So @Python_Dv dropped a tweet listing the full /usr/bin/bash AI stack for 2026 and it blew up — 68K views, 1.4K likes. If you haven't seen it, the gist is: you can build a production AI system without spending a dollar. Ollama for LLMs, LangGraph for orchestration, LlamaIndex + Chroma for RAG, MCP for tool use, Claude Code for coding, Next.js on Vercel free tier, SQLite for data, Langfuse for observability, Docker for deploy.
Honestly? Most of this is solid. I've used nearly every tool on that list at some point. But after spending the last few months building agent infrastructure — specifically an SDK that handles memory, identity, and payments for AI agents — I think the tweet is missing a couple of things that matter once you move past the demo stage.
Let me walk through what I agree with, where I'd push back, and what's genuinely absent.
The stuff that actually works at /usr/bin/bash
Ollama is a game-changer. Running Gemma 4 or Llama 3.3 locally was painful even a year ago, and now it's basically and you're off. For prototyping, local dev, anything where you don't need to worry about rate limits — it's unbeatable.
LangGraph and CrewAI are fine for orchestration, though I'd say LangGraph has gotten complex enough that you should really know what you're doing before reaching for it. CrewAI is simpler but the abstractions can get in the way. Neither is bad. Pick one and commit.
MCP deserves its own callout. The Model Context Protocol is genuinely the most important thing on this list because it solves a real interoperability problem — agents need to talk to tools, and MCP gives them a standard way to do it. I've built MCP servers for my own projects and the developer experience is legitimately good.
SQLite and DuckDB — yes, absolutely. If your first instinct is to spin up Postgres for an AI side project, you're overengineering it. SQLite handles more than people think.
Where it gets shaky
Here's where I start squinting.
"Deploy: Docker / Cloudflare Workers / HuggingFace Spaces" — sure, technically free, but Cloudflare Workers have CPU limits that'll bite you the second you try to do anything compute-heavy. HuggingFace Spaces are great for demos, terrible for anything you'd call production. Docker is free but you still need somewhere to run it. That somewhere usually costs money.
Vercel free tier is generous but it has bandwidth limits, serverless function timeouts, and if your AI app gets any real traffic, you're upgrading fast. Same with Supabase — their free tier pauses your database after 7 days of inactivity, which is a fun surprise when your agent tries to query a sleeping DB at 3am.
Langfuse self-hosted is cool in theory. In practice, you're now maintaining an observability stack on top of your actual application. For a side project or learning — fine. For production? You're trading dollars for hours, and your hours aren't free.
None of this means the tweet is wrong. It's just that "/usr/bin/bash" has a footnote, and the footnote says "assuming your time is worthless and traffic stays low."
The thing nobody mentioned
Here's what really jumped out to me though — there's no layer for agent memory or payments. And I don't mean "store embeddings in Chroma." I mean: when your agent takes actions across sessions, how does it remember what it's done? When it needs to spend money on behalf of a user — call an API, purchase something, pay for a service — how does that work?
This is the gap I've been living in. I built MnemoPay specifically because I kept running into this wall. You'd have an agent that could reason, use tools, generate great responses... and then it had no concept of trust, no spending limits, no memory of past transactions, no identity. It's like building a self-driving car with no odometer and no wallet.
MnemoPay handles the memory + payments + identity layer for AI agents. It's got 14 modules — Agent FICO scoring (think credit scores but for agents), behavioral anomaly detection, Merkle integrity for tamper-proof transaction logs, support for Stripe and Paystack payment rails. It runs as an MCP server so it plugs right into the stack from that tweet. And yeah, it's open source — .
I'm not saying every project needs this. If you're building a chatbot that answers questions about your docs, you don't need agent payments. But the tweet said "production AI system," and production systems eventually need to handle money. The agentic commerce wave is already here — agents booking flights, agents managing subscriptions, agents negotiating prices with other agents. That stack has no answer for any of it.
What I'd actually recommend
If I were starting from zero today with the goal of building something real, I'd take about 80% of that tweet's stack and add two things:
- A persistence layer that's agent-aware — not just a database, but something that understands sessions, trust, and state across interactions
- A payments/identity layer that gives agents spending authority with guardrails
The rest of the tweet holds up. Ollama + a good open model for the brain, MCP for tool connectivity, LlamaIndex for RAG if you need it, Next.js or Streamlit for the face. That's a legit starting point.
Just know that the jump from "it works on my laptop" to "it works in production with real users and real money" is where most of these /usr/bin/bash stacks quietly fall apart. The tools are free. The complexity isn't.
go build something
The best part about 2026 is that the barrier to entry genuinely is near zero. Two years ago you needed API keys and a budget to even start. Now you can run a capable LLM on a laptop with 16GB of RAM. That's incredible and the tweet is right to celebrate it.
Just don't confuse "I can start for free" with "I can ship for free." They're different problems, and the second one is where it gets interesting.
If you want to poke around the payments + memory layer I mentioned: github.com/mnemopay/mnemopay-sdk. PRs welcome, issues welcome, "this is dumb and here's why" also welcome. That's how open source works.
Top comments (0)