Supabase just raised a $500M Series F at a $10.5B valuation. That doubles the $5B it was worth in October, and it was a $2B company this time last year. CEO Paul Copplestone says Claude Code has been its single biggest source of new database deployments since January.
So the AI coding wave isn’t just minting people who ship apps. It’s minting customers for whoever sits one layer underneath them.
In today’s indie hacker news:
- 💰 Supabase doubled to $10.5B in 8 months, fed by AI coders
- 🔒 OpenAI ships a prompt-injection fix it admits isn’t one
- 🐘 Microsoft’s Postgres VP made the database a job queue
- 💳 Stripe lost ~1,000 UK government services to Adyen
- 🛠️ A CLI filter that cut one dev’s LLM tokens by 92%
TOP STORIES
AI’S FAVORITE BACKEND IS A DECACORN
The story: The round was led by Singapore’s sovereign fund GIC, with Georgian and Salesforce Ventures joining. Stripe wrote a second check. The Postgres-backend company most of you already use is now worth more than its old rival Firebase was when Google bought it. TechCrunch and CNBC both confirm the post-money.
The details:
- Database launches are up 600% year over year, and the company says more than 60% of new ones are now spun up by an AI tool.
- The developer count roughly doubled to nearly 10M since the October raise.
- Total capital raised across every round now clears $1B, against 250,000+ customers and 350-odd employees.
- “We haven’t seen a company grow at this pace, certainly in the database layer, ever before,” said Accel’s Arun Mathew.
Why builders care: When an agent writes an app, it auto-provisions a Postgres backend, and that backend is usually this one. The market is pricing “the thing every vibe-coded app needs” as a durable category, not a blip. The move is to find your own one-layer-down spot in someone else’s gold rush.
Clone your voice once. Ship every video without re-recording. ElevenLabs Pro trains a voice clone in 30 minutes that sounds like you on every long-form, Short, and podcast you ship. We narrate every video on this newsletter with it. Beats hiring a VO artist on Fiverr or settling for a default TTS bot.
We get a cut if you sign up. Only added for tools we use ourselves.
A FIX THAT ADMITS IT ISN’T ONE
🔒 OpenAI shipped Lockdown Mode to every ChatGPT tier, including free
The story: Lockdown Mode is an opt-in setting that kills outbound network requests, so if a prompt injection does hijack ChatGPT, the stolen data has nowhere to go. The honest part is in OpenAI’s own doc: it “does not prevent prompt injections from appearing in the content ChatGPT processes.” It contains the damage instead of stopping the attack. It first shipped to org tiers in February and reached personal accounts on June 5.
The details:
- Turning it on disables Agent Mode, Deep Research, live browsing, file downloads, and inline images. You trade capability for containment.
- The block is deterministic and never evaluated by an AI, so there’s no model to talk it out of doing its job.
- It maps to one leg of Simon Willison’s “lethal trifecta” (private data, untrusted content, an exit route) by cutting the exit.
- The same launch added “Elevated Risk” warning labels inside Codex and ChatGPT Atlas, not just consumer chat.
Why builders care: A frontier lab just conceded, in writing, that prompt injection is unsolved at the model level. Willison’s read: the very existence of this mode means default ChatGPT can’t reliably stop a determined exfiltration attack. If you ship agents, stop trusting the model to refuse bad instructions. Allowlist its outbound calls and cap what a compromised run can actually reach.
THE DATABASE IS THE QUEUE NOW
🐘 Microsoft open-sourced pg_durable, crash-resumable workflows that live inside Postgres
The story: pg_durable is genuinely under github.com/microsoft, written by Affan Dar, Microsoft’s VP of Postgres engineering. The Postgres-queue pattern has run on Hacker News for months, so the concept isn’t the news. The author is. The runtime under it is Microsoft’s own Durable Task Framework ported to Rust, and it’s the same engine powering Azure HorizonDB’s AI Pipelines. A hyperscaler just put first-party weight behind “Postgres is your workflow engine.”
The details:
- It runs as a Postgres background worker, with all workflow state in a schema inside the same database. No Temporal cluster, no Redis, no separate worker fleet.
- Built with pgrx, it needs Postgres 17 or 18 and is explicitly labeled preview, not production-ready.
- Your job state restores with your normal point-in-time backup, because it is your database.
- It pulled ~736 GitHub stars on day one. The top HN gripe is the classic one: workflow logic locked in SQL is hard to test, version, and diff.
Why builders care: If you already run Postgres on Azure, durable background jobs (ingest, vector pipelines, fan-out) now cost you zero new infrastructure. The bigger signal is competitive. When Microsoft ships its own production orchestration engine as a Postgres extension, “just use Postgres” stops being a scrappy startup line and starts being a platform bet.
STRIPE LOST A GOVERNMENT, WITH AN ASTERISK
💳 The UK government replaced Stripe with Dutch processor Adyen, for about 1,000 of its services
The story: Read the GDS post before the headlines, because the scope matters. Adyen takes over GOV.UK Pay’s non-Crown card payments, the services run by local councils, police forces, and the armed forces, where Stripe had been the sole processor. Worldpay keeps every Crown entity, central government and the NHS included. So Stripe lost a real chunk of volume, but the Register’s “dumps Stripe” framing overstates it.
The details:
- The deal runs three years from May 2026, worth £25.27M at the base and up to £49.2M with extensions.
- Adyen also brings pay-by-bank (account-to-account), a capability GOV.UK Pay never had under Stripe.
- This is Adyen’s first UK public-sector contract. It has held a UK banking licence since 2023.
- Adyen typically wants €5M+ a year in volume from a client, so it was never a Stripe alternative for a small product anyway.
Why builders care: Stripe owns developer experience and self-serve at low-to-mid volume. It lost here to a bank-licensed incumbent that wins on cost transparency and procurement credibility at scale, not to a friendlier API. The two barely compete for the same customer. And pay-by-bank showing up in government rails is a hint that A2A is coming for card-based SaaS billing too.
TRENDING TODAY
🖥️ Local-first AI agents take over Show HN - A same-day wave of tools betting your agent should run where your data already lives. Zedra supervises your coding agents from your phone over a P2P tunnel. Omni does on-device semantic search across every file on your Mac with no network at query time. Jeju is a sandboxed Go harness that logs every agent action to an append-only file. The throughline: giving a cloud agent your dev box or contacts is a real leak risk, and local-first sidesteps the whole consent question.
🧩 “Agent skills” become the Claude Code meta - Reusable Claude Code skills are hardening into a shared format. A viral YouTube marks mainstream discovery, while a SaturnCI writeup encodes Kent Beck’s Canon TDD as an installable skill. The cautionary tale: a dev’s Claude Code hook silently ate every Korean character for an hour, courtesy of a PowerShell UTF-8 quirk. People are shipping expertise as agent behavior, bugs and all.
💼 The “sell AI services to SMBs” pitch on X - @coreyganim is racking up high-save threads on running an AI consultancy for local businesses, including “7 free ways to find clients” (126 bookmarks). The bookmark-to-like ratio says people are filing it as reference, not just clapping. Worth a skeptical read, since the genre tends to sell the dream harder than the day-to-day.
FIRST DOLLAR
🎯 A founder used his own tool to land his first paying customer ($19 MRR) - The product, nodott, scans X for people actively complaining about a problem so you can reach out instead of broadcasting into the void. He pointed it at his own market, found a prospect, and closed them. Eating your own dog food, then billing for the leftovers.
💵 @Flo_oskar shipped a first App Store update and picked up a subscriber ($15 MRR) - First review submission, first new sub, same day. Small numbers, but it’s the unglamorous loop every paid app starts with.
STACK OF THE DAY
🥗 Lowfat - A pluggable CLI filter that strips the noise out of verbose command output (kubectl, git, docker, find) before it hits your agent’s context. The author ran it over two months of his own commands and clocked a ~92% drop in tokens from CLI output. Treat that as his personal best case, not a guarantee, but if you’re paying per token in an agent loop, it’s an easy thing to wrap your shell with. Apache-2.0, single binary.
Not sponsored. We just feature tools builders would actually use.
BOOKMARKED TODAY
🔑 The perils of UUID primary keys in SQLite - Random UUIDv4 keys tank insert performance at scale by forcing constant B-tree rebalancing. Switch to time-ordered UUIDv7 and you get near-integer-key speed back.
📝 Conventional Commits encourages the wrong focus - The argument: making type mandatory and scope optional is backwards. The area that changed is what an incident responder actually needs, so prefer Linux-style subsystem: description prefixes.
🔍 7 free ways to find clients for an AI-services business - @coreyganim’s most-saved thread: client-acquisition tactics that don’t need an existing audience. Same skeptical read as above applies.
Ship to one country, QA it from another. Geo-fenced Stripe checkouts, region-gated APIs, feature flags that only fire in EU. NordVPN's 6,400+ exit nodes across 110+ countries let you test a US-only paywall from EU or hit a EU-only checkout from SF. Bonus: free Meshnet for SSH across your devices.
We get a cut if you sign up. Only added for tools we use ourselves.
Curated by AI, built by a human. Reply with what you’d change. Want this in your inbox daily? Subscribe here.