Mastra Run VC Minimal Agent

What’s included (final minimal scope)

• Prebuilt run.vc website Q&A: prebuilt-runvc-qa (answers strictly from indexed site chunks)
• Portfolio company Q&A: portfolio-static (static JSON list)
• FAQ Sheet Q&A: faq-sheets-qa (Google Sheet tab)
• Minimal pitch intake (CSV append only): pitch-intake

Getting started

• Prereq: Node 20+.
• Option A — Mastra playground (agents): npm run dev (unchanged; for reference only).
• Option B — Website + CometChat widget page: npm run web then open http://localhost:8080.

Website + CometChat widget

• File: public/embed.html — Minimal page that:
  • Iframes any site (default https://run.vc, override with ?u=<url>)
  • Mounts the CometChat Chat Embed widget for chat
• The page loads: https://cdn.jsdelivr.net/npm/@cometchat/chat-embed@latest/dist/main.js
• Configure inside public/embed.html:
  • COMETCHAT_CREDENTIALS (appID, appRegion, authKey)
  • COMETCHAT_LAUNCH_OPTIONS (targetElementID, isDocked, width/height, chatType, defaultChatID, variantID)
  • Provide the user UID via query param: ?uid=<yourUserUid> or edit COMETCHAT_USER_UID placeholder
• Preview: http://localhost:8080?u=https://run.vc&uid=<yourUserUid>

Embed elsewhere

• You can embed public/embed.html itself in another site via iframe, or host it directly.
• Note: Some sites disallow iframing via X-Frame-Options or CSP frame-ancestors. If that happens, consider a proxy renderer approach; I can add a simple proxy mode if needed.

Notes

• This setup does not have the agent scraping; the page just displays the site and your CometChat widget handles Q&A.

FAQs from Google Sheets

• Set the env var in .env:
  • RUNVC_FAQ_SHEET_URL=<your Google Sheet link> (include gid= for the tab to use)
• Make the sheet accessible: Share with “Anyone with the link can view” or Publish to the web.
• The agent calls the faq-sheets-qa tool first for every question and uses its answer if a close match is found.
• Cache TTL: set RUNVC_FAQ_TTL_MINUTES (default 10). Set to 1 for near real-time updates as you edit the sheet.

Static portfolio data

• The agent can answer portfolio questions from a static file: data/runvc_portfolio.json.
• Format: { "companies": [ { "name": "Company", "website": "https://...", "logo": "(optional)", "description": "(optional)" } ] }
• Edit this file any time; no rebuild required. The portfolio-static tool reads it at runtime.
• If running via mastra dev where the working dir is .mastra/output, the tool looks in:
  • .mastra/output/data/runvc_portfolio.json, then ../data/runvc_portfolio.json (repo root)
  • Or set RUNVC_PORTFOLIO_PATH=/absolute/path/to/runvc_portfolio.json to override.

Answering behavior

• Portfolio questions: use portfolio static data (list or details). If not found, reply "Not found on run.vc. Please check the Portfolio page." (as configured in agent instructions).
• FAQ-style questions: try FAQ sheet; if low match or missing, fallback to website index.
• General run.vc questions: website index.
• Pitch intent: trigger minimal pitch intake flow (see below).

Pre-crawl and index (website data)

• Crawl run.vc and build a local vector index:
  • npm run crawl:runvc — fetches pages into data/runvc_pages.json
  • npm run index:runvc — builds embeddings index into data/runvc_index.json
• The agent tool prebuilt-runvc-qa uses this index to answer questions strictly from run.vc content.
• You can re-run these commands whenever the site changes.
• Auto-run on agent start: npm run dev and npm start will run src/mastra/scripts/ensure-index.mjs first.
  • It builds the index if missing or older than RUNVC_INDEX_TTL_HOURS (default 24h).
  • Override: set RUNVC_INDEX_TTL_HOURS=0 to force rebuild every run, or increase to reduce frequency.

Minimal pitch intake (strict) The agent collects at most 3 short items ONLY when the user clearly expresses intent to pitch:

1. Startup name (if not already given)
2. One-line description
3. Contact email Optional: website ONLY if explicitly provided already; not a separate question. After required fields present it appends a row to data/pitch_intakes.csv with: id,timestamp,startupName,oneLiner,contactEmail,website(optional) and replies exactly: Thanks! We will be in touch. If Google Sheets access is configured (GOOGLE_SERVICE_ACCOUNT_KEY, RUNVC_PITCH_SHEET_ID, optional RUNVC_PITCH_TAB_NAME), each submission is also appended to that sheet. No pitch generation, no extended questionnaire, no emailing.

Removed / deprecated (intentionally not present anymore)

• Email tooling & outbound pitch emails
• Rich multi-question startup submission (submit-startup)
• Middleware prototypes (question limiting / description capture)
• JSON submission mirroring (CSV only now)

Environment variables in use

• OPENAI_API_KEY - required for embeddings & model
• RUNVC_FAQ_SHEET_URL - public/shareable Google Sheet link
• RUNVC_FAQ_TTL_MINUTES - optional cache TTL (default 10)
• GOOGLE_SERVICE_ACCOUNT_KEY - base64 Google service account JSON for Sheets append (optional)
• RUNVC_PITCH_SHEET_ID - Sheet ID to append minimal pitch rows into (optional)
• RUNVC_PITCH_TAB_NAME - Tab name for append range; defaults to Sheet1 when unset (optional) Optional / can remove if present: legacy email vars (EMAIL_PROVIDER, SENDGRID_API_KEY, EMAIL_FROM, RUNVC_PITCH_TO).

Regenerating site index Commands remain the same: npm run crawl:runvc -> updates data/runvc_pages.json npm run index:runvc -> updates data/runvc_index.json Index tool looks for data/runvc_index.json. Override path with RUNVC_INDEX_PATH.

Development npm install npm run dev (agent playground) or npm run web (embed demo page)

Data files summary

• data/runvc_index.json (website chunks + embeddings)
• data/runvc_pages.json (raw crawled pages)
• data/runvc_portfolio.json (portfolio companies)
• data/pitch_intakes.csv (minimal pitch submissions)

License: Internal / experimental.