AI capabilities
The AI stack behind every VegaDūta agent
VegaDūta is an agentic-AI platform: you describe a job in one sentence, and an agent takes it on with real tools — across WhatsApp, voice, and ten other channels, in 22 languages, with guardrails that make it ask before doing anything risky.
These pages explain each layer of that stack — agents, Model Context Protocol, LLM connectors, retrieval, guardrails, memory, voice, and on-device models — as it is actually built.
Model Context Protocol (MCP)
Connect any MCP server to your agents, and expose your agents as MCP tools to Claude or your IDE.
AI Agents
Agents with ~35 built-in tools, deployed across 10+ channels, with guardrails that ask you before risky actions.
LLM Connectors
Bring your own key for 13 LLM providers, with live model catalogs, per-user budgets, and automatic fallback.
Knowledge & RAG
Upload documents, get grounded answers: extraction, OCR, local embeddings, reranking, and transcription in one pipeline.
AI Guardrails
Detection for PII, toxicity, and jailbreaks, a configurable rule engine, and human approval gates on sensitive actions.
Graph Memory
Agents remember across conversations: entities and facts extracted into a knowledge graph they can recall as a tool.
Voice AI
Agents that answer the phone: Whisper and Sarvam transcription, ElevenLabs and Sarvam voices, Twilio calls, LiveKit rooms.
Edge AI
Run LLMs in the browser via WebLLM with a model marketplace, and pair desktop or mobile devices as companion agents.
Try it before reading further
The sandbox provisions a real tenant — describe an agent in one sentence and test it, no account, no card.
