Vernix — AI Video Call Agent

AI that joins your video calls, creates embeddings from the meeting content, and acts as a voice or text agent participant with RAG capabilities.

Supported platforms: Zoom, Google Meet, Microsoft Teams, and Cisco Webex.

API Documentation

Tech Stack

• Next.js 16 (App Router, Turbopack, TypeScript)
• Shadcn/ui + Tailwind CSS v4
• Drizzle ORM + PostgreSQL
• Qdrant for vector storage
• OpenAI for embeddings and voice
• Recall.ai for meeting bot integration
• Vercel AI SDK for streaming chat with tool calls
• Minio / S3 for knowledge base file storage

Quick Start

Prerequisites

• Node.js 24+
• pnpm
• Docker

Setup

# 1. Install dependencies
pnpm install

# 2. Copy env file and configure
cp .env.example .env.local

# 3. Start Postgres, Qdrant, and Minio
docker compose up -d

# 4. Push database schema
pnpm db:push

# 5. Start dev server
pnpm dev

Open http://localhost:3000 to see the dashboard.

Environment Variables

See .env.example for all environment variables.

Recall.ai Webhook Configuration

Two webhook endpoints, each serving a different purpose:

1. Realtime Transcripts — /api/webhooks/recall/transcript (automatic)

Configured per-bot in code when creating a bot via POST /api/agent/join. Recall streams live transcript.data events here during the call. No manual setup needed — just make sure NEXT_PUBLIC_APP_URL is set correctly.

2. Bot Status Events — /api/webhooks/recall/status (manual)

Go to Recall Dashboard → Webhooks → Add Endpoint and configure:

• URL: https://your-app.up.railway.app/api/webhooks/recall/status
• Events: bot.call_ended, transcript.done

bot.call_ended sets the meeting to "processing" status. transcript.done triggers summary generation after all transcript data has been delivered.

Scripts

Command	Description
pnpm dev	Start dev server
pnpm build	Push DB schema + production build
pnpm validate	Format, lint, typecheck, and test
pnpm test	Run unit tests
pnpm typecheck	TypeScript type check
pnpm lint	Run ESLint
pnpm format	Format with Prettier
pnpm db:push	Push schema to DB
pnpm db:studio	Open Drizzle Studio

Architecture

• Meetings are stored in PostgreSQL with metadata (title, join link, status, participants, summary)
• Each meeting gets its own Qdrant collection for vector embeddings
• The meeting bot (Recall.ai) joins calls, streams audio, and generates transcripts
• Transcripts are embedded with OpenAI's text-embedding-3-small and stored in Qdrant
• After a meeting ends, an LLM summary is generated from all transcript segments
• The RAG agent searches current and past meeting transcripts to answer questions
• The knowledge base lets users upload PDF, DOCX, TXT, and MD files — parsed, chunked, embedded, and searchable alongside meeting transcripts via unified RAG. Documents can be scoped to specific meetings for context-aware boosting
• Meeting agenda field embedded into Qdrant for RAG, injected into agent prompts and summary generation
• Action items are auto-extracted from transcripts via LLM and stored as discrete tasks — viewable per-meeting and across all meetings on the dashboard
• MCP Server exposes meeting data (search, transcripts, tasks) to Claude Desktop, Cursor, and other MCP clients via API key auth
• MCP Client connects to user-configured external MCP servers, routing their tools to the chat and voice agents
• Streaming chat on meeting detail and dashboard pages with tool-call visualization and source citations
• The dashboard provides meeting notes, transcript timeline, search, and filtering
• Silent Mode — a per-meeting toggle that changes the agent from a voice participant to a passive text listener. The bot joins without audio or screen presence, monitors the live transcript server-side, and responds via the meeting's built-in chat when addressed by name (Vernix). Enable it in the "New Meeting" dialog or on the meeting detail page before the agent joins. Responses are rate-limited to 1 per 30 seconds to prevent chat spam.

Agent Readiness

Vernix advertises its API surface to AI agents through several standard discovery conventions. These are all served by the Next.js app — no separate infra, no build-time steps, no extra cron jobs.

• Link headers (RFC 8288) on / and /docs advertise the api-catalog, OpenAPI spec, MCP endpoint, and llms.txt. See next.config.ts.
• /robots.txt emits Content-Signal: ai-train=yes, search=yes, ai-input=yes alongside the standard allow/disallow rules.
• /.well-known/api-catalog — RFC 9727 linkset pointing at the OpenAPI spec, docs, llms.txt, and the MCP server card.
• /.well-known/mcp/server-card.json — MCP Server Card (SEP-1649 draft) describing the server at /api/mcp, its transport, and auth model.
• /.well-known/agent-skills/index.json — Agent Skills Discovery v0.2.0 index with a live sha256 of /llms.txt.
• /llms.txt — hand-written markdown guide for LLMs covering pricing, API, MCP tools, and lifecycle. Regenerated from source data on every request.

Markdown for Agents

Public pages respond to Accept: text/markdown with a markdown version of the page. Convention: the browser gets HTML, agents asking for markdown get markdown.

curl -H 'Accept: text/markdown' https://vernix.app/pricing
curl -H 'Accept: text/markdown' https://vernix.app/docs   # rendered from OpenAPI, not HTML

How it works: src/middleware.ts intercepts requests carrying Accept: text/markdown and rewrites them to /api/agent-md?path=<original-path>. That route fetches the HTML version from our own origin, strips nav/footer/scripts, and converts to markdown via node-html-markdown. Results are cached in-process for 1h per path (200-entry LRU cap).

/docs is special-cased: it renders markdown directly from the OpenAPI spec via src/lib/api/openapi-md.ts because Scalar's client-rendered UI doesn't survive HTML-to-markdown conversion.

Why on-demand instead of pre-built files:

• Works on any Next.js deployment (Railway, Vercel, self-host) with zero extra steps — no build hooks, no cron, no writable filesystem required.
• First request per path per instance pays ~100-500ms (fetch + convert); subsequent requests are cache hits (~1ms).
• Cache is per-instance. When Railway scales to multiple instances, each one warms its cache independently as agents arrive. No coordination needed.
• Content stays fresh automatically (1h TTL), so changes to marketing pages show up in markdown within an hour without redeploying.

If you ever want to pre-warm the cache (e.g. before a press launch): loop through the sitemap and hit /api/agent-md?path=<each> from any machine. But for normal operation you don't need to — agents warm their own paths.

Deferred agent-readiness features

See TODO.md for auth/payments items that need product decisions before shipping: OAuth / OIDC discovery, OAuth protected-resource metadata, Web Bot Auth JWKS, WebMCP, x402, UCP, ACP.

License

Vernix is licensed under the Elastic License 2.0 (ELv2). You can use, copy, distribute, and modify the software, with the exception of offering it as a hosted or managed service. See the LICENSE file for full terms.