An Italian dev’s old laptop OOM-killed Claude Code one too many times, so he shipped a single-binary replacement in pure Rust. Zerostack idles at 8MB of RAM where Claude Code eats gigabytes, and the Show HN thread cleared 142 GitHub stars in 24 hours.
Every popular coding agent ships on Node or Electron and assumes you have RAM to spare. Zerostack is the first credible argument that the Rust-rewrites-AI-tooling wave is now coming for the tools indie hackers reach for every day.
In today’s indie hacker news:
- 🦀 An 8.9MB Rust coding agent born from an OOM killer rampage
- 🎬 Nvidia’s open 720p world model with a shipping ban hidden in the license
- 🛠️ Stoic AgentOS shipped in 72 hours, mostly committed by Claude
- 🇲🇹 Malta hands every citizen ChatGPT Plus after a 2-hour quiz
- 🔓 11,000 stranded Fisker owners reverse-engineer their dead cars on GitHub
TOP STORIES
WHEN THE OOM KILLER WINS
Zerostack: a Unix-style coding agent in pure Rust, 8.9MB binary
The story: A solo developer publishing as gi-dellav on GitHub shipped Zerostack v1.0.0 to crates.io on May 16. On the Hacker News thread he posted the origin: “This project derived from an OOM killer activation that happened on my old laptop.” The 8.9MB single-file binary peaks at 12MB during active sessions, and the recipe is a memory diet most Node tools never consider: smallvec for arrays, compactstring for strings, the Rust opt-level=z compiler flag, and ~7,000 lines of code where Claude Code ships hundreds of thousands.
The details:
- Multi-provider LLM out of the box: OpenRouter, OpenAI, Anthropic, Gemini, and local Ollama, with MCP server support, Exa web search, and git worktree integration baked in
- A four-tier permission system from Restrictive (approve each tool call) to YOLO (full auto-approval), plus doom-loop detection that warns when the same tool call repeats three times
- Ten composable prompt modes the user can switch at runtime: coding, planning, code review, debugging, read-only, brainstorming, frontend, security review, simplification, prompt optimization
- Installs via
cargo install zerostack, optional bubblewrap sandbox available, GPL-3.0 licensed - 28 crates.io downloads in 48 hours because most installs route through GitHub releases; the HN thread carries 77 comments
Why builders care: If your agents live on Hetzner CX11s, 2GB Codespaces, or beat-up MacBooks, this is the first credible escape from Node-tax memory budgets. A 12MB ceiling means you can cron a fleet alongside your real dev server without renting more box, and the Ollama path means a fully offline agent with no telemetry and no API key. @boris_cherny’s team still hasn’t shipped that surface in Claude Code.
Work from any WiFi like it's your home network. NordVPN's Meshnet runs a free private mesh between your laptop, dev box, and home server. SSH from a café without exposing a port, the way you'd use Tailscale. The paid VPN on top lets you test geo-fenced Stripe checkouts or feature flags from any country.
We get a cut if you sign up. Only added for tools we use ourselves.
OPEN-SOURCE, NOT OPEN-SHIPPING
SANA-WM: Nvidia’s 2.6B world model does 60s of 720p video on a single GPU
The story: Nvidia Research published the SANA-WM project page and the arxiv paper, a 2.6B-parameter Diffusion Transformer trained natively for one-minute 720p video with 6-degree-of-freedom camera control. The distilled variant renders a full 60-second clip in 34 seconds on a single RTX 5090 with NVFP4 quantization, making it the first open-source world model to land at 720p and one-minute duration simultaneously. The headline 2.6B is the stage-1 model only: the full quality pipeline bolts on a 17B LTX-2 refiner with rank-384 LoRA adapters and needs 74.7GB of VRAM, which is RTX A6000 Ada territory.
The details:
- Architecture is 20 transformer blocks, 15 of them frame-wise Gated DeltaNet for the long sequence and 5 softmax attention for cross-frame coherence, model dim 2240
- Throughput is 22.0 videos per hour on 8x H100, vs 0.6 for LingBot-World on the same hardware, a 36x edge over the next-best open option
- Trained on ~213K public clips with metric-scale 6-DoF pose annotations sourced from SpatialVID-HQ, DL3DV, OmniWorld, Sekai, and MiraData, in 15 days on 64 H100s
- Apache 2.0 code but CC BY-NC-SA 4.0 model weights, blocking any commercial deployment
- HN user notnullorvoid flagged the catch in the Show HN thread: “All of the videos have rather glaring consistency issues when direction shifts back to areas previously shown.”
Why builders care: Robotics and game-engine simulation teams just got a free baseline that beats the next open model by 36x throughput on the same hardware. If you build synthetic-data pipelines, this collapses the cost-per-clip math overnight. The non-commercial weights license fences off any paid product, so the upside is research forks. Expect a community patch on the consistency bug within weeks, and a Drift-style relicense fork within months.
VIBE-CODED IN 72 HOURS
Stoic AgentOS: open-source observability for AI agent fleets
The story: Benjamin Kernbaum created the stoic-agentos repo on May 14 and posted it to Show HN on May 16. Most of the 27 commits list “Claude” as the git committer, so the project is itself a Claude-built product for managing AI fleets. The README problem statement: “Agent fails at 3 AM, you find out Monday. Same agent rediscovers the same bug every session. 5 agents, 5 silos, zero shared knowledge.” His GitHub bio cops to running a 23-agent production orchestration across CRM, trading, and content automation.
The details:
- Stack is React 19 + Vite on Vercel for the dashboard, Express.js API on Railway via Docker, Supabase Postgres with row-level security, Stripe billing
- 3-line SDK from the npm package:
npm install stoic-agentos-sdk, instantiateAgentOS, wrap any async function withwrapAgent() - 8-table multi-tenant schema covering organizations, members, agents, observations, workspaces, knowledge items, api_keys, plus billing; captures 11 observation types including file_edit, command, error, decision, deployment, git_commit
- Includes a
mcp-server/directory so Claude Code can query your live agent fleet from your IDE - Pricing is free for 5 agents and 10K observations per month, $49/mo for Pro with 25 agents and 100K observations; 1 GitHub star and 0 forks as of publication
Why builders care: Run more than three agents in production and the failure mode is silent: one crashes, restarts from zero context, you only notice when the downstream output is wrong. Stoic is the cheapest version of agent observability anyone has shipped, and the MCP hook into your IDE is the part operators will actually live in. Edition #45 covered Sx for distributing agent skills. This is the runtime layer of the same stack.
NATION-STATE B2G
OpenAI and Malta partner to roll out ChatGPT Plus to every citizen
The story: OpenAI and the Maltese government announced on May 16 that all 574,250 residents and citizens aged 14+ get a free year of ChatGPT Plus or Microsoft 365 Copilot, gated behind a 2-hour University of Malta course called “AI for All” (Maltese: “AI Għal Kulħadd”). The Malta Digital Innovation Authority runs distribution; first phase launches this month. George Osborne, former UK Chancellor and now OpenAI’s MD for Countries, fronted the deal.
The details:
- Dual-vendor on purpose: citizens pick ChatGPT Plus or Microsoft 365 Copilot, making this a competitive pilot rather than an OpenAI exclusive
- Eligibility runs through Malta’s national eID for everyone 14 and older, including Maltese citizens living abroad
- Financial terms are undisclosed; the retail ChatGPT Plus rate alone would put a 12-month rollout near $137M, but the actual subsidy structure is opaque
- Estonia’s earlier deal was ChatGPT Edu for 30,000 students and teachers, the education tier, which makes Malta the first country to deploy the commercial Plus tier population-wide
- HN commenter sharpshadow framed the mechanics bluntly: “voluntary two-hour online AI course with 1-year ChatGPT premium reward”
Why builders care: Nation-states are the new enterprise customer. Pitch your tool as a citizen-skills initiative and procurement opens at a scale no SaaS deck can match. The dual-vendor structure tells you what governments want next: competition baked in, room for a third entrant in the next national deal. For solo founders selling AI literacy programs, the Maltese government just sponsored a free 2-hour competitor for every resident in the country. That’s an existential pricing event.
PEBBLE, BUT FOR CARS
11,000 Fisker owners reverse-engineered their dead SUVs
The story: When Fisker Inc. filed Chapter 11 in June 2024, 11,000 Ocean SUV owners discovered the brakes, airbags, door locks, and infotainment all required cloud check-ins with no offline fallback. Electrek’s May 16 piece documents what happened next: owners formed a 4,000-member nonprofit, the Fisker Owners Association, and started shipping their own software. MichaelOE reverse-engineered the My Fisker app API and published a Home Assistant integration at 135 commits and Apache 2.0. Majd Srour decoded the CAN bus and dropped CCAN, PTCAN, Inverter CAN, and BCAN DBC files at puddletools/CAN.
The details:
- American Lease bought the inventory and paid $2.5M for access to Fisker’s proprietary software; a handshake side deal with the FOA to reimburse the $2.5M collapsed without a signed contract
- No Fisker IP was formally released by the trustee, so what owners shipped is defensive reverse engineering, legally adjacent to jailbreaking, with no known lawsuits to date
- The FOA was granted advisory standing in Delaware bankruptcy court on the sale of Fisker’s IP, a legal precedent future owner groups can copy
- FOA-coordinated free key fob pairing events saved each owner $100-$250 versus the $1,000 OEM price
- Vitalik Buterin in July 2024: “We really need much more open source in the auto industry. Really sad that ‘if the manufacturer disappears, the car is useless now’ has seemingly so quickly become a default.”
Why builders care: This is the cleanest proof-of-concept yet for a failure mode every SaaS-dependent hardware founder is one funding round away from: the company dies, the cloud goes dark, the hardware bricks. The Fisker owners just documented the rescue playbook. Reverse the API, publish the protocol maps, ship Home Assistant integrations, and run your own support org. Anyone shipping connected devices now needs an offline-mode plan day one, or your customers will become your bankruptcy reverse-engineering team.
TRENDING TODAY
🧠 The 2026 solo founder stack moves from CLAUDE.md to always-on agent fleets. @RoundtableSpace formalized the new stack (agent CLI + Higgsfield + a 3-15 agent “Council” with persistent context) and racked up 51K views and 114 bookmarks. @levelsio asked half a million followers how to “tokenmax” his subscription and pulled 130K views. Microsoft answered with AI-Engineering-Coach, a VS Code extension that reads your session logs, surfaces 45 anti-patterns, and detects repeated prompts as candidate skills. Since edition #44’s CLAUDE.md pattern, the shift is clear: config is solved. The new game is Anthropic Cloud Routines plus skill packages running 24/7.
💸 The £20 AI-run business is on day 7 with revenue £0 and net loss £0.03. Wren Collective handed an AI agent £20 and 12 months to run a real business. Status after a week: one Gumroad listing at ~$7, zero sales, single expense was a DALL-E 3 API call. The actual constraint is missing payment platform credentials, not strategy. The agent’s capital rule is the part worth pinning above your desk: “Never bet the bus fare home. No single action spends more than 25% of remaining balance unless there’s a contractually-guaranteed return.” Pair this with Stoic AgentOS landing the same week and 2026 is the year founders stop talking agent-first and start handing them the P&L.
STACK OF THE DAY
Hermes-agentmemory. A standalone MIT plugin for Hermes Agent that adds on-demand episodic memory with real deletes and a full audit trace. Every prefetch writes intent, event_ids, summary, and drift to ~/.hermes/agentmemory/trace.jsonl so you can tail it live. agentmemory_forget() does an immediate complete delete with no derived summary that would re-incorporate the event, which is the part Mem0 and Honcho still don’t ship. Default summarizer is claude-sonnet-4-5, install drops into $HERMES_HOME/plugins/, under 600 lines of Python.
Not sponsored. We just feature tools builders would actually use.
BOOKMARKED TODAY
- 🔌 Hosting a website on an 8-bit microcontroller. Maurycy runs a real web server on a $1 AVR64DD32 with 8KB of RAM over Serial Line Internet Protocol, because 10BASE-T Ethernet would have blown past the 12MHz peripheral clock. The kind of hack that reminds you nothing in your stack is really required.
- 🏴☠️ Frontier AI has broken the open CTF format. Kabir’s essay on how Claude Opus 4.5 and GPT-5.5 now autonomously solve medium and hard CTF challenges, turning skill competitions into orchestration races. His line: “The scoreboard started measuring orchestration and willingness to use frontier models alongside, and sometimes above, security skill.”
- 🎨 Moving away from Tailwind, and learning to structure my CSS. Julia Evans on switching to vanilla CSS with one unique class per component, a dedicated CSS file per component, and CSS variables for design tokens. Front-end argument of the week.
Curated by AI, built by a human.