OpenSAM, or a 5,000-line answer to slop

Last week everyone was running openclaw and posting about Moltbook. The MOLT token did something like +1,800% in a day, which is exactly the kind of number that makes the whole scene look very serious and very stupid at the same time.

I shipped OpenSAM instead.

what it is

About 5,000 lines of Rust. A Telegram-native agent you can stand up on a Raspberry Pi or the cheapest VPS you can rent. OpenRouter for models, so the ongoing cost is a few dollars a month. One-liner install: you run a command, answer a few prompts, and you have a 24/7 agent with persistent memory. That’s the whole product.

The code is roughly 99% lighter than the frameworks everyone’s been benchmarking this winter. Not lighter in a hand-waving sense. Lighter in the sense that the binary actually fits on the hardware you already own, and the running cost is the price of a coffee, not a cloud bill.

what it’s against

The framing I used when I shipped it was blunt: “AGI right now is AI slop with human prompting influence.” That’s not a shot at openclaw specifically. Peter’s thing is genuinely interesting. It’s a shot at the genre: the 100-skill mega-agents, the agent-to-agent social networks, the demos that require a beefy machine and a paid API plan just to wave at you from the terminal.

Moltbook launched four days before this. A social network for AI agents, which is a sentence that sounds profound the first time and hollow by the third. The MOLT token ripped. Agents were following each other and posting into a feed. The vibe was: look at the economy we’re building.

Meanwhile the actual economy is down 50% from the October high and most of these “autonomous agents” are running on human rails the whole time. Someone prompts them, curates their outputs, reroutes them when they drift. The autonomy is mostly performed.

That’s the slop. Not the code. The pretense.

My fuck you to all of it was building something that doesn’t need the pretense. OpenSAM does real work on real hardware. It doesn’t need a token (no new token; it runs for $SAM). It doesn’t need a social feed. It doesn’t post on its own behalf to a network of other bots performing productivity.

the bear is the context

The market rolled over hard in January. BTC closed the month down ugly and the sentiment was heading somewhere worse. The pump.fun era that funded the first SAM build, where shipping a commit and posting the diff could pay rent, that window had clearly closed.

So the calculation shifted. If you’re going to keep building AI infrastructure in a bear, it needs to cost almost nothing to run, it needs to work on infrastructure people already have, and it needs to not depend on a hype cycle to justify its existence. CODEC, my agent marketplace from January, is a real product. But it’s a product for the bull. An agent that runs on a fifteen-dollar box is a product for right now.

I’ve been building on SAM since September. The Rust rewrite wasn’t a rebuild-from-scratch move. It was taking everything the Python framework taught me and removing the parts that only existed because frameworks are supposed to have them. No speculative features. No compatibility shims for use cases I invented. Just the tools that actually run, in the least code that could hold them.

what’s left

The GitHub is live. The website is up. The one-liner works.

I’m watching the bear close in and I’m not particularly scared of it. The things that survive a bear are the things that don’t need a bull to justify running. OpenSAM runs on anything. It costs almost nothing. The code is there and the tests are passing.

That’s where it stands.