The Builders Group and Its Coding Tools
AIFS Builders, the AI for Shlichus subgroup where shluchim who build apps, sites and agents compare work, and the coding tools they use, with October 2026 prices.
On this page
Most of the apps on Community-Built Chabad Apps were first shown in one place: AIFS Builders, a WhatsApp subgroup of the AI for Shlichus community for shluchim who build software, websites and AI agents.
The group
The group admin opened it on February 22, 2026, in the middle of the Megillah app sprint, so that long technical threads would stop crowding the main group. The first poll favored keeping one group (32 to 11); a second poll a few days later favored a separate one (14 to 7), and the split went ahead.
To join, ask in the main AI for Shlichus group. This page does not publish invite links.
Members range from shluchim who had never opened a terminal to people who write software for a living. A typical thread: someone posts a working link, others test it on their phones, and requests and bug reports come back within the hour.
What the group talks about
A sample of the threads from February to October 2026:
- Show and tell. The Megillah app, maamorim.app, the Shluchim Exchange Haggadah creator, two Kiddush HaChodesh models, a Pesach sources index, CountOmer and a Tanya map were all posted here first.
- Getting Chassidus text into apps (March 2026). The admin called a "Sefaria for chassidus" the key to the next stage. Members pointed out that much of Kehot's library has never been typed in, and that existing plain-text sources such as dach.dev have typos. They argued over whether an AI can be trusted to fix OCR errors "based on meaning" (several said no). One member built a crowd-transcription site, Sofrim B'yachad, that same night: each line goes to three people and is kept only if all three agree. The copyright side of this is on Chassidus Copyright and Access.
- WordPress or a custom site (March 2026). One side wanted an AI-written WordPress theme that office staff can update through a simple form. Two builders said they had left WordPress for Claude Code builds. A third, with more than ten years on WordPress, keeps it because staff already know it. The conclusion: decide based on who will update the site after you build it.
- Personal agents (from March 2026). Setting up and fixing agents on a rented server took over a large part of the group. See Personal AI Agents, GabAI and Agent Hosting.
- Which model to code with (September 2026). Answers varied. One builder codes with Opus and calls in Fable for long, complicated jobs and a final review. The admin found Fable 5.1 strong but hungry for usage, and set a rule in Claude Code that sub-agents never run on Fable. Another builder has a second company's model review the code at checkpoints and said it catches bugs Claude misses. His summary: "the model isn't the whole story." For the current models, see Claude.
- Duplicate projects (September 2026). A member asked everyone to post what they were building, after several people turned out to be making similar reading and Rambam tools. Some worried about ideas being taken; nothing formal came of it.
- A shared Chabad House website (October 2026). A poll asked who had rebuilt their Chabad House site with AI (one had, two had rebuilt pages). Members floated an open-source site that every Chabad House could use, built together like the Megillah app. An experienced builder was skeptical: the Megillah app was a short, one-time job, and a site you maintain all year, alongside Chabad.org's systems, is far harder.
How the builders work
Practices that came up again and again. None is required; they are what experienced members do.
- Plan before you build. Ask the AI for a plan, read it, and fix it before any code is written. Some builders send the plan to a second model for review first.
- Work on a copy. Make changes on a separate branch, review them, then merge. Keep numbered versions so a bad update can be rolled back.
- Make it write scripts. For repetitive work, have the AI write a small script once and reuse it, instead of doing the job by hand every time. One builder also has Claude keep a checklist of its own past mistakes and check it before and after each task, and says he stopped running out of usage.
- Stop it when it wanders. If a coding agent starts doing something you did not ask for, stop it (Escape or Ctrl-C) and ask what it is doing.
- Let it keep going while you are away. Claude Code's Remote Control lets you continue a session from the Claude phone app. Members run long sessions inside
tmuxso closing the laptop's terminal window does not end them. - Use auto mode with care. In Claude Code's auto mode, a second model reviews each action instead of asking you every time; Shift+Tab switches modes. One builder preferred it to a blanket "always allow" list, because a command can be safe in one situation and harmful in another.
- Share what works. Reusable agent skills go to the community's GitHub organization, chabad-commons, by pull request.
A debugging prompt a builder posted in April 2026, still in use:
What it's doing wrong: [describe]
What I want it to do: [describe]
Rubber-duck this process and everything related to it. First decide whether
we should be doing it this way at all, or whether there is a simpler way.
Then make a plan to fix it. Add logs at the key points so we can see what
happens. No magic numbers.
The tools, October 2026
Prices checked on each vendor's page on 2026-10-21. "Coding agents" work on real code in a folder or a GitHub repository; the beginner builders (Lovable, Base44, Replit, Bolt) are compared on Vibe Coding.
| Tool | Who makes it | Price | Notes from the group |
|---|---|---|---|
| Claude Code | Anthropic | Included with Claude Pro ($20/month, or $17/month billed yearly); not on the free plan | The group's main tool. Also runs the "host" that maintains many members' agents. |
| Codex | OpenAI | Included in ChatGPT plans, from Free up | Praised from May 2026 for debugging and code review; some use it as a second reviewer next to Claude. |
| GitHub Copilot | GitHub | Free tier (2,000 completions and 50 chat requests a month); Pro $10/month; Pro+ $39/month | Members called the $10 plan the best value for someone already in an editor (February 2026). |
| Cursor | Anysphere | Free tier; paid plans from $8 to $200/month, with model usage billed separately | Strong, but watch usage. |
| Kiro | Amazon (AWS) | Free (50 credits); Pro $20/month | One builder writes his specs in Kiro before coding elsewhere. |
| Jules | Base tier of 15 tasks a day; Google's paid AI plans raise it | Works from a phone against a GitHub repository. |
For hosting what you build:
| Service | Free plan | Watch out for |
|---|---|---|
| Vercel | Hobby plan, $0 | Vercel says Hobby is "for personal, non-commercial use." For a Chabad House site, check whether you need Pro ($20/month). |
| Netlify | 300 credits | Each production deploy costs 15 credits, so frequent small updates add up. Personal plan $9/month. |
| Cloudflare Pages | Yes | Members called Cloudflare "basically free" for many small projects (October 2026). |
| GitHub Pages | Yes | GitHub says Pages is not meant for sites whose main purpose is commercial transactions. |
If your app stores anything in a Supabase database, turn on Row Level Security for every table. Supabase's own documentation says a table without it "is readable and writable by any role with a grant on it," which on many projects includes anyone holding the app's public key. In April 2026 Supabase warned one member that several of his tables, donations among them, had it switched off.
What changed in 2026
- Claude Code stopped being the only answer. Through spring 2026 the group's line was "all you need is Claude Code." By May, Codex was in daily use for review and debugging, and several builders now run more than one company's model on the same project.
- Prices and plans moved. Lovable can now import an existing project as a ZIP file (it still cannot connect to an existing GitHub repository). Kiro left preview and has paid plans. cto.new, once listed here as a free coding tool, is now a business-building service, so it is no longer in the table.
- From apps to agents. After March 2026 more of the group's energy went into personal agents than into apps.