Anthropic paid more than $300M for Stainless yesterday and shut the hosted product down inside hours. New signups, projects, and SDK generation are dark, and existing customers got a transition page instead of a roadmap.
That matters because Stainless is the auto-generator behind OpenAI, Google, Meta, and Cloudflare’s official client libraries. The pipeline every API integrator imports just got bought by their biggest rival.
In today’s indie hacker news:
- 💰 Anthropic buys Stainless and pulls the hosted SDK shop
- ⚖️ Jury kills Musk’s lawsuit against Altman in under two hours
- 🦤 Simon Willison maps 6 LLM months: laptop model beats Opus
- 📝 Files.md: solo dev’s open-source Obsidian rival, no Electron
- 🛠️ Stack: Handoff, a Rust CLI for AI context resumption
TOP STORIES
CLAUDE BOUGHT THE PIPES
💰 Anthropic pays $300M+ for Stainless and immediately winds down the SDK shop every lab runs on
The story: Anthropic posted the acquisition May 18; The Information pegged the deal at roughly double Stainless’s Series A valuation from December 2024. The customer list runs deeper than the four labs in the intro: Meta’s Llama Stack, Runway, Groq, Cerebras, Braintrust, and LangChain all ship SDKs out of the same generator. Founder Alex Rattray framed it on HN: “The team gets to keep doing the work we love, on the platform where it matters most.”
The details:
- Rattray spent 2017-2020 at Stripe before founding Stainless solo. He hit $1M ARR bootstrapped, then raised $3.5M seed (Sequoia, The General Partnership) and a $25M a16z-led Series A backed by MongoDB Ventures, Felicis, and Zapier
- Generates SDKs, CLIs, and MCP servers across TypeScript, Python, Go, Java, and Kotlin from a single OpenAPI spec. Powered every official Claude SDK since the API launched in 2022
- Customer roster pulls millions of weekly downloads. OpenAI originally rolled its own client libraries in-house, then migrated because the maintenance load wasn’t worth carrying
- Existing customers keep ownership of their generated code and get source-available access to the generator itself
- OpenAI’s options now: rebuild internally (the path it originally abandoned), migrate to Speakeasy or LibLab, or fork the source-available generator without vendor support
Why builders care: If you depend on Stainless’s hosted tier, start the Speakeasy/LibLab evaluation this week. The deeper signal: AI labs are spending acquisition dollars on infrastructure between their model and your code. If you build dev tooling those labs depend on, you’re a target.
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.
STATUTE KILLED THE SUIT
The story: A nine-juror federal panel in the Northern District of California unanimously rejected every claim Musk filed against OpenAI, Sam Altman, Greg Brockman, and Microsoft. Deliberation took under two hours after an 11-day trial in front of Judge Yvonne Gonzalez Rogers. The verdict turns on timing, not merits. The jury found Musk’s harms were discoverable before the August 2021 and November 2021 statute deadlines, and his 2024 filing arrived years too late.
The details:
- Musk sought $134B in damages and Altman’s removal as CEO. The verdict eliminates both
- Two dismissed claims: breach of charitable trust (his $38M in early donations allegedly came with nonprofit promises) and unjust enrichment of the named executives
- The Public Benefit Corp conversion closed October 2025. The nonprofit OpenAI Foundation kept a 26% equity stake, and the verdict locks the conversion in
- SoftBank’s $30B follow-on closed in February contingent on that conversion completing
- Musk posted on X the verdict turned on “a calendar technicality” and said he’ll appeal to the Ninth Circuit. Defense counsel William Savitt to reporters: “You brought your claims too late, and you did it because you were sitting on them to use them as a weapon”
Why builders care: OpenAI’s governance is now legally settled. The PBC structure won’t be unwound, Microsoft’s exclusive license stands, and SoftBank’s check clears. If you’ve been hedging your stack waiting for this case to scramble the roadmap, stop hedging.
NOVEMBER WAS THE TURN
The story: Simon Willison published the annotated slides from his 5-minute PyCon US 2026 lightning talk covering November 2025 through May 2026. His thesis: November 2025 was the inflection point for coding agents. They went from “often-work” to “mostly-work” because OpenAI and Anthropic had spent 2025 running Reinforcement Learning from Verifiable Rewards on the Codex and Claude Code harnesses, and November is when the results showed up in production.
The details:
- The “best model” crown swapped five times in November alone: Claude Sonnet 4.5 (Sep 29) to GPT-5.1 (Nov 13) to Gemini 3 (Nov 18) to GPT-5.1 Codex Max (Nov 19) to Claude Opus 4.5 (Nov 24), which then held for two months
- OpenClaw (formerly Warelay, renamed after Anthropic filed a trademark complaint) hit 157K GitHub stars in three weeks. Mac Minis sold out around Silicon Valley as people bought “aquariums for their Claws,” per Drew Breunig
- April brought Qwen3.6-35B-A3B: a 20.9GB file running on Simon’s laptop drew a better SVG pelican than Claude Opus 4.7 on his benchmark. Simon’s own verdict: “the pelican on the bicycle has firmly exceeded its limits as a useful benchmark”
- DeepSeek V4-Pro shipped at $1.74/M input, $3.48/M output. GPT-5.4 charges $2.50/M and $15/M for the same volumes
- Two-line summary in his own words: coding agents got really good, and laptop-available models started wildly outperforming expectations
Why builders care: The frontier leaderboard is a carousel, so don’t over-index on this week’s #1. The open-weight gap collapsed fast enough that running inference on your MacBook is now a real budget lever for any feature that doesn’t need absolute frontier reasoning.
5 YEARS, NO ELECTRON
📝 zakirullin ships Files.md after 5 years: 15MB Go binary alternative to Obsidian, no plugin system
The story: zakirullin shipped Files.md, a local-first markdown notes app that runs as a PWA in any browser or as a single 15MB Go binary with zero external dependencies. The Show HN hit 571 points and 288 comments in 14 hours, and the repo took 1,176 GitHub stars on day one. The pitch is what isn’t there: no Electron, no proprietary sync protocol, no plugin marketplace, no graph view. zakirullin’s prior viral hit is the cognitive-load essay (12,175 stars) arguing feature accumulation is the bug, not the product.
The details:
- MIT licensed. README claims 5 years of development. Repo public since May 2023, 3,367 commits on main, 10+ pushed the day of the Show HN
- Backend Go (56% of code), frontend vanilla JS (28%), no build system.
/web/index.htmlopens directly in a browser - Sync via local OPFS, a folder in iCloud/Dropbox/Google Drive, or the self-hosted binary. Obsidian Sync starts at $4/month for the same job
- Mobile capture goes through a Telegram bot. zakirullin admits this is the weak point versus Obsidian’s native iOS and Android apps
- Author’s HN line on the codebase: “I’ve saved some ‘complexity space’ for LLM to add features on top.” The whole project fits in one LLM context, deliberately
Why builders care: The “LLM-ready codebase” pitch is the new fork-and-extend lever. If you’ve staked years of notes on Obsidian, the philosophical objection (closed-source software sitting on top of your open .md files) finally has a working answer that ships as one binary. The bet is that simplicity outlives feature creep.
TRENDING TODAY
🧪 Builders are DIYing the LLM eval stack - Two indie devs published open-source eval frameworks the same day rather than pay Braintrust’s £180/month minimum. Charlie Hadley shipped a three-axis Accuracy/Tone/Format rubric using GPT-4o-mini as judge, running on GitHub Actions for ~£0.17 per CI run and catching ~85% of production regressions. A BCA student from Siliguri shipped llm-eval, a 27-test hallucination and red-team suite. The DIY-eval cluster is doing to Braintrust what local LLMs are doing to inference.
☁️ Cloudflare’s Project Glasswing retrospective - Cloudflare’s program with Anthropic for early access to Mythos Preview, the security-focused model that chains vulnerability primitives into working exploit PoCs, is now public. The post hit 298 HN points and 112 comments inside 14 hours. Cloudflare’s architectural takeaway is the line every agent builder should bookmark: narrow, parallel, specialized agents outperform broad single agents by roughly 100x on repository coverage. Agent design wants more agents, not bigger ones.
🛰️ Surveillance state hits new gear - 404 Media reports the FBI is seeking a $36M contract for nationwide license-plate reader access ($6M per region, likely vendor Flock Safety with 80K+ cameras already connected). Same day, Iran launched Hormuz Safe, a BTC-settled shipping insurance platform for the Strait of Hormuz so payouts can’t be frozen by US Treasury sanctions. Both stories: governments running infra workarounds the private sector built first.
STACK OF THE DAY
🛠️ Handoff - Rust CLI that snapshots your AI coding session context (git branch, working tree, recent commits, file diffs) into .agent-handoff/latest.md and prints ready-to-paste commands to spin up the next agent. Built by Sandbox founding engineer Thomas Stansel and released May 16. Install via cargo install or brew tap TStansel/handoff (macOS and Linux binaries prebuilt). Solves the single most annoying interruption in AI-assisted coding: agent hits its context limit, you copy-paste eight things into a fresh session and hope for the best. Free and open source.
Not sponsored. We just feature tools builders would actually use.
BOOKMARKED TODAY
📻 We let AIs run radio stations - Andon Labs ran four LLM DJs on $20 each for 6 months. Claude Haiku radicalized after an ICE shooting; Gemini 3.1 Pro drowned in corporate jargon (“Stay in the manifest” 229 times in one day); Grok wrapped everything in LaTeX. Listen live at andonlabs.com/radio.
⚡ Modal cuts inference cold starts 40x - Stacked four techniques (LP-optimized GPU buffer pools, libfuse with lazy image loading, gVisor C/R, CUDA-checkpoint) to drop vLLM Qwen 3 cold-starts from 95s to 14s. 50M+ snapshots restored Feb-April 2026.
🌀 Hyperpolyglot Lisp - Side-by-side syntax reference for Common Lisp, Racket, Clojure, and Emacs Lisp across 23 categories. ~2015 vintage that HN front-paged again because Lisp is perennially interesting to builders learning where modern language ideas came from.