ppm — GitHub
Pascal Password Manager
A password manager I built and used between 2016 and 2018. All cryptography runs locally in your browser — nothing is transmitted or stored.
Python and Android implementations are also available.
Generate password
Because there are no third-party dependencies or remote services, the trust surface stays
minimal. The same strong password is deterministically derived each time from your master
password and a per-site keyword (e.g. blue-sky or orange-tiger).
Under the chosen scrypt parameters, a brute-force attack would take roughly 1019 years on a million MacBook Pros to reach the expected halfway point of the keyspace. (I wrote that estimate about 10 years ago.)
Revisiting the estimate in 2026
Hardware has moved on, so it's worth redoing the math. The parameters here are
logN = 15, r = 8, p = 1, which means each scrypt evaluation touches
128 · r · 2N = 128 · 8 · 32768 ≈ 32 MB of memory and performs on the
order of 4 · r · 2N ≈ 1.05 · 106 Salsa20/8 core operations.
Because scrypt is memory-hard, throughput is bounded by RAM bandwidth as much as by raw CPU.
2016 baseline. A mid-range MacBook Pro of that era (Intel Core i7, 4 cores,
~25 GB/s memory bandwidth) ran a single evaluation in roughly 300–400 ms, giving about
10 evaluations/second when fully parallelised across cores. A million such
machines therefore deliver around 107 evaluations/second, or
~3 · 1014 per year.
Halfway-point assumption. The original figure assumed a master-password
keyspace of roughly 2110 ≈ 1.3 · 1033 candidates, with
brute force expected to find it after searching half:
6.5 · 1032 evaluations. Dividing that by 3 · 1014
evaluations/year gives ~2 · 1018 years — within an order of magnitude
of the 1019 figure (the original was a back-of-envelope upper bound).
2026 hardware. A current MacBook Pro (Apple M-series, 10–14 performance cores, ~400 GB/s unified memory bandwidth) completes a single evaluation in roughly 15–25 ms per core. Multi-core throughput lands around 500 evaluations/second per machine — call it a ~50× speed-up over the 2016 baseline, dominated by the jump in memory bandwidth rather than clock speed.
Updated estimate. A million 2026 MacBook Pros deliver
5 · 108 evaluations/second, or ~1.6 · 1016
per year. Searching half the keyspace then takes
6.5 · 1032 / 1.6 · 1016 ≈ 4 · 1016 years —
still about three million times the age of the universe.
In other words: ten years of hardware progress shaved roughly two orders of magnitude off the number, and the answer is still "not in any human time frame." The memory-hardness of scrypt is doing its job — speed-ups track memory bandwidth, which scales far more slowly than transistor counts.