← All work

Twin apps for a 3,500-title Plex library

Postarr chooses a poster for every film and show and puts it on Plex with one click and an undo. Reconcilarr checks that five systems agree about every title, and turns each disagreement into a fix you preview first.

Status
In daily use · v2 live
Runs
Postarr · Reconcilarr
Started
Apr 2026 · rebuilt Sep
Stack
FastAPISQLiteReact 19TypeScriptSSEPlex APIShared design system
The Arrs: main view
3,581Titles managed
617Commits across 3 repos
787Automated tests
37Brand rounds

The Problem

My media library spans five systems that each keep their own copy of a title's name, year and ID, and they drift. A wrong ID in one place puts the wrong poster on the wrong film. The bug that started it all: the old poster tool reported "all items already covered" while its network share was down and it was reading nothing.

How It Works

The workflow, step by step.

  1. 01

    An inbox, not a to-do list

    New seasons, set updates, new titles and anything Plex refused land in Postarr's Inbox, fed by Radarr and Sonarr webhooks and a 15-minute scan. A push notification says when to come.

    An inbox, not a to-do list
  2. 02

    Compare every candidate

    Posters from ThePosterDB, verified against TMDB so the page really is the right film, beside TMDB's own images and your uploads.

  3. 03

    Use it, with an undo

    One click downloads, finds the title on Plex, uploads and locks it. The toast says "Chosen · Undo", and Undo puts the old file back on Plex too. Use Set does a show and all its seasons at once, and says exactly what it replaces first.

  4. 04

    Check five systems agree

    Every night Reconcilarr reads Plex, Radarr, Sonarr, the media folders, Seerr and Postarr, then compares each title's folder, name, year, ID and poster.

    Check five systems agree
  5. 05

    Preview, apply, undo

    Each disagreement becomes a card with a fix named for what it does. The preview shows Now and After with IDs resolved to real film titles, and the fix writes through Postarr's API with an undo.

    Preview, apply, undo
  6. 06

    Native on the phone

    A tab bar, rows that swipe and press-and-hold menus that open at the finger, an A–Z index, pull to reload and swipe back. Built once in the shared repo to the family's phone standard, checked on an iPhone round by round.

    Native on the phone
  7. 07

    One library, at a glance

    A wall of 3,500-plus posters, filterable and sortable. Every night each poster is checked against Plex by its file hash, so anything Plex lost comes back to the Inbox.

    One library, at a glance

The Output

The brand: one unit, thirty-seven rounds.

Every dimension of both marks is derived from one measurement of the typeface, and the logo is drawn by code. Postarr's o is a poster; Reconcilarr's co is a Venn of two sources that should agree.

Round 1: ten directions
Round 1: ten directions
Round 4: a mixer with sliders
Round 4: a mixer with sliders
Construction review: the six rules
Construction review: the six rules
Every derived value, in a table
Every derived value, in a table
Every token with its contrast ratio
Every token with its contrast ratio
The twin marks
The twin marks

On The Phone

A real phone layout, not a shrunk desktop.

Reconcilarr's Inbox: only what needs me
Reconcilarr's Inbox: only what needs me
Postarr's library wall, with an A–Z index
Postarr's library wall, with an A–Z index
A finding, explained, and what it can't fix by itself
A finding, explained, and what it can't fix by itself
Every title, with one status line
Every title, with one status line

The Details

Where the rigor shows.

01

"A dead source is never a clean result"

Collectors raise an error rather than return a short list, and anything under a sanity floor counts as down. Rules missing an input are skipped and reported as skipped. An unmounted share and an empty one look identical, so the app refuses to guess.

02

"Never assert an authority you do not have"

Plex is wrong about IDs often enough that a conflict is shown as a choice between two resolved titles ("64691: a 1946 nutrition film" vs "64690: Drive, 2011"), not an automatic correction.

03

Matching that verifies, never guesses

The old matcher picked the wrong page for 3 of 77 sampled films and flagged none. Now a page is used only if its reference poster appears in the film's TMDB images. Otherwise the app asks.

04

Undo you can trust

Undo is captured at apply time from the value actually there, so it can only reverse what this process did. A fix whose value already matches doesn't clear the finding, because that would hide a live discrepancy.

05

Safety bugs caught before deploy

An audit found Delete Everywhere permanently removing media. It now moves files to a trash and needs explicit confirmation. Three GET routes that changed state became POST.

06

A migration rehearsed on a copy

4,841 posters were copied with every file hashed, refusing on any mismatch, into a database built in a temp file that only goes live when everything is in. The script refuses to run twice.

07

26 MB to 300 KB

The startup snapshot shrank 99% on the wire, the item list went from 737 ms to 71 ms, and thumbnails are cached by source fingerprint and pre-warmed after each scan.

08

Polite by design

ThePosterDB is read at 1.5 requests a second with retry on 429. Big set pages fill in as they arrive and are cached for 30 days. Background work yields to people.

09

Dates recovered by hash

Posters carried over from the old tool had lost when they were chosen. A script matched each file to its original by SHA-1 and took the old file's date: 4,930 of 4,930 matched on a copy, as a dry run, before it was applied on the box.

10

787 tests on fakes

Postarr has 282 pytest and 154 vitest tests, Reconcilarr 193 and 148, shared code 10, all on throwaway data with a fake Plex and a fake Postarr. They run before every commit, and a deploy ships nothing if one fails.

Architecture

How the pieces fit.

  1. Radarr · Sonarr · Plex · disk · Seerr
  2. Reconcilarr
  3. Postarr API
  4. SQLite
  5. Plex

Each app is one FastAPI + React container behind Cloudflare Access, built from one design system and one set of phone components synced from a shared repo. One owner per file: Reconcilarr never touches Postarr's database, it asks Postarr's API for every change.

How AI Built It

The process behind the product.

  1. Audits before rebuilds

    A design audit screenshotted about 40 states per app and found 26 issues. A functional audit ranked 14 bugs by harm and posed nine calls as trade-off cards with mockups. An architecture review found 27 more. I took all nine calls.

  2. Planned like a program

    Each phase got a model, an effort level and a role. Two Sonnet builders ran side by side, one on each app, while Opus ran the refactor. Each step logged its measured usage against the estimate.

  3. A 37-round brand

    Twin marks built from one unit (13.03 font units): one weight, one clearance, one circle. 734 lines of decisions, including the rounds I rejected and why.

  4. Fourteen steps, all shipped

    The overhaul plan ran to fourteen steps, each with an owner, a model and a measured result. The last ones put both apps on the phone standard, merged the worktrees and renamed the brand repo arrs-common, since it now holds shared code too.

  5. Handoffs a fresh session can pick up

    A progress file per phase, a plan page that marks each step done and measured, and one explicit "Next step" line, so nothing depends on a single long chat.

In Motion

Postarr: the wall, every title with its chosen poster
Reconcilarr: three findings, then the library's checks (requesters other than me shown as Friend)

More Screens