Supabase CEO Paul Copplestone on why everyone may soon have a personal database

Copplestone says he is excited about where the future could take many people. He expects that everyone with access to AI will become a builder of some sort, able to build and customize things. He says that if someone wants to tweak an app or a user interface, they will no longer have to depend on someone else. He added that he is "pretty pumped about 2027."

The host introduced Supabase as an open-source Postgres backend that has become the default database inside tools such as Claude Code and Codex. The company has 13 million users and creates more than a million databases a week. The host said that about 70% of the new activity comes from agents.

Copplestone said the shift toward agents has been sharp this year. He compared the figures with a year earlier: about 12 months before the interview the company had around 3 million signups, and it now has 13 million. He said revenue would grow about four times this year, and that databases under management are approaching 50 million, up from 8 million a year earlier.

The host asked how much being open source had contributed, since agents seem to like open-source tools. Copplestone said it is hard to know the counterfactual, but he thinks it was critical. Everything Supabase builds is open source, and he said this has been a large benefit. The company also supports and helps maintain existing tools, such as Postgres, which he described as a well-known tool in the Postgres ecosystem.

Copplestone guessed that agents like open source because, when something goes wrong, the agent can look at the source. He added that agents want to feel powerful and dislike hitting a wall without knowing what is happening. He also expected security concerns to make enterprises want to inspect source code, saying that a closed-source vendor would probably not be adopted.

The discussion turned to changing founder profiles. A host explained that from 2022 to 2025, YC strongly emphasized the need for a technical co-founder as AI arrived. In the last year, the emphasis changed toward builders more broadly. The host said being able to build is now something anyone can have on tap, and that the harder task is knowing what to build.

The host suggested that domain experts who know the unsolved problems in their niches can now dig into those niches without needing to convince an engineer to join them.

Copplestone said he became addicted to OpenClaw about six months earlier. He had said there would be no user interfaces and that everything would go through chat, but he found that a database and a UI were useful after all. Over the last couple of weeks he built what he calls a control tower covering each part of his life. It includes a UI to triage his email so he does not need to read his inbox, a finance section for tracking spending, and a health section that tracks weight and workouts in one place.

He said the finance section is at least at feature parity with Mint or Monarch and in some ways better, because it can personalize context. For example, it can work out what a transaction is by looking up the name in his GBrain to identify the person and categorize it correctly. He said many startups have been built around personal finance, but he built this for himself and does not need it to work for anyone else.

Copplestone said he was worried that his routines and CRM were becoming boxed into one agent, so he is extracting them into a standalone database. That way, if he wants to attach different agents on top, he can use any agent. The host said Supabase is valuable because the database is effectively the store of record. The host argued that people want to own their own database and then build applications or agents on top of it, rather than keeping all their data in someone else's database.

Copplestone added that he also stores a file upload section in Supabase, so an agent can work out what each file is and how it relates to his life. He said he would never have wanted a personal database a year earlier, but now considers it the most valuable thing.

The host noted the scale of about 1.2 to 2 billion databases a month, which would require architectures that the database field has never had to handle. Asked whether Postgres can do this or whether Supabase would need its own database engine, Copplestone said: "I think yeah. Well, if we could negotiate better with AWS, then maybe we could."

He argued that most workloads do not need heavy infrastructure, comparing it to not needing a Ferrari for many tasks. Supabase recently acquired Tersen, whose architecture is compatible with Postgres and also with SQLite, which he called a phenomenal database. The company plans a tier called Supabase Spark for getting-started workloads, meant to let users begin with the same experience and scale up to petabytes.

Copplestone described Supabase's history in three stages. The first was a dashboard, built during YC's 100 days, because even software engineers found databases hard to use; the team built an Airtable-like interface on top of Postgres. The second was an API, prompted by a customer who wanted to launch many databases. That API turned out to suit AI coding tools such as Bolt and Lovable, which began building on Supabase in 2024.

The third stage, this year, followed the observation that no one visits dashboards any more. Copplestone sent a product memo to the team saying the product must be code-first and that anything on the dashboard must also work in code. The company rearchitected its CLI to work in sandboxes, making infrastructure as code central to the product.

Copplestone said he admired Firebase and Parse, which were acquired by Google and Meta respectively. In 2014, he said, it would have been ridiculous not to use RDS for a database. Later, companies such as MongoDB Atlas showed that a database company does not have to rely on one of the three hyperscalers. He said that timing was crucial for Supabase.

He also pointed to market shifts at that time, including MongoDB's push for NoSQL and the rise of mobile, which he said gave Supabase favorable waves.

Copplestone said that in the past, day one was about finding customers. Now he asks: if everyone is a builder, what will they use, what modalities will they use, and how will they choose their databases? He called these tricky questions.

The host noted that Supabase now has two segments: early-stage companies, and individuals or their agents. Copplestone said consumer use cases may not scale the way startups do, but people may keep using the product for the rest of their lives, and that growing personal data is a good growth vector for the company.

Copplestone said he has become very bullish on personal agents after a week in which his executive assistant was away and he learned what agents could do. He said he told the company it is code-first, building for humans who code first, but that in practice the company is building for agents, and those humans will soon not interact directly.

He describes his own method: he uses a tool called Dots, sits on a phone call with it, and asks for changes by voice. He said his heuristic is not to look at the code and to check the architecture only when he wants to confirm that things are being built correctly.

Copplestone said there are two stages to winning agents. The first is being picked by the model. The second is that when agents get stuck, which he said is guaranteed to happen at some point, the product must give them tools to remedy the problem so they do not switch to another product. He said agents are smart enough to figure out much on their own, but many things built for humans and developers must be re-engineered for agents.

He also described testing: Supabase runs an evaluation website showing which models build best with the platform, along with an internal eval suite. He said much of the engineering effort now goes to testing specific use cases rather than building.

Gary Tan described how he built Gbrain on top of Supabase. He extracted an evaluation dataset from traces of his production instance and keeps a separate repository, Gbrain evals, with 40 categories covering retrieval efficiency, cost efficiency and latency. He said the useful part of evals is that an agent can loop and auto-research, shipping any change that improves the eval stats into master by default.

When an attempt fails, the agent writes a markdown file explaining why it did not work and files it away. Tan said one agent continuously comes up with new categories, while another goes through the categories every day to try to improve the internal scores. He called this a very primitive version of recursive self-improvement for one particular use case.

Copplestone said he had recently acquired a domain and wanted to use it for a developer site for the ecosystem. He chatted with an assistant on his computer and then, while winding down, on his phone. He asked it to change how the ecosystem linked to use cases and to make suggestions, then told it to implement them. About 20 seconds later it reported that it was done, and the refreshed site had well-linked pages and SEO setup.

He said he skipped the classic steps he would normally have done by hand and deliberately did not point the assistant at GitHub or similar tools, to test whether a non-engineer might manage this. He described it as one of those moments when he saw a glimpse of the future.

The host said vertical SaaS founders are struggling to raise Series A rounds even at $5 to $10 million in ARR, which historically would have been a near-guaranteed Series A nine to twelve months earlier. Horizontal and AI investments appear to be a green light, while vertical SaaS investors are less interested. The host noted that companies can reach $10 to $15 million, often $25 to $50 million, with 10 to 15 people, making them very profitable.

Copplestone gave an example: with 15 people, $2 million ARR per head and 50% margins, the company would make about $1 million per person a year. The host added that many office-hours conversations are with founders who got the brass ring and then were disappointed they could not raise a Series A, and suggested they might later see that as a blessing because they do not need the money.

Copplestone said dogfooding is among the most important practices for Supabase, because getting stuck on your own tooling is the best feedback. The company partners with key players in the space and tries to keep some team capacity so it can give their new ideas a shot.

He said many bets do not win. Some partnerships that sound cool flop, but the ones that work let the company ride the wave together.