- Event: Web Summit 2026
- Dates: November 9-12, 2026
- Location: MEO Arena and FIL, Lisbon
- Host: Web Summit
I’m taking SoftDevStudio.com to Lisbon. We’re a small independent studio that builds databases, queues, proxies and network tools, for clients and in the open. All 28 of our projects are on GitHub under Apache 2.0, and each one has its own site. For Web Summit I’ve picked five of them to present: Abbreviatory, RadarModel, CodeInstances, HyperCrux and Watchkeeper. None of them is past version 1. Every one of them runs, and every one publishes its test numbers.
Why Web Summit
A few weeks ago I wrote about taking five infrastructure projects to Cloudflare Connect, and afterwards what Connect had to say about agents becoming the web’s busiest visitors. Connect is a room full of engineers. Web Summit is a different room. More than 70,000 people come through Lisbon that week, and a lot of them are founders, investors and journalists who have to decide quickly whether a thing is real.
That’s a good test for a studio. If I can’t explain a tool to someone who’ll never read its code, it probably isn’t finished. So this time the lineup leans toward problems anyone shipping software with AI in it already has, plus one that anyone with a browser has.
What I’m bringing
Abbreviatory makes tool output short enough to read. A coding agent runs the tests, gets thousands of lines back, and pays to read every one of them, often twice. Abbreviatory turns a test run, a log, a diff or an HTTP response into a short report with failures, errors and warnings first, all within a byte budget, and everything it left out stays one command away. There’s no language model inside it, so the same input always gives the same report. One sample log went from 63,369 lines and 5.6 MB to a 3,075-byte report. Version 0.3 added an MCP server so agents can call it directly, and the browser demo lets you paste your own output in.
RadarModel tells you when something your app depends on changes underneath it. RadarModel AI reads the model lists of OpenAI, Anthropic, Google and OpenRouter on a schedule and warns you when a model your app calls is retired, gets a retirement date or changes price. RadarModel Schema does the same for Postgres, MySQL and SQLite, catching the index someone added by hand at 2 a.m. without a migration. It scanned 2,000 Postgres tables in 0.41 seconds and found a dropped one on the next pass. The case study on stopping a deploy is the quickest way in.
CodeInstances keeps a run of a program as a file. Your code is versioned and kept forever. Its runs get thrown away. CodeInstances records a run with every HTTP and HTTPS exchange it makes, replays it offline, and tells you whether it came out the same. It can put a failing run next to a passing one, shrink the failure to the answer that causes it, and keep that as a test. In the self-checking demo, 23 calls were recorded into one 20,631-byte file and replayed ten times with the live servers switched off. All ten matched. A bug report that arrives as a runnable file is the part I most want people to see.
HyperCrux puts SQL, key-value, graph and vector in one native database. Four databases cross at one record, in one file and one transaction, so they can’t disagree. You can query a record with SQL, fetch it by key, walk its links and search it by vector. In testing, a writer was killed mid-write 200 times and nothing came out of step. The new Beta engine in Go searches vectors 11 times faster than 0.1 and shipped its first pre-release for Linux this month. The use case I’ll lead with is an AI assistant’s memory that forgets properly, because deleting a fact should delete its embedding too.
Watchkeeper is the newest of the five and the only one you don’t need to be a developer to use. It’s a tiny browser for Windows, macOS and Linux with a panel beside the page. Open a news site and the panel lists the other companies the page told about your visit, the cookies it left, what it wrote to your computer and which scripts read your device’s fingerprint. The session is wiped when you close it. We tested it against a page built to fool it, which made 73,827 junk events to hide behind. Watchkeeper still found the tracker, and all 14 of 14 addresses a test tracker was asked for showed up in the panel. That’s at how it was tested.
Five projects, one habit
Put them side by side and they share a habit. Each one shows its work. Abbreviatory points to every line it left out. RadarModel says what changed and when. CodeInstances keeps the run so nobody has to argue about what happened. HyperCrux can’t drift out of step with itself. Watchkeeper shows you the traffic a page would rather you didn’t see. Each project’s site has its test results on the home page, with the logs on GitHub.
That’s what I’m bringing to Lisbon. If you build software, I’d like to hear which of these you’d use on Monday morning. If you invest in it or write about it, ask me which one I’d bet on.
I’ll have an answer.