The Tools I Use, What They Cost, and What I Dropped

Work covered: January to February 2025 and June to September 2026.

One page, sorted by job

On 2026-09-16, 875 of the 1,003 commits in my main workspace repository carried a line crediting Claude as co-author, counting from the first commit on 2026-06-29. That figure has two limits. It covers one repository only, since a few projects keep repositories of their own, this website among them. And it does not mean one agent built everything, because a handful of commits credit other agents.

The tools below are the main ones doing real work across my projects: the apps, a trivia channel that renders and publishes on its own, a game prototype, and this site. They are sorted by the job each one does, and each entry says which project it worked on and what it cost when I have a record.

The subscription and credit figures come from the expense log I keep for these projects. Its latest entry is dated 2026-08-03, so it shows charges from July and early August only, and each figure is what was charged on the date shown, not a list price from a vendor's pricing page. I have no recorded charge for Remotion, Meshy, Supabase, Sentry, Vercel or Postiz. The Apify figures are usage readings from Apify itself. Three.js, Rapier, Next.js, Keystatic and Blender are free to use.

Build: two coding agents from two companies

Claude Code does most of the engineering, on Anthropic's Claude Max 20x plan, carried in my log at $216 a month as of 2026-07-17.

On 2026-07-04 I also bought $49 of Claude API credits, which are billed per use. A subscription is not a developer API plan, so code that calls Claude directly through Anthropic's SDK (the code library for calling its models), as the It Quiz engine can, needs metered credits of its own.

Codex, OpenAI's coding agent, is the second agent I run. It builds too, but its most distinctive job is cross-model review, which means a model from a different company reads work the first one produced. A second model misses different things than the first. On Helpings, my recipe and meal-planning app, a Codex review found that an old version of the function for joining a household was still callable and skipped the limit on wrong join-code guesses. Revoking it became its own database migration. Codex also co-authored Helpings migrations in July and review fixes on Rack Up in August. The only OpenAI charge in my log is ChatGPT Pro, at $108 a month as of 2026-07-11, and I have not split out how much of that goes to Codex. I wrote up what agent review caught on Rack Up, and what it could not separately.

Voice and video: Remotion, ElevenLabs and CapCut

It Quiz What It Is is a faceless trivia channel on YouTube Shorts, meaning no presenter on camera. Publishing went automatic with the videos for 2026-07-10, and since mid-July the channel has posted two Shorts a day. Claude writes the questions ahead of time and checks the facts against the web, in batches from agent runs I start myself, and I get a warning when fewer than seven days of questions are left. A nightly task renders each queued item into a narrated vertical Short and schedules it to publish.

It did not start that way. In early July the engine wrote questions and commentary itself, through Claude's API or a free local model. On 2026-07-11, the same day a commentary line from the local model told viewers a disaster involved fireworks storage when it was a propylene tanker, I took generation out of the nightly run. Claude now writes the whole queue ahead of time.

Remotion is the renderer. It builds video out of React components, so each Short is code plus one row of data, with no design tool and no hand editing in between. Before the rebuild went ahead, a Week-0 spike (a deliberately small test before the real build) rendered one complete video in code on 2026-07-02. Remotion's license, as I read it, turns on the size of your company rather than how much you render, so check the current terms against your own situation.

I’m an independent ElevenLabs affiliate and may earn a commission if you purchase through this link.

ElevenLabs (paid link) is the voice for It Quiz. Each question and answer reveal gets its own narration clip. The Shorts read the question and correct answer aloud, with the explanation left on screen; the long-form episodes also read the explanation. The Shorts' opening hook stays on screen unread.

The clips are saved on disk, so re-rendering with the same narration and those saved clips does not generate and bill the lines again. That depends on keeping the clips: the full Shorts pipeline clears them before a rebuild. If narration fails, the renderer falls back to a music-only cut and flags the problem.

I paid $65 for a year on 2026-07-04 and $176 for a one-year upgrade on 2026-07-19. The same plan also covers audio for other projects.

ElevenLabs and the ElevenLabs logo are registered trademarks or trademarks of ElevenLabs, Inc. Used with permission. Up & Adam is not sponsored by, endorsed by, or affiliated with ElevenLabs except as an independent participant in the ElevenLabs Creator Affiliate Program.

CapCut is the editor for videos I film myself. In July 2026 I started a challenge to post one short video a day for 31 days, in one take with the stumbles left in. I edited them in CapCut, and from Day 3 its transcript was the starting point for the captions. It was $13 a month as of 2026-07-08.

3D: Meshy, Blender and free model packs

Elsewhere is a browser action game I am building with Three.js for rendering and Rapier for physics, on Vite and TypeScript. It is a prototype and has not been released, so when I say a model shipped below, I mean it went into the game build, not in front of players.

Its base art comes from Quaternius: a downtown city kit, a stylized nature kit and a universal animation library, all from the free tier and released as CC0, a public-domain dedication that allows commercial use without attribution.

Meshy generates the props those packs do not cover. The agents drive it through its MCP server (MCP is a standard way for an AI agent to call another service's tools directly), live in both Claude Code and Codex since 2026-09-12. That day, 415 credits went into generation and 12 of the resulting models shipped: a water tower, a drone, a van, a dumpster, a newsstand, a fire hydrant, a conduit, a fire escape, a scaffold, a rail signal, a crate stack and a crystal pickup. That left 685 of the 1,100 credits budgeted for September.

The first 45 credits bought the lesson. Three previews on the meshy-5 model, at 5 credits each, came back lumpy and organic and looked wrong beside the Quaternius kit. A low-poly preview at 20 credits plus a 10-credit refine (Meshy's second pass, which paints the texture), 30 credits a prop, matched it, and that refined water tower was the first prop to ship. Every prop after it went the same way. Two prompts still missed. The first try at the crystal pickup, which the game calls a focus, produced a carved head. The note I kept says to ask for a gem with no figures on small objects, and the second try shipped. A subway entrance came out with stairs that climb instead of descend, so it is parked.

One prop came from Blender instead. On 2026-09-07 an agent working through the official Blender Lab MCP server modeled, exported and rendered a door lantern that went into the game.

Research: Apify

Apify runs ready-made scrapers, which it calls actors. On 2026-09-07 a Reddit scan collected 855 posts from game development and AI coding communities. The scan, including a second keyword run that was stopped for being too slow, came to about $1.60 in Apify usage. Around the 2026-07-20 research for Helpings, which pulled 146 one-star and two-star Google Play reviews, the account's usage for that billing cycle read $1.67 against a $10 cap.

Ship, host and post

Helpings is built with Expo, and Expo's EAS service runs its cloud builds on the free plan (as of 2026-09-16), which caps how many Android builds you get each month. Supabase handles sign-in, the database and server-side functions behind it, and Sentry reports crashes from both Android and iOS.

This site runs on Next.js 16 with Keystatic as its editor, hosted on Vercel. Keystatic saves content straight into the site's code repository, so every edit, including the next update to this page, is a commit to the main branch.

Postiz is the least tested tool here: a seven-day trial I started on 2026-09-16. It is the scheduler I picked for some of the posts I approve in my own dashboard, and nothing goes out without my click. As of that date I had not connected any accounts to it.

Tools I dropped

Around January and February 2025 I built the first version of It Quiz. Questions lived in a Google Sheets database, Make.com and Zapier moved them along, three Canva templates turned them into videos, and FFmpeg added the audio before each upload. It stalled for two reasons: the automations got complicated and expensive, and building the videos by hand in Canva became a lot of work. I abandoned the channel. The Remotion rebuild exists to remove that editing step and the glue around it. I have no record of what Make.com, Zapier or Canva cost in 2025. FFmpeg is the one piece that carried over, and the new engine still uses it to measure voice clips and grab each thumbnail frame.

The second departure came during the rebuild, when I dropped Adobe Stock as the source of the music beds (the looping track under the narration). From 2026-07-03 the Shorts used licensed Adobe Stock tracks, and those tracks still drew Content ID claims or blocks. Content ID is YouTube's automatic copyright matching, and a license to use a track does not stop it from flagging that track. On 2026-07-15 the engine briefly fell back to a synthesized bed and, the same day, moved to tracks from the YouTube Audio Library, which YouTube provides for free use in videos on its platform.

Make.com, Zapier, Canva and Adobe Stock get no links from me.

Affiliate programs, as of 2026-09-16

An affiliate program pays a commission when a reader buys through a tracked link. Postiz approved me for its program on 2026-09-16, the day my trial started. That is far too little use to recommend it on, so this page does not link to it.

Update, 2026-10-11: I have since joined the ElevenLabs affiliate program. My link and disclosure are in the voice section above.

Apify, Meshy, CapCut and Vercel run affiliate programs too. As of 2026-09-16, I had not applied to any of them.

For OpenAI, Anthropic, Expo, Remotion, Supabase and Sentry, I found no cash affiliate program. OpenAI runs time-limited referral campaigns, Anthropic's guest passes appear to pay in product credit rather than cash, Supabase's partner programs are for integrations and agencies, and Sentry offers credits to startups.