's done, a sound plays. It has four color themes: Peach, Midnight, Cloud, Moss.
It has 1,300+ stars on GitHub and I think the reaction it gets — genuine delight, not just "cool project" — tells you something.
The Clippy joke is doing a lot of work here
Everyone who writes about lil agents mentions Clippy. I get it, the visual parallel is obvious. Animated character on your desktop, connected to something that's supposed to help you. But Clippy is remembered as a failure of a specific kind: it interrupted you, it assumed you needed help when you didn't, and it had no off switch that felt respectful. The joke landed because everyone had been annoyed by it.
lil agents inverts that. You go to it. It doesn't come to you. Bruce and Jazz walk around on the dock doing their thing, and they stay there until you click. That's a meaningful design decision, even if it sounds small. A lot of AI tools right now are competing to be more present, more proactive, more anticipatory. Clicky, for instance, is an AI that sits next to your cursor and can see your screen — it's watching, waiting to jump in. That might be exactly what some people want. But it's a fundamentally different relationship than "I'll be here when you need me."
Ryan Stephen talked about this in an interview — asking himself "what's the most playful way to interact with AI?" and thinking about AI as material to explore, not just a service to consume. That framing matters. When you think of AI as material, you start asking different questions about form. You're not just asking "how do I make this more powerful?" You're asking "how does this feel to use? What kind of presence does it have?"
The privacy architecture is a design choice, not just a feature
Here's the part I find most interesting, and it's easy to miss if you're just looking at the cute characters. lil agents runs entirely locally. No user accounts. No analytics. No data collection. Your AI interactions go through whichever CLI you've installed — Claude Code, Gemini CLI, whatever — and those interactions are governed by that provider's policies, not by anything lil agents itself is tracking.
In a category where most tools are racing to give AI agents with more access to more of your data, this is a real editorial stance. The Future of Privacy Forum has written about the data protection challenges that come with AI agents that have persistent access to your environment — the more an agent knows about you, the more useful it can be, but also the more exposure you carry. lil agents essentially says: we're not in that business. You get the interface. Your data stays between you and the model provider you already chose.
That costs something in features. lil agents can't learn your preferences over time the way a more data-hungry tool could. It can't anticipate what you're about to ask. It can't build a profile of your work patterns. Whether that trade-off is worth it depends on what you're trying to do — but the fact that Ryan Stephen made it explicit, in the architecture rather than just in a privacy policy, is worth noticing.
I'll be honest: I usually don't think hard about this stuff until something goes wrong somewhere and it's in the news. lil agents made me think about it proactively, which is the better time to do it.
What "AI as material" actually means for how we build tools
The quote from Ryan Stephen that I keep coming back to is the one about "AI as material." Most of the conversation about AI tools right now is about capability — what can it do, how fast, how accurately. The interface is treated as a wrapper around the capability. You build the thing that works, then you make it look okay.
lil agents flips that priority, or at least balances it differently. The interface is the argument. The animated characters aren't decoration on top of a terminal — they're making a claim about what it should feel like to have AI in your workspace. Present but not intrusive. Useful but not surveilling. Playful without being condescending.
That's a harder design problem than it sounds. The Forbes piece on gamification in the workplace raises a real concern: when you add delight and personality to tools, you can tip into manipulation, into making people feel more attached to a tool than they should be, into engineering dependency rather than genuine usefulness. I don't think lil agents crosses that line — the characters are charming, not addictive, and the whole thing is opt-in in a deep way — but the concern is real and worth keeping in mind as this category grows.
The question Ryan Stephen seems to be asking is whether you can make AI feel human-scale without making it feel manipulative. Whether presence can mean "available" rather than "watching." That's not a solved problem. lil agents is one attempt at it, and the fact that it's open-source under an MIT license means other developers can look at the attempt, learn from it, fork it, and try their own answer.
Why this matters for creators and freelancers specifically
Most of us aren't developers. I'm not. But the tools that developers build for themselves usually show up in our workflows six to eighteen months later, either directly or as influence on the products we end up using. lil agents is worth paying attention to now because it's working through questions that are going to affect everyone who uses AI in their work.
The AI moving from destinations to ambient presence is already happening. Claude has a desktop app now. Gemini is baked into Google Workspace. Microsoft Copilot is embedded in Windows. The browser tab model is getting replaced by something more ambient. The question is what values get embedded in that ambient layer — who it serves, what it watches, how it behaves when you're not actively prompting it.
lil agents is a small, free tool made by one developer who wanted something that felt less like infrastructure. It has two tiny animated characters and four color themes and a sound that plays when your query is done. And it's asking better questions about the future of AI interfaces than most of the venture-backed products in the same space.
That seems worth paying attention to. You can find it at lilagents.xyz or check out the lil agents GitHub repo—the code is public and right there if you want to see how it works.
Generated Images
Seven variants below — three standard compositions, one documentary (foreground bokeh), and three dynamic-angle "spatial" compositions for parallax video. To request a fix on any one, add a checkbox under## Image Touch-upslike:- [ ] spatial-square: remove the random hand on the right
landscape — 1920×1080
