Global Claw GLOBAL-CLAW MASTER BIBLE Version: v0.1 (Feb 2026) Mission: Sell instant, preinstalled AI automation tenants globally. User pays → tenant is live → Telegram bot + dashboard works with zero setup. Non-Negotiables 100% Cloudflare (compute, queues, workflows, storage, auth edge, routing) Telegram-only to start (no WhatsApp/SMS/email v1) 1-click “plugins” (OAuth connectors + prebuilt packs) Cheap-LLM first with enforced budgets + failover No human-in-loop after payment Isolation fail-closed: compromised tenant can’t impact others A) Product Definition What the customer buys A tenant with: A Telegram bot pre-wired to default “packs” A web dashboard A plugin marketplace with 1-click connect Usage + billing + limits LLM routing policy (cheap-first) What “preinstalled” means Tenant provisioned with: Default Telegram workflow pack enabled Demo data + templates Safety limits pre-set Subdomain created API key minted Router config attached B) Cloudflare-Only Stack (Default) Compute / Orchestration Workers: API control plane, webhooks, router Durable Objects: per-tenant agent state + isolation Workflows: provisioning state machine + retries (OBSERVED Workflows GA) Queues: async jobs (message handling, plugin sync, audits) Data D1: tenant metadata, config, usage counters R2: templates, exports, logs, assets AI plumbing Cloudflare Agents SDK for agent/message patterns + tool handling (OBSERVED changelog) AI Gateway (optional) for central logging/ratelimiting of LLM calls (keep provider-agnostic) C) High-Level Architecture (Multi-Tenant) What is this? 1) Tenant Isolation Model (Recommended) Mode C (Hybrid): Shared Control Plane + Per-Tenant Durable Object Agent Why: maximum “agentic” behavior with Cloudflare primitives + fast provisioning + strong isolation boundaries. Isolation Guarantees (INFERRED) Each tenant has a dedicated Durable Object instance: Separate state, rate limits, budgets, connector tokens All cross-tenant calls must go through Control Plane: Verified tenant_id → DO namespace routing Fail-closed: If tenant mapping is missing/invalid → 403/404 If budget exceeded → hard stop If abuse signals high → auto-suspend tenant Provisioning time target < 10 seconds from webhook “paid” → tenant usable (Workflows + DO constructor + D1 inserts) 2) Telegram-Only v1 (Zero Cost Messaging) Telegram Bot topology One bot per tenant (recommended for white-label feel) PRO: easy branding, separation, easier abuse containment CON: bot creation requires token → must automate creation or use a shared bot initially Practical v1 choice (INFERRED) Start with “Shared Bot + Tenant Routing” for zero manual steps: One global bot token Each user links to tenant via /start <tenant_code> Later: upgrade plans can “bring your own bot token” (BYOBOT) Telegram webhook handling Telegram → POST https://api.global-claw.com/tg/webhook Worker validates Telegram secret header (if configured) Worker extracts: chat_id, user_id, message, tenant_code Worker routes to tenant DO: env.TENANTS.get(tenant_id).fetch(...) 3) One-Click Plugins (Cloudflare-Only Pattern) “Plugin” = Connector + Pack + UI Toggle, activated with one click. Plugin framework (INFERRED) Each plugin has: manifest.json (metadata) scopes (permissions) oauth (optional) jobs (background sync) packs (workflow templates) ui (dashboard toggles) Plugin types No-OAuth plugins (1 click enable) Webhook inbound Google Sheets via “copy API key” (later) OAuth plugins (true 1-click) Google (Gmail/Calendar/Drive) Microsoft (Outlook/Graph) Slack, HubSpot, etc. OAuth on Cloudflare (INFERRED) /oauth/:provider/start Worker → redirect to provider /oauth/:provider/callback Worker → exchange code → store refresh token (encrypted) in D1 Tenant DO marks plugin active Token storage rule: only store refresh tokens encrypted; short-lived access tokens cached in DO memory. 4) Cheap LLM Router (DeepSeek-First) Provider shortlist DeepSeek API (primary cheap) — has cache hit/miss pricing (OBSERVED) Kimi / Moonshot (secondary) — UNKNOWN pricing in this doc; verify current API pricing page before enabling MiniMax (secondary) — UNKNOWN pricing in this doc; verify current API pricing page before enabling Premium fallback optional (OpenAI/Anthropic) — disabled by default DeepSeek pricing (OBSERVED) DeepSeek docs list per-1M token prices and separate cache hit/miss input + output. Router policy (INFERRED) Task classes: chat_low_risk → DeepSeek chat long_context → Kimi/MiniMax (if needed) high_accuracy → premium (only if tenant allows) Guardrails: Per-tenant daily token budget Per-tenant per-minute rate limit Circuit breaker on provider errors Caching: Semantic cache for FAQ-like messages (R2 or KV-like; prefer R2+D1 index) DeepSeek cache features used where supported Router pseudocode function route(req, tenantPolicy, task) { enforceBudgets(req.tokens, tenantPolicy); const candidates = [ {name:"deepseek", weight:70}, {name:"kimi", weight:20}, {name:"minimax", weight:10}, ]; const ordered = prioritizeBy(task, tenantPolicy, candidates); for (p of ordered) { const r = callProvider(p, req); if (r.ok) return r; } if (tenantPolicy.premiumFallback) return callPremium(req); throw new Error("All providers failed"); } 5) Billing + Provisioning (Zero-Touch) Stripe flow (INFERRED) Hosted Checkout Webhook → Workflows provisioning Idempotency keys everywhere: stripe_event_id + customer_id Provisioning state machine What is this? Outputs tenant_id tenant_subdomain tenant_api_key telegram_start_link 6) Dashboard (Fast to Build, Minimal Surface Area) UI principle 3 tabs only: Overview (status, usage, limits) Plugins (1-click enable + OAuth connect) Packs (toggle workflows on/off) Tech Cloudflare Pages (Next.js/React) Auth: magic link/email optional later; v1 can be token session after Stripe receipt (INFERRED) 7) Abuse Controls (Telegram-Specific) Telegram has $0 messaging cost → abuse shifts to LLM + workflow loops. Mandatory controls (INFERRED) Per-tenant: max messages/day max tokens/day max tool calls/day max concurrent workflows Bot-farm indicators: many new chats/hour repeated prompts high failure rates Enforcement: auto-throttle → auto-suspend → require payment method revalidation 8) Repo / Service Layout (Monorepo) global-claw/ apps/ web/ # Pages: marketing + app shell workers/ api/ # Control plane API tg-webhook/ # Telegram ingest llm-router/ # provider-agnostic router oauth/ # 1-click plugins OAuth handlers durable/ tenant-agent/ # per-tenant DO packages/ plugins/ # plugin manifests + pack templates packs/ # preset workflows shared/ # types, auth utils, crypto, logging db/ migrations/ # D1 SQL infra/ wrangler/ # wrangler configs ci/ # GitHub Actions 9) D1 Schema (Minimal v1) -- tenants CREATE TABLE tenants ( id TEXT PRIMARY KEY, created_at TEXT, plan TEXT, status TEXT, subdomain TEXT, token_budget_daily INTEGER, msg_budget_daily INTEGER ); -- users CREATE TABLE users ( id TEXT PRIMARY KEY, created_at TEXT, email TEXT, stripe_customer_id TEXT ); -- tenant_users CREATE TABLE tenant_users ( tenant_id TEXT, user_id TEXT, role TEXT, PRIMARY KEY (tenant_id, user_id) ); -- subscriptions CREATE TABLE subscriptions ( tenant_id TEXT PRIMARY KEY, stripe_subscription_id TEXT, status TEXT, current_period_end TEXT ); -- plugin_tokens (encrypted refresh tokens) CREATE TABLE plugin_tokens ( tenant_id TEXT, provider TEXT, encrypted_refresh_token TEXT, metadata_json TEXT, created_at TEXT, PRIMARY KEY (tenant_id, provider) ); -- usage_daily CREATE TABLE usage_daily ( tenant_id TEXT, day TEXT, tokens INTEGER, messages INTEGER, tool_calls INTEGER, PRIMARY KEY (tenant_id, day) ); 10) “Latest AI Tools” (Cloudflare-native + practical) Cloudflare Agents SDK is explicitly aimed at production agent interfaces and tool handling (OBSERVED) Durable Objects + Workflows are positioned by Cloudflare as a foundation for agents and multi-step processes (OBSERVED) INFERRED recommendations (tooling): Cloudflare Agents SDK (agent/message patterns) Workflows for provisioning + retries Queues for async executions AI Gateway for centralized LLM governance (optional) Sentry for exception aggregation OpenTelemetry export if/when Cloudflare supports your pipeline cleanly 11) MVP Plan (14 Days, Telegram-Only) Day 1–2: Domain + Pages skeleton + Worker API scaffold Day 3–4: Stripe checkout + webhook → Workflow provisioning Day 5–6: Tenant DO + Telegram webhook routing Day 7–8: LLM router + DeepSeek integration + budgets Day 9–10: Packs v1 (Lead capture + FAQ + simple CRM logging) Day 11–12: Plugin framework scaffolding (manifests + toggles) Day 13: Abuse limits + throttling + audit logging Day 14: Launch + programmatic SEO templates (top 5 locales) 12) Verification Plan (What must be confirmed) UNKNOWN → verify before production Kimi 2.5 API pricing + rate limits + terms (check official docs) MiniMax API pricing + rate limits + terms (check official docs) Telegram per-tenant bot creation automation feasibility without human steps (BotFather constraints) Cloudflare account limits for DO/Workflows at your plan tier (confirm current quotas) OBSERVED (already sourced) DeepSeek token pricing published in DeepSeek API docs Cloudflare Workflows GA + DO position for agents per Cloudflare press release Agents SDK release notes (v0.1.0+, v0.3.0) in Cloudflare changelog Copy-Paste “North Star” Statement Global-Claw is a Cloudflare-native, Telegram-first agent automation SaaS: pay → provision → bot works instantly. Plugins are 1-click connectors. LLM is cheap-first with hard budget caps. Every tenant is isolated via its own Durable Object agent. If you want, I’ll convert this Bible into actual implementation artifacts next: wrangler.toml (Workers + DO + Queues bindings) Workflow definition (provisioning) Telegram webhook Worker Tenant DO skeleton LLM router Worker with DeepSeek wired (using the exact endpoints/pricing assumptions) now write a business plan (must have resellers) must focus on not usa market, webiste should be in different languages GLOBAL-CLAW Business Plan (Non-US, Reseller-Led, Multilingual) 0) Operating thesis Telegram-first AI automation SaaS has structurally better unit economics outside the US than WhatsApp-first (no per-conversation messaging fees, faster onboarding, fewer compliance gates). The fastest global distribution is resellers (agencies, MSPs, local “digital shops”) selling packaged outcomes to SMBs in their language. Product must be localized by default: website, onboarding, pricing, templates, invoices, and in-bot UX. 1) Target markets (explicitly non-US) Primary regions (Telegram adoption + SMB density + payment rails) DACH (Germany/Austria/Switzerland): strong Telegram usage + high SMB spend Southern Europe (Spain/Italy/Portugal): high SMB density, strong messaging culture Eastern Europe (Poland/Romania/Czechia/Baltics): Telegram strong, reseller ecosystems active MENA (UAE, Saudi, Egypt, Morocco): Telegram adoption high, business messaging normal LATAM (selected) (Mexico, Brazil, Colombia, Chile): Telegram meaningful; WhatsApp dominates but Telegram still viable for many verticals South Asia (India): huge SMB base, strong Telegram usage Positioning rule: “Not a chatbot.” It’s instant automation (lead capture, appointment routing, FAQ triage, internal ops). 2) Customer + reseller ICP End customers (B2B only, by design) 1–50 employee SMBs that already live in messaging: clinics, salons, gyms, tutors, real estate teams, restaurants, auto service, home services, ecommerce operators Resellers (must-have channel) Local agencies and consultants who already sell: websites, social ads, SEO, CRMs, “digital transformation” MSP-like operators (IT services) Industry-specific operators (e.g., salon booking consultants) Reseller value prop: sell outcomes without building software; recurring revenue; local language advantage. 3) Product offer (what you sell) Core SKU: “Instant Telegram Automation Suite” Tenant provisioning in seconds Telegram bot + dashboard live immediately 5 preset packs (start with 2–3 enabled by vertical) 1-click “plugins” (OAuth connectors) as upsells Preset packs (v1: Telegram-only) Lead capture + qualification + routing Appointment request intake (calendar later via plugin) FAQ + escalation (to human contact) Review capture / feedback collector Internal reminders / simple ops assistant 4) Pricing (reseller-first, non-US localized) End-customer pricing (example, adapt per country PPP) Starter: €29/mo — 1 bot, 1 pack, basic limits Pro: €79/mo — all packs, higher limits, basic integrations Business: €149/mo — multi-operator, analytics, advanced packs, higher limits Usage add-on: LLM overages / premium model toggle (rare if you keep cheap-first) Reseller pricing + margin (mandatory) Two models; offer both: Model A: Revenue share (simplest) Reseller gets 30% recurring for 24 months Payout monthly Reseller link attribution + last-touch Model B: Wholesale (best for serious agencies) Wholesale: 40–50% margin Agency sets retail price (within floors) Optional onboarding fee: €300–€1,500 (kept by reseller) Rule: No “enterprise negotiation” early. Keep it productized. 5) Website strategy (multilingual, non-US) Domain structure (global-claw.com) global-claw.com/{lang}/... /de/, /es/, /pt-br/, /fr/, /it/, /pl/, /tr/, /ar/ Currency + pricing localized by geo (IP + user selection) Content strategy: programmatic + vertical landing pages For each language: /de/telegram-bot-friseur /es/automatizacion-telegram-inmobiliaria /pt-br/bot-telegram-clinica Each page includes: Localized benefits 2–3 “pack” descriptions ROI calculator (hours saved/week) 60-second setup promise reseller CTA (“Find a Partner”) Translation quality (non-negotiable) Human-verified for top 5 languages first Back-translation checks Lock legal pages per language 6) Go-to-market (100% automated + reseller-led) Stage 1: Reseller acquisition (the engine) Automate partner recruitment: Landing page: “Become a Global-Claw Partner” Self-serve signup → partner dashboard → unique link Automated email drip for onboarding In-product: partner can spawn tenants in 1 click Acquisition channels: Search: “white label chatbot”, “telegram automation”, “agency automation tool” in target languages LinkedIn outreach (localized) Local agency directories YouTube/TikTok shorts in each language (templates, use cases) Stage 2: Reseller enablement Provide: 10 vertical pitch pages per language proposal generator (PDF) demo Telegram bot flows (shareable) case study templates (localized) Stage 3: Customer acquisition loops In-bot “Powered by Global-Claw” badge (toggleable for reseller plans) Referral code for end customers Template marketplace (packs = shareable links) 7) Operations (minimal support, self-serve) In-app guided “Enable Pack” toggles Automated diagnostics: “Bot health” + “Limits” Hard budget caps prevent runaway costs Abuse detection → auto-throttle → auto-suspend Support model: No tickets by default Knowledge base per language Resellers provide first-line support (you support resellers, not SMBs) 8) Compliance + taxes (practical for non-US) B2B-only posture: collect business name + VAT ID where applicable For EU complexity reduction: consider Merchant of Record later (Paddle/Lemon Squeezy) if you go heavy B2C Privacy: GDPR-ready baseline (data export/delete) 9) Milestones + KPIs (90 days) Month 1 5 languages live (DE, ES, PT-BR, FR, IT) Partner portal + tracking 20 reseller signups 50 trials started Month 2 100 resellers signed (long tail) 100 paying tenants total Activation rate > 40% (pack enabled + first automation) Churn < 5% monthly Month 3 300 paying tenants 30 resellers producing ≥3 sales/month CAC approaching near-zero via programmatic SEO + reseller channel 10) Financial model (simple targets) Target blended ARPU (non-US): €49–€89 With 30% reseller payout: Gross margin remains strong if LLM is cheap-first and Telegram-only Break-even target: ~200–400 tenants depending on fixed costs + reseller ramp 11) Risks (and mitigations) Low Telegram adoption in some countries → prioritize Telegram-heavy markets first; add WhatsApp later as premium Resellers selling garbage / abuse → partner scoring, minimum standards, automated throttles, kill switch Localization quality → lock top markets with human QA; measure conversion by locale Payment failures by region → add local methods over time; optional MoR

ID: 0212644d-7134-4805-9de6-68928a7edd55

Status: running

Phase: N/A

Quality Score: N/A

Revenue Potential: N/A

Mode: cloud

Artifacts: 0