VegaDūta Help & Docs

VegaDūta is an agentic-AI platform: you create AI agents that plan, use tools, work across messaging channels, and ask a human before high-stakes actions. You can start entirely free — the Free plan needs no card and runs on a platform-hosted model.

Your first ten minutes

1. 1Create your account at app.vegaduta.xyz (email, Google, or GitHub sign-in). A short onboarding introduces the console the first time you sign in.
2. 2Open Agents and create your first agent — give it a name and a system prompt describing its job.
3. 3Talk to it right away in the built-in chat to check its behaviour.
4. 4Ground it in your own material: upload documents under Knowledge and attach them to the agent.
5. 5When it behaves the way you want, connect a channel (WhatsApp, Telegram) so it can talk to the outside world.

Where things live

• Build — Agents, Workflows, and Quick Task: the things you author.
• Resources — Knowledge, Tools, Code Intel, and Channels: what your agents draw on and reach through.
• Plans & Help — your plan's entitlements, billing (admins), and this guide.
• Administration — tenant-admin-only operations like channel management, billing, and the audit tail.

The home dashboard shows what your workspace is actually doing right now. Every figure on it is real telemetry — nothing is illustrative.

• Telemetry tiles — tokens per minute, cost today (and per conversation), LLM calls over 24 h, and process uptime.
• Pending approvals — the single most actionable item: actions your agents are waiting on a human to approve, with a one-click path to review them.
• Latency and model routing — p50/p95/p99 response times and which models handled the last 24 hours of traffic.
• Channels live — a dot per connected channel showing its connection state.
• Live event tail — the same real audit stream the admin console reads, so you can watch actions as they happen.

An agent is a configured AI worker: a model, a system prompt that defines its job, plus the knowledge, tools, and channels you attach to it. Where a chatbot only replies, an agent plans steps, calls tools, and carries a task to completion.

Creating and shaping an agent

• System prompt — the most important control. Describe the agent's role, tone, boundaries, and what it should refuse or escalate.
• Model — Free-plan agents run on the platform-hosted model. Paid plans unlock choosing the provider and model per agent, including bringing your own provider API keys (BYOK); keys are stored encrypted.
• Knowledge — attach document collections so the agent answers from your material instead of general memory.
• Tools — give the agent capabilities beyond text: connected MCP tools let it call real systems.
• Channels — decide where the agent talks: in-console chat, WhatsApp, Telegram, and more.

Staying in control

Guardrails keep agents inside the boundaries you set, and the approvals system pauses high-stakes actions until a human confirms them — pending approvals surface on the dashboard and in the admin console. Every action lands in the audit trail.

Quick Task is the fastest path from "I need this done" to a result: describe a one-off job and run it immediately, without creating and configuring a persistent agent first. Use it for ad-hoc work — a summary, a draft, an extraction — and graduate to a full agent when the same job keeps coming back.

Workflows chain steps into a repeatable automation: route an incoming message, call an agent, branch on the result, notify a human. Runs are tracked with status, and recent runs appear on the dashboard.

Learn by example — industry recipes

The console ships a large library of industry workflow recipes — real, buildable automations for dozens of sectors (retail, healthcare, finance, logistics, creators, and many more), each mapped to the platform's workflow nodes with suggested prompts, agents, tools, and guardrails. It is the fastest way to see what a production workflow looks like before building your own.

Upload your documents — PDFs, office files, text — and the platform's knowledge pipeline extracts, chunks, and embeds them so agents can search and cite your material when answering. This is retrieval-augmented generation (RAG): the agent answers from what you know, not just what the model remembers.

How to use it well

• Organize uploads into collections per topic or per agent — attach only what an agent needs, which keeps retrieval sharp.
• Prefer source documents over exports of exports; cleaner input means better answers.
• Knowledge capacity varies by plan — see the plan comparison for your tier's allowance.

VegaDūta speaks MCP (Model Context Protocol) in both directions. Outbound, you connect external MCP servers as tools your agents can call — turning them from talkers into doers. Inbound, the MCP gateway exposes platform capabilities to MCP-compatible clients.

• Browse the public MCP registry from the Tools screen and connect servers to your workspace.
• Tool credentials are stored encrypted and redacted from logs.
• Connected tools become available to agents you grant them to — capability is opt-in per agent, not global.

Code Intelligence indexes source repositories so you — and your agents — can ask questions about the code itself: where something is implemented, how a subsystem fits together, what a change touches. It powers engineering workflows like onboarding to an unfamiliar codebase or grounding an agent in your team's actual implementation.

Agents aren't limited to the console. Connect messaging channels and your agent answers incoming messages there directly — WhatsApp-first, which matters if your users live on it.

• WhatsApp — pair by scanning a QR code from the Channels screen; the session stays connected with automatic keep-alive, so you don't re-pair after every idle period.
• Telegram — bridge a bot to your agents.
• Channel management lives in the admin console's Channels screen; connection state is visible there and on the dashboard.
• Channel availability varies by plan — check the plan comparison for what your tier includes.

Three tiers: Free ($0), Premium ($19/mo · ₹399/mo), and Unlimited ($79/mo · ₹1,599/mo). India gets ₹ pricing with UPI and card payments via Razorpay.

• Free — build and run real agents on the platform-hosted model, no card required. Model/provider selection is locked on this tier.
• Premium — unlocks model choice and bring-your-own-key providers, plus higher limits.
• Unlimited — the highest limits for teams running agents in production.
• The Plans screen in the console shows your tenant's live entitlements — what your current plan actually includes right now; the public pricing page has the full comparison table.

Billing management (invoices, upgrades) is a tenant-admin surface inside the console.

• Sign-in is handled by a dedicated identity service (Keycloak) with Google and GitHub social login; idle sessions are logged out automatically.
• Each workspace (tenant) is isolated — your agents, knowledge, and keys are yours alone.
• Provider API keys and channel secrets are stored encrypted, and credentials are redacted from tool traffic.
• Guardrails constrain what agents may do; the approvals system routes high-stakes actions to a human before they execute.
• An audit trail records platform actions — the same stream you can watch live on the dashboard.

The Academy at academy.vegaduta.xyz is the platform's free learning portal: a structured, hands-on curriculum that takes you from "what is an agent?" through building, grounding, connecting, and evaluating real agents — including a code-intelligence hands-on and a platform-architecture deep dive.

Lessons ladder through tiers — Builder, Operator, Engineer, Architect — so you always know what to learn next. Everything you learn is practised on the real platform: you learn agentic AI by shipping an agent.

"I can't change my agent's model."

Model and provider selection is a paid-plan feature — Free-tier agents run on the platform-hosted model. The Plans screen shows what your tier unlocks and how to upgrade.

"WhatsApp pairing isn't available on this deployment."

QR pairing has to be enabled by the deployment's administrator. If you see this message, the pairing secret hasn't been configured on your deployment — ask your admin.

"Knowledge upload or search is temporarily disabled."

The knowledge service is briefly unavailable; the console gates uploads rather than accepting files it can't process. Try again shortly.

"We'll be back" screen

The console can't reach the core service — usually a restart or maintenance window. The page recovers on its own once the service is back.

"I was signed out."

Idle sessions are ended automatically for security. Sign back in and you're where you left off.

Still stuck?

The in-console Guide (Plans & Help → Guide) covers every screen in more depth, and the Academy's hands-on lessons walk through the same flows step by step.