Agentic AI Platform
Read the code, fork it, run it on your own servers. Nothing is locked away.
Say what you need in plain English. VegaDūta builds the agent, picks its tools and model, and puts it on WhatsApp, voice, email and six more channels. Then it shows its work: replay any run step by step, test a change before customers see it, and approve the risky steps yourself. It builds the phone app around the agent too.
Feed it a workload:
About a minute · no card · your dashboard opens immediatelyor try it first — no signup
Simulated build log: the typed spec compiles into tools, a model pick, guardrails and channels.
Your conversations and your data are protected at every layer of VegaDūta's agentic AI platform.
Provider keys & channel secrets are AES-encrypted in the database — never stored in plain text.
Sign-in runs on Keycloak (OAuth2 / OIDC with PKCE). We never see or store your password.
Every workspace's data lives in its own isolated schema — no cross-tenant leakage.
Full data export and erasure on request — you stay in control of your information.
For your business, your studies or your side project — staff that speaks your customer's language, finishes the job, and asks you first on the big calls.
Open-source agentic AI · BYOK or free edge models · hosted pool for experiments
About a minute · no card · your dashboard opens immediately
What it does
No slides and no demo data: each of these is a shipped capability you can switch on in the free sandbox right now.
Connect your number by QR. It replies from your price list and documents, in the customer's language, and hands over to you when it should.
You type “Reply to every WhatsApp message with prices from my list, and pass complaints to me.”
It answers the phone in the caller's language, books the slot or takes the order — and every call is transcribed and logged.
You type “Answer calls to the clinic, book appointments, and keep the day's list up to date.”
Lay out a process — on a schedule, a webhook or an event — with branches, approvals and parallel steps. It runs while you sleep.
You type “For every new order: check stock, and if it is over ₹5,000 ask me before sending the invoice.”
Point it at your repository. It maps callers and callees, works in a sandboxed workspace, and shows you the diff before anything merges.
You type “Add rate limiting to the payments API and open a pull request with tests.”
About a minute · no card · the agent you build in the sandbox comes with you.
Just shipped · October 2026
Building an agent takes a sentence. Knowing it will behave takes evidence. Everything below is live on vegaduta.ai today.
Agent Lab draws any past run as a graph of every model call and tool call. Re-run it from the step you choose, remove parts of the prompt to see which ones the answer depends on, and repeat an input to check that it answers the same way twice.
Find it: Agent workbench → Lab
Fourteen built-in skills, from customer support and SQL to Word, Excel and PDF files, each a plain SKILL.md folder. An agent reads the full instructions only when a task calls for them. It can also draft new skills from its own successful runs, and those stay off until an admin approves them.
Find it: Skills · Agent workbench → Skills
Give an agent two or more specialist colleagues and it splits a job between them at the same time, then combines the answers. Delegation goes one level deep, each branch has a time limit, and the whole team shares one run budget.
Find it: Agent workbench → Team / Swarm
Pin a step's output while you build, so the steps after it run on the same data. Expressions pass values between steps, draft runs stay separate from production, and long jobs run in the background while you do something else.
Find it: Workflows → canvas
Score an agent against your own test set with six automated judges, with confidence intervals so one lucky run doesn't fool you. Set a pass bar, and a version that falls short is not promoted.
Find it: Agents → Quality
One page lists what your agents have learned about you. Edit what is wrong, delete what you don't want kept, and write the notes you want every conversation to start with.
Find it: Build → Memory
Status comes from the deploy log, not a brochure. Not claimed here, because it is not true yet: outbound calling, agents collecting UPI payments into your own account, and store listings for the mobile apps.
Your path - three stops
Try it with nothing. Understand what it did. Then bring it home - the account takes a minute, and the agent you built comes with you.
Try
Open a private workspace with a working agent. Talk to it, hand it a task, watch it finish. Nothing is asked of you - not even an email.
Understand
Every reply shows its steps: the plan, the tool it picked, the check before it acted. Academy turns that into lessons, from your first agent to expert.
Implement
We ask for the minimum. Your trial agent is copied into your account, and the dashboard unlocks - create agents, build workflows, connect channels, all on your own.
Everything we ask for
That is it. No card, no company documents, no sales call.
The dashboard unlocks: create agents, build workflows, connect Telegram and Email free - 5 agents on the free plan, hosted model included.
Not sure where to start? Start at 01. Nothing is asked of you there.
Everything inside
Thirty-six things your free account opens on day one — each one shipped and counted in the code that runs this page. Tap any tile for the detail, and where it lives in the console.
Live todayOpt-inBetaIn design36 of 36
Senses
Send a photo, a scanned bill or a screenshot in any chat and it reads what is in it — up to three files, 10 MB in total per message, on every model that supports images.
Where it lives: Console → Playground → attach; any channel that carries images.
Every feature is unlocked for every account — permanently. Opt-in items are switches, not extra charges.
Where this sits
Same idea: you describe an outcome, and a graph of steps, tools and models carries it out. The difference is where the work happens. Those tools are workflow-first. Here the agent is the product and workflows are a supporting capability inside it — so it arrives already speaking on WhatsApp, Telegram and voice, in Indian languages, on a multi-tenant backend.
Better there
Integration breadth — thousands of integrations against VegaDūta's dozens.
Better here
Your product is an agent that answers customers on WhatsApp, Telegram, email or a phone call, in the language they write in.
Better there
RAG UX polish — Dify's knowledge-base experience is more mature.
Better here
Your agents must meet customers on WhatsApp, Telegram or a phone call, in Indian languages.
Better there
Conversation-design UX depth — years of chatbot-specific product polish.
Better here
You are an Indian SMB or agency and roughly ₹63,000 a month at ₹84 to the dollar for a team plan is not a serious option.
Every row on those pages carries its own citation, including the ones where we lose.
Why VegaDūta
ChatGPT, Claude Desktop, Hermes, OpenClaw — all good tools. The difference isn't intelligence, it's jurisdiction: an assistant helps you inside its window; an agentic platform works for you, on your customers' channels, under your rules.
Assistants help you inside their window. They won't answer your customer on WhatsApp, wait three days for your approval, or keep a workflow running while you sleep.
Then you know the catch: you are the platform. Hosting, model keys, guardrails, approvals, budgets — and wiring real channels — is a DIY project. VegaDūta is the same idea productised: going open source too, with the platform parts built in.
What you actually control here
Every inbox, one agent
WhatsApp, Messenger, Instagram, Slack, Discord, X, Telegram, Email and Voice. 9 is the whole list — every channel the platform routes, not a selection — and three of them are open the minute you create a free account.
What it takes to switch each one on
The 9 channels a free account can run today, with no upgrade and no plan to buy. WhatsApp needs your own number — pair it by QR, or connect Meta Cloud API or Twilio.
WhatsApp
Messenger
Instagram
Slack
Discord
X
Telegram
Email
VoiceCounted from the platform’s own channel list rather than a brochure: 9 channel types the message router accepts. The 0 bring-your-own-account channels carry no plan restriction of their own — they need credentials from the service you are connecting. The marks above are our own drawings in each brand’s colour, used to identify the service; no partnership or endorsement is implied.
Use cases
Every card is a real brief you can run verbatim — “Run this spec” drops the sentence into the live studio on this page, no signup. The recipe link opens the full step-by-step in the workflow-recipes guide.
“Answer customer questions on WhatsApp and email from our own docs, and hand off to a human when unsure.”
“When an alert webhook fires, dedupe it, write the incident summary and page the on-call — after my confirm.”
“Review every pull request against the call graph, post the findings, and hold the merge for my approval.”
“Chase each unpaid invoice on WhatsApp with a UPI link — disputes come to me, receipts go out on their own.”
“Take one research question, fan it into sub-questions, search in parallel and return a brief with sources.”
“Answer every call, book the slot on the calendar and text the confirmation. Spoken-language coverage depends on the speech provider you connect.”
Browse the full catalogue — 89 industries of workflow recipes
Not just chat
An agentic workflow is a durable run: it branches, fans out to specialist agents, calls your systems, waits for a person or a reply for as long as that takes, then carries on. Pick a process and watch how it is built.
Hundreds of supplier invoices a week: read, matched to purchase orders, exceptions chased, payment released only after sign-off.
Runs for days per invoiceTo run it on your systems: Email inbox, your ERP or accounting API. You connect those after you sign in; nothing here asks for a credential.
Live studio — no signup
Describe the outcome — the planner parses your intent, searches the live MCP registry for tools, picks a model with a failover chain, and builds a real agent around it. Run it here, then ship it as a public one-tap link.
Free agentic-AI case study
Tell us how one job actually gets done in your business today, in your own words. We write back with the agent we would build for it: the steps it would take, the tools it would touch, and exactly where it would stop and ask you. No charge, no card, no sales call — and the dashboard you would build it in is free too.
Case file — your process, your agent
Pick the one closest to your business
What you tell us
Customers message us all day asking price, stock and delivery. My two staff answer the same twenty questions, copy orders into a notebook, and we still lose a few every week.
One paragraph like this is enough. You do not need to know anything about AI to write it.
The agent we would draft
Picks up the message the moment it lands and replies in the customer's own language.
Telegram + Email — free planAnswers price, stock and delivery from your own price list and notes — not from guesswork.
Your own knowledge baseWrites the order up and hands it to you to confirm before anything is promised.
Human approval gateEvery reply and every action is written down, so you can read back exactly what it did.
Audit logAn illustrative sketch, assembled only from things the platform already does today — channels, your own knowledge base, workflows, the approval gate and the audit log. No numbers, no promises: your real draft is written from your process, not from this one.
How the free case study works
You email us one process. A paragraph is plenty — no form, no NDA, no discovery call.
We write back with the agent we would build for it: its steps, the tools it needs, and the points where it should stop and ask a human.
You build it yourself in your free workspace — Academy teaches you how — or you ask us to sit with you while you do.
And the dashboard itself is free
Not a trial. Sign up and a real workspace switches on: build agents, wire up workflows, connect Telegram and Email, and learn the whole craft in Academy at no cost.
Prefer to just start? Nothing is stopping you. Create the free account, follow Academy from your first agent to your first workflow, and take the business a level up on your own — we are here if you want a second pair of eyes.
Questions, answered straight
Yes — free, and open source. Every feature is unlocked permanently: unlimited agents, workflows and seats, WhatsApp, Telegram and Email, white-label branding. Three ways to run models: bring your own keys, run on-device edge models at no cost, or use the platform pool on fair-use limits so your first agent answers without an API key.
No. Every plan includes the platform's pooled hosted model, so your first agent works with zero setup. On paid plans you can bring your own key for any supported provider — including Sarvam for Indian languages — and your agents switch to it.
Yes, on every plan including Free — you connect your own number by QR, or via Twilio or Meta's WhatsApp Cloud API, and the agent answers there directly. We do not charge for the channel; the number and any Meta conversation fees stay yours. The conversations playing on this page are clearly-labelled simulations of that experience.
You decide how much rope it gets. Set an agent's autonomy tier so consequential actions wait in your approval queue instead of just happening, guardrails and content moderation are built in, and every action lands in an audit log you can review any time.
Talk to a human
No question is too simple — even "how does agentic AI actually work?" gets a real answer from a person. Tell us what your business does and what you want an agent to handle. We will tell you plainly whether the platform can do it — including when the answer is no.
Will it work for my business?
Describe your process. We will say which parts an agent can take today, and which it cannot.
Bigger than the free plan?
More seats, more agents, or a workload you would rather talk through before committing to it.
Security and data questions
Where your data sits, what the audit log records, and how to keep a model running on your own device.
Help us keep it free
We are a very small group of developers building VegaDūta on our own bootstrap savings — and we have made everything free for everyone. If it is useful to you, even the price of a coffee genuinely helps us keep going.
Write to us at [email protected]
No paywall, no nag screens, no feature held back behind this. It is a thank-you, not a toll.