`MARKETING.md`

94 lines
# VenturaClaw — launch comms drafts

## ⚠ Pre-flight before posting any of this

Verified 2026-05-07 against the live site:

- `https://venturaclaw.com/` currently renders a **"Coming soon · 67 SaaS tools" email collector** — not a live demo widget. The Show HN body below promises "Try the live demo on the homepage" — that promise is **false against current prod state**.
- The full app (chat, connectors UI, admin) lives at `https://app.venturaclaw.com/` and is **admin-gated** (login required).
- Connector registry confirms **67** tools (`server/connectors.json` → 67 keys). Use 67 consistently; an earlier draft said "56-connector grid" — fixed below.

**Two ways to make this draft postable:**

1. **Ship the demo widget back onto the homepage** (preferred — that's the strongest hook). Embed the same classify endpoint the hidden-door app uses; gate writes behind the existing approval queue; reads (list/get/search) can run anonymously with a per-IP rate limit. Then this draft works as-is.
2. **Rewrite the demo paragraph** to point at a screen recording or a write-up of the architecture instead. Weaker hook (HN expects to *click and try*) but truthful.

Don't post until one of those is done.

## Show HN draft
**Title:** Show HN: VenturaClaw – Type one sentence, AI runs the right SaaS tool

**Body:**
> I built VenturaClaw because every workflow-builder I tried — Zapier, Make, n8n — felt like more work than the work it was supposed to replace. Drag, wire, debug, maintain. The actual job ("refund this order", "post in #launch", "purge the cache") is one sentence; the canvas was a tax.
>
> So this tool sells the outcome, not the canvas. You type what you want; a local LLM (qwen3:14b on a Mac Studio in my office) picks the connector and runs it. 67 SaaS tools pre-wired out of the box. Sensitive actions are gated by an approval queue with the parsed intent shown in plain English. API keys are AES-256-GCM encrypted at rest with a versioned key envelope.
>
> Try the live demo on the homepage — type a command, watch the connector grid light up the matched tile. No login needed for the demo. The classifier runs on my hardware, not OpenAI/Anthropic.
>
> https://venturaclaw.com/
>
> Stack: Express + JSON store (will be PG when user count crosses 5), local Ollama qwen3:14b for routing (zero per-request inference cost), AES-256-GCM at rest, Browserbase for OAuth-app registration sandbox. Single-file HTML pages, no SPA, no SDK to install.
>
> Happy to answer technical questions in the thread.

**Top-of-thread seed comments to drop after posting:**

1. *(Why local LLM)* "Routing runs on a Mac Studio at the edge of my home network — qwen3:14b. Eliminates per-request inference cost, removes the 'is OpenAI training on my business commands' question, and the latency is fine because the prompt is small (~3KB system + 200 tokens user). Hot model serves classify in 13s warm, 30-60s cold (Mac1 has OLLAMA_MAX_LOADED_MODELS=1 so it gets evicted by other workloads). 10-min response cache + sequential preset prime on boot keep the demo widget instant for repeat queries."

2. *(Why not just function-calling on a hosted LLM)* "Two reasons: (1) my business commands aren't training data for someone else's model; (2) per-task pricing on hosted LLMs makes the unit economics ugly when you're routing thousands of cheap reads. Self-hosted means the marginal cost of routing is electricity, not API fees. Trade-off: I can't ship a managed multi-tenant SaaS as easily — every customer would either share my hardware (privacy regression) or run their own (deployment burden). Currently single-tenant by design."

3. *(How sensitive actions work)* "Connectors carry a `sensitive: bool` and per-action risk classification. Reads (list, get, search) execute immediately. Writes (refund, post, dns_edit, delete) land in `approvals.json` with the parsed intent shown in plain English: 'Refund order_id=1234 on shopify · $42.50 · customer steve@example.com'. Admin clicks Approve or Reject. The LLM never decides whether to execute — it only decides what would be executed if approved."

---

## Twitter / X announce (280 chars)

> VenturaClaw is live: type one sentence ("refund order 1234 on shopify"), the local LLM picks the right SaaS tool from 67 wired connectors, runs it. Sensitive actions gated by approval queue. No workflow builder. No third-party LLM. https://venturaclaw.com/

---

## IndieHackers (slightly longer, more product-led)

**Title:** Built an AI agent that replaces Zapier with one sentence — and ships running

**Body:**
Hey IH,

I just shipped VenturaClaw at https://venturaclaw.com/. Quick summary:

**The problem:** Zapier, Make, n8n all sell you the *canvas*. You drag triggers to actions, wire them up, maintain them. The actual job is one English sentence; the canvas is a tax.

**The product:** You type "refund order 1234 on shopify" or "post in #launch on slack saying we shipped". A local LLM (qwen3:14b on my Mac) picks the right one of 67 wired connectors and runs it. Sensitive actions queue for approval. Reads execute immediately.

**The differentiation:**
- Per-user OAuth + API key vault, AES-256-GCM encrypted at rest
- Local LLM means your commands aren't third-party training data
- Sensitive-action approval gate with parsed intent in plain English
- 67 connectors pre-wired — no DIY app builder, no Zaps to maintain
- $0/request inference cost (because Ollama on a Mac, not OpenAI/Anthropic)

**Try it without signing up:** the homepage demo widget calls the live LLM router. Type any command, watch the connector grid light up. No login required for the demo.

**What's next:** registering OAuth apps for Stripe, Slack, Notion, HubSpot, Discord, Mailchimp, Google so users can connect with one click. Today the connector vault accepts pasted API keys. Pricing is free during early access — eventual model is flat workspace fee, no per-task metering.

Looking for feedback on: the demo UX (does it feel responsive enough?), the homepage copy (does it sell the outcome over the canvas?), and the threshold question — would you trust a sensitive-action approval queue, or do you want stricter human-in-the-loop?

---

## Search Console / Bing Webmaster — sitemap submit

After verifying domain ownership in Google Search Console (DNS TXT or HTML meta tag — DNS is cleaner; add at Cloudflare):

1. Search Console → Sitemaps → Add: `https://venturaclaw.com/sitemap.xml`
2. Bing Webmaster Tools → Sitemaps → same URL
3. IndexNow (faster Bing/Yandex indexing): the GET below pings them with the changed URLs

```bash
curl "https://www.bing.com/indexnow?url=https://venturaclaw.com/&key=KEY-FILE-NAME"
```
(IndexNow needs a key file at `/<KEY>.txt` containing the key as proof of ownership — set up later)

## Sitemap auto-ping (already shipped)
- /robots.txt declares Sitemap line ✓
- Nothing more to do server-side; the manual SC + Bing submit is the only step.