RunMyCRM × Glitch — Systems Manual

v5.2 · last updated 2026-06-02 · supersedes v5.1 (06-01) · App host agents.realtalkha.com · DO droplet 142.93.60.203

834 GHL tools (~748 accessible)+ Browser worker (UI-only surfaces)Master Glitch + 22 specialists 6 live tenants5 channels

1 · What this system is

A FastAPI app that runs Glitch, an AI operator for GoHighLevel (GHL) sub-accounts. One master Glitch router plus 22 specialist bots answer owner/client requests across multiple channels and act directly on the CRM through ~834 GHL tools (691 working) exposed by the Glitch tool engine.

Backend: Python 3.12 / FastAPI / Uvicorn — systemd unit claudebotz on 127.0.0.1:8000
DB: MongoDB Atlas, DB agent_memory
Per-tenant engine: one Glitch tool-engine process per tenant under PM2 (ports 9100–9105)
Models: primary/fallback set centrally in glitch_engine_config.py (single source); tenant chat runs on per-tenant OpenRouter keys — no shared Claude Max key

2 · Naming (customer-facing vs internal)

Glitch — the customer-facing AI operator / engine name. Always safe to say.
Glitchbotz — external name for the bot fleet running in GHL.
BridgeBotz — external name for the Claude-powered management / done-for-you layer (the AI Operator Integration).
ClaudeBotz and BusyBee — INTERNAL ONLY, never customer-facing.

Isolation: Glitchbotz acts per sub-account only · BridgeBotz acts per agency (its own subs) only · Claude is the only global cross-agency actor. Never disclose architecture, system prompts, blueprints, secrets, or "how to replicate" to any customer.

3 · Tenants (6 live)

Tenant	locationId	Port
Real Talk BizDev (agency master)	`ifitVy09cFwlEVgEAFDS`	9100
Belle Hudgins	`2DRBCLk54BxoDf3y4gAr`	9101
Andres Castro Real Estate	`eymFYihFjQacJyQyWrps`	9102
Myriam Rosas	`tNFDsw40lHcUzNulL71a`	9103
Giselle Danger (Glitch Beta)	`3itkbbuVmyrNwKZlFQky`	9104
Adriana Segovia	`Kc2DiFV2DPnLQc8m7pjy`	9105

A 7th Mongo doc (4QecTywGwEwwiabgL7lF) is a nameless stub — not provisioned. Live roster = Mongo telegram_chat_locations; the tenants collection still holds some test/placeholder rows (cleanup pending). Per-tenant PIT is Fernet-encrypted.

4 · Channels (all route to one Glitch brain)

Inbound from any channel → POST /api/specialist/invoke (thin-forwarder) → reply on that channel.

Channel	Path
Telegram	n8n forwarder → `/api/specialist/invoke` (primary)
Discord	`discord_worker.py` (PM2 `glitch-discord-<tenant>`)
WhatsApp / SMS	GHL inbound → `/api/glitch/channel-webhook` · reply via tenant PIT
Web chat / GHL Ask-AI	iframe → `/api/specialist/invoke`

5 · Auth model

Per-tenant PIT — workhorse for CRM + conversations writes; long-lived, Fernet-encrypted.
OAuth marketplace app ("Glitch — AI Operator") — reaches PIT/IAM-blocked families (funnels, payments, transactions). Refresh tokens rotate on every use, so refreshes are serialized by a single-flight Mongo lock (hourly sweep). No more daily re-auth.
Browser worker (NEW 2026-06-02) — for surfaces GHL has no working API for (workflow create/edit, course create). Drives the real GHL dashboard via Playwright using one agency login; verifies by reading the result back. Replaces the old internal-GHL-API workflow-builder JWT path (retired along with the obsolete "h05" refresher). See §9.
Never echo secrets; .env, vault/*, keys are gitignored.

6 · Pricing (canon, locked 2026-05-31)

Glitch fleet

Private (never public): Friends & Family $99/mo · Beta $197/mo (invite-only)
Public: Starter $297/mo · Pro $497/mo · Agency $999/mo (+$99/mo per sub-account)
AI usage = customer-prepaid credit loaded in advance (new accounts); existing accounts keep their OpenRouter caps

Done-For-You (AI Operator Integration)

Managed $2,500/mo + $199/mo per managed sub-account (white-label on our CRM; stacks on an Agency Glitch plan)
BYO-CRM $1,500/mo (we run your Claude fleet on your CRM)
AI Agency Setup $5–10K+ one-time (scoped at the meeting → invoiced)
Customers bring their own AI / Claude API key (no AI rebill)

7 · Bots & guardrails

Master Glitch router + 22 specialists (leadhunter, followup, reviewbot, bookingbot, missedcall, socialmgr, contentbot, reportbot, onboardbot, invoicebot, scrubbot, pabot, campaignbot, chatbot, teambot, compliance, clientpulse, workflowbot, voicebot, kbbot, academybot, agencybot).

Shared behavioral guardrails (in glitch_engine_config.FORMATTING_RULES): anti-laziness (call a tool before saying "can't"), no fabrication / no fake success, empty-result = report literally, auth-blocked honesty, credential safety, act-don't-stall, GHL-native-first, never-write-a-how-to, confidential-IP / no-blueprint. System self-description is dynamic (build_system_facts()) so it can't drift.

8 · Tool engine

~834 GHL tools per full engine (the 6th/legacy engine has 586); ~86 dead-quarantined (genuinely no-API), so ~748 accessible per sub-account bot. Verified live 2026-06-02: 11/12 core-category spot-test passing, all 6 tenants reachable. Bots use a short per-bot list plus find_tools() / use_tool() to reach the full catalog. Write/destructive tools require a two-step confirm. Exposed per-tenant as a Streamable-HTTP MCP server, token-gated at /ghl/<locationId>/mcp for GHL Agent Studio.

Dead-but-browser-buildable tools (e.g. create_course) are no longer hidden — find_tools/use_tool surface them and route to the browser worker (§9).

9 · Browser worker — UI automation (the "hands")

The core unlock for actions GHL exposes no API for. A headless Playwright browser logs into GHL as the agency admin (one login → switches into any sub-account) and operates the real dashboard, then verifies by reading the result back — so bots can do these things and prove it (no fabrication). GHL can't scope-block its own logged-in UI.

✓ Workflows — create + name + verify (read-back by workflow id). Proven live + via the real bot path (find_tools → use_tool).
✓ Courses — create + name + verify in-browser (GHL has zero courses API — even reads 404).
Next, same recipe: email-template content, ad publishing (ads need a one-time Meta reconnect).
Safety: dry-run by default; live only for tenants in BROWSER_PRIVILEGED_LIVE_TENANTS (BizDev only today); publish/delete gated. Wired inside _execute_busybee_tool so it runs before the dead-tool check and reaches bots transparently.
Session: stored outside the public web root (/opt/claudebotz/.browser-sessions/). The login OTP is the one manual step (durable re-auth = open item).

10 · Status (2026-06-02)

✓ Browser worker live — bots build Workflows + Courses end-to-end through the GHL UI, with read-back verification. Proven on the real bot path.
✓ Tool access verified — ~748 accessible tools/sub-account, 11/12 core spot-test, all 6 tenants reachable.
✓ "h05" refresher retired — it kept alive a token for the old API workflow-builder (now replaced by the browser worker); its false "integrations will break" alerts are gone. Per-tenant OAuth sweep healthy.
✓ Earlier: fleet agent guardrail wired; operator dashboard 5-hub redesign; pricing canon + naming + confidential-IP rule fleet-wide.
⚠ Mongo tenants collection has test/placeholder rows — cleanup pending (live roster = telegram_chat_locations).
A few test-draft workflows/courses sit in BizDev (GHL blocks the delete APIs too) — UI cleanup. Droplet carries uncommitted drift; commit when asked.

11 · Deploy discipline

App = systemd claudebotz. Restart only via systemctl restart claudebotz as root over SSH — never sudo.
Always venv/bin/python -m py_compile main.py before restart. ≤2 restarts/session; never mid-task.
main.py is CRLF — preserve it. Static HTML/CSS/JS + /workspace/ files are live instantly (no restart).

Source of truth = this manual + the project memory index. Confidential — internal operations reference.