GabAI

GabAI is a personal AI assistant for shluchim that runs on a server you rent and answers you on WhatsApp or Slack. It is a fork of NanoClaw, maintained by volunteers under the Chabad Commons GitHub organization. In the project's own words it "helps shluchim with scheduling, answering questions, managing WhatsApp groups, voice transcription, CRM automation and more."

The project calls itself "secure, free, open-source." The license is MIT, but the main repository is private: you request access before you can install it.

Who it is for

A shliach who wants an assistant that is always on, works inside WhatsApp groups or Slack channels with a team, and keeps donor and family data on a server the shliach controls rather than with a startup.

It is not yet for someone who has never opened a terminal. The builders themselves said so: in April 2026 the admin held off a wider rollout because "conversations have been about fixing bugs," and a non-technical tester said it needs a way to set up and troubleshoot "not using a terminal" before it can spread. By October 2026 the installer had improved a lot, but you still log into a server over SSH and you still debug by asking Claude Code to read logs. If that sounds like too much, look at Claude Cowork or the hosted agents on Personal AI Agents, with the privacy cautions there.

How it differs from plain NanoClaw

Members who run both describe GabAI as NanoClaw plus shluchim-specific pieces:

• A zmanim engine and a Shabbos mode, so reminders and outbound messages hold for Shabbos and Yom Tov (see AI and Shabbos).
• Extra hardening in the installer: firewall, fail2ban, automatic security updates, Tailscale, and binding the credential vault so it is not exposed to the internet.
• An owner-group model: only your own private chat with the agent gets Google Workspace, email and similar access. Agents in shared groups are unprivileged and must ask the owner agent for approval.
• The community skills library (below).

GabAI v2 is built on NanoClaw v2. As of August 2026 it was still a release candidate. In October 2026 the admin said the latest build supports the newest Claude models and includes an upgrade skill (/gabai:upgrade). If you customized heavily and only want newer models, the admin's advice was to ask the host to update the Claude Agent SDK instead.

The setup path

Steps as the group's guides and members described them in 2026:

1. Request access. Open the "Request access" form linked from the Chabad Commons page. You need a free GitHub account; accept the organization invitation when it arrives. The installer fails with "permission denied" until you are approved.
2. Rent a server. Ubuntu 22.04 or 24.04, 2 vCPU and 8 GB of RAM recommended. Shluchim with Microsoft's nonprofit grant mostly use Azure; others use Hetzner, DigitalOcean or AWS. Options and prices are on Agent Hosting.
3. Run the installer. The link comes with your access. It is an interactive full-screen program for standard 64-bit (amd64) Linux servers; ARM servers are not supported yet. It sets up the server, Docker, the credential vault and your first agent.
4. Connect a model. Run claude setup-token on the server and use the token it gives you, so the agent runs on your Claude subscription instead of pay-per-use API billing. The installer accepts the token up front so you sign in once.
5. Connect a channel. Scan a QR code for WhatsApp (a dedicated number is better than your own) or create a Slack app.
6. Learn the host/agent split (next section) before you start adding integrations.

How long it takes depends on you. One beta tester in April 2026 needed "about an hour or two" once he had the right server size; another member's first install, with a nonprofit-grant wait and token trouble, took about a week of on-and-off work.

Host and agent: the one concept to learn

The group repeated this more than anything else:

• The agent is the bot you message on WhatsApp or Slack. Use it for daily work: "summarize this group," "draft a reply," "remind me after Shabbos."
• The host is a Claude Code session in the terminal on your server, started from the GabAI install folder. Use it to install skills, connect Google or a CRM, change settings, read logs and fix problems.

In the admin's words: "Integrations on the host (terminal). Operations ('read my emails and answer them') on WhatsApp." Connecting Google Workspace from inside WhatsApp is the wrong way round. Skill commands such as /gabai:skills are host-only on purpose, because a container that can install software is a security hole.

A workflow the admin recommends for new features: describe what you want to the agent, let it interview you, ask it to "write this up as a feature request," then tell the host there is a feature request waiting and to build it.

If an answer is over your head, tell the agent "ELI5." To make that permanent, ask it to add "I'm not a software engineer; always explain things in simple non-technical terms" to its CLAUDE.local.md file.

The skills library

Skills are packaged abilities you install from the host. The community library ("GabAISkills") takes contributions by pull request, and the maintainers review before publishing. The admin said in September 2026 that the skills also work on other agent platforms, not only GabAI.

Skills members reported using in 2026 (check the current list on the host; names change):

• Voice-note transcription, and voice replies through ElevenLabs or OpenAI
• Daily or weekly WhatsApp group summaries, and search across message history
• Google Workspace (Drive, Docs, Sheets, Calendar, Contacts) and a separate Gmail skill
• Perplexity research, web reading, and PDF and image reading
• An inter-agent messaging skill, so a group agent can ask your private agent for a calendar check or an approval
• Emoji reactions as status (eyes means "seen, working on it")
• A Karpathy-style wiki for shared notes (see The LLM Wiki Pattern)
• "ADHD-Friendly Replies," the first skill a shliach contributed back (July 2026): replies that lead with the next action, numbered single steps, and lists capped at five

To share your own: in a fresh host session use /customize to describe the skill, test it, then ask the host to open a pull request on the skills repo. Describe it the way you would to a new human assistant who has never worked for you: what triggers it, the steps, where results go, and the exceptions.

Slack or WhatsApp?

The group's working answer after six months:

• Slack for your private agents. It is a proper bot API, so it is far more stable. Threads keep projects apart, search is good, and the agent can show what it is doing as it works. Slack's free plan keeps 90 days of history in Slack; the agent's own data stays on your server regardless. Use Socket Mode so you do not need a public web address.
• WhatsApp where the people are. Community groups, family, and anyone who will not install Slack. It runs over an unofficial library, so expect disconnects and re-pairing, use a dedicated number, and keep the library patched. Details on WhatsApp Automation.

Many members run both: one general agent on both channels, specialist agents on Slack only. One heavy user dropped WhatsApp entirely in June 2026; the main thing he lost was forwarding voice notes to the agent.

Two Slack mistakes members made: running /install-slack-app (that installs Anthropic's own Claude app, not your bot; GabAI's setup starts with /add-slack), and forgetting to invite the bot into each channel.

Costs

• Software: free.
• Server: $0 if your Microsoft nonprofit Azure grant ($2,000 a year) or AWS credits cover it; otherwise roughly $10 to $80 a month depending on provider and size.
• Model: your existing Claude Pro, Max or Team subscription through claude setup-token. Some members switched agents to a ChatGPT subscription through the Codex provider during the mid-2026 billing scare and stayed.
• Extras: an OpenAI or ElevenLabs key for voice, at a few dollars a month for typical use.

The expensive mistake: one shliach's agent had been set up to reach Claude through AWS Bedrock. His nonprofit AWS credits covered the server but not the Claude usage, which AWS bills as a Marketplace charge, and the bill ran to several hundred dollars a month before he noticed (July 2026). Full story on Agent Hosting.

What it does well, and what it does not

Wins members reported: CRM cleanup and donor logging through APIs, student lists built from a campus CRM, a voice-note "capture" channel, event photos sorted by face recognition (stored locally, with an open question about biometric privacy law), a website built from a single prompt and edited by voice notes, and an agent that audits an events spreadsheet every three days. One shliach said it costs "a literal fraction" of his virtual assistant.

Problems that came up again and again: agents going silent after a crash or an expired token, WhatsApp "Waiting for this message" errors, forwarded WhatsApp messages the agent cannot read, and agents confidently describing features they do not have. The group's rule: trust the host, which can read the code, over the agent's description of itself.

What changed in 2026

• March: The group formed and chose the name GabAI. Repos lived under a ChabadLabs GitHub organization at first.
• April to May: Beta testing, a skills marketplace, and the move toward NanoClaw v2.
• June to July: The repos moved to Chabad Commons; GabAI v2 release candidates; the host/agent rule was written down after many beginner mix-ups.
• August to October: Installer hardening, Shabbos and zmanim features, the first community-contributed skill, and an upgrade skill.

Related

• NanoClaw, the upstream project
• Hermes Agent, the alternative some members chose
• Agent Hosting and Agent Safety
• Personal AI Agents
• The ChabadLabs AI agents webinar