Blog

Notes from building the tools

Deep dives on the open-source tools behind slaps.dev, with real runs and real numbers.

Slapify · A five-part series

Building a browser agent that can do anything a person can

Slapify takes a goal in plain English and works a real browser until it's done: reading, clicking, typing, logging in, handling files and tabs, for a minute or for days. This series is how it got there, with real runs, real failures and real costs.

Start with part 1

Slapify shopping a demo store: it opens a jacket, rejects it on price, picks a cheaper waterproof one, selects size M, ticks gift wrap and adds it to the cart. — A real run on a demo store, 79 seconds end to end.

26.8 hlongest unattended run

$0.22per reply at Sonnet 5 API prices

348 msmedian Jev decision

0stealth tampering detected (CreepJS)

The series

Part 1 · Getting started

Your first autonomous task, start to finish

Install, one sentence of goal, and a step-by-step replay of the agent shopping a demo store, with what it cost.

Part 2 · Case study

We built a Twitter reply agent without Jev. Here's what broke.

Nine failures from days of unattended running, each fix, and the bill at API prices.

Part 3 · Decision model

Adding Jev: small decisions, made fast

A 58-decision benchmark against Sonnet 5 on real pages, calibration you can play with, and a live demo with review.

Part 4 · Browser

Stealth mode: patch nothing that isn't already a tell

Why patching less beats patching more, before-and-after on bot.sannysoft.com, and where script-level stealth stops.

Part 5 · Harness

A browser agent you can leave alone

Contracts, admission, review, recovery and compaction: step through a tool call's life and watch an outage play out.

All runs shown are real. Demo sites are fictional and local; third-party content from the Twitter runs is never shown, only counts and timings. Costs use list API prices.