📙 Glitch / RunMyCRM — Systems Manual v6.1 (2026-06-10, rev: brain repoint + portable onboarding)

Canonical current-state architecture + system manual. Supersedes v5.4 (Jun 03). Written 2026-06-10 after a major architecture session: the GHL-independence pivot, the polling watcher, per-account spend caps, monitoring dashboards, and the bot-endpoint security lockdown. New engineer or future-you after a gap — start here.

0. What changed since v5.4 (the big shifts this session)

Strategic pivot — Glitch is now GHL-INDEPENDENT by design. Glitch owns the brain (LLM); GHL is a swappable system-of-record it reads/writes through stable APIs. Browser automation only where no API exists, minimized. Down the road this means other CRMs, not just GHL.
KILLED: the AI-Employee bridge ("Direction B" — run the AI natively inside GHL Agent Studio to use AI-Employee-unlimited). GHL reshaped Agent Studio out of the bundle into pay-per-use + it's documented-flaky. Cost lever replaced by BYO-brain routing + Glitch Connect + the $100/account cap.
KILLED: GHL trigger-workflow self-publish. Research confirmed GHL has no public workflow write API (scope is workflows.readonly); the builder is UI-only + ephemeral-draft. Replaced by →
NEW: the Polling Watcher — GHL-reshape-proof reactivity. Per-account poller on our scheduler diffs the GHL read API for new leads / stage changes and fires Glitch (gated). Proven E2E.
NEW: Glitch Connect per-account spend cap ($100/mo included credits, self-enforced) + two monitoring dashboards (owner Usage & Activity, operator Fleet Monitor) + fleet health-check cron.
SECURITY: the bot endpoint is locked. /api/specialist/invoke (+async) now authenticate (was fully public — the #1 audit exposure). Enforce mode live.
NEW (milestone): 3rd Floor's brain is now the Glitch engine, not GHL Agent Studio. And portable onboarding stands up a working Glitch on ANY location (agency or BYO/external) from just a PIT — no snapshot, no menu links, no marketplace install. The "one Glitch engine, two front doors" architecture is real + proven (see §6 + §9).

1. The product in one paragraph

Glitch is an AI employee that lives inside a GoHighLevel sub-account and runs that business's CRM 24/7. The sub-account owner talks to it from their phone (Telegram today; WhatsApp / Discord / Web chat planned) and tells it what to do; Glitch goes into the CRM and does the work — follows up leads, books appointments, posts social, fills forms, sends invoices — without the owner logging in. It's modeled on the OpenClaw pattern (brain + senses + memory + skills + hands + a safety harness) but it is our engine, not GHL's AI.

2. The two layers (the core distinction)

	ClaudeBotz	Glitch
Who sees it	The agency / operator (Heather). Customers never see it.	The customer. The only thing they see.
Surface	Operator dashboard at `agents.realtalkha.com` (fleet management, approvals audit, AI cost per tenant, Fleet Monitor, skills, config undo)	Telegram chat + the Bot Safety page + Glitch Bots page (GHL custom-menu links)
Identity into GHL	Shared agency super-admin (`glitch-299` automation user) — full access to every sub-account	Its own scoped GHL Admin user assigned to ONLY that one sub-account • its own AgentMail inbox + capped OpenRouter key. Isolation proven.
Role	The factory / back office	The product / the bot

Important: it's mostly one codebase and one brain loop (main.py → run_specialist_agent_loop). The difference is identity + scope and surface, not separate engines. A sketched third tier — BridgeBotz (per-agency dashboard) — is planned, not built.

3. Architecture — the durable spine

Everything below depends on none of the GHL surfaces that keep getting reshaped (no workflow webhooks, no Agent Studio).

Two kinds of hands into GHL: