THE DOCS.
FROM ZERO TO YOUR FIRST AGENT.

01 / GETTING STARTED

Three steps from nothing to a running agent. No cloud account, no deployment pipeline, no per-seat math.

$ babuba deploy --license LIC.TEAM-XXXX-XXXX
gateway booting .......... OK
channels: whatsapp · telegram · gmail · web CONNECTED
agents: make your own — ∞ available

1. Get the license — pay once at pricing.html, receive your key instantly. UPI, bank transfer or card.
2. Install the gateway — one command on your server or desktop. The installer boots the gateway, the admin panel and the channel adapters.
3. Activate & build — redeem your key at redeem.html, then create your first agent. It boots with an identity (SOUL), long-term memory (MEMORY), tools and channels wired — in about two seconds.

02 / ARCHITECTURE

Babuba is one gateway process — one port (18793) — hosting everything:

babuba gateway (127.0.0.1:18793)
├─ router       — reads every message, picks the agent + workflow
├─ agents       — one per department; own memory, tools, rules
├─ channels     — whatsapp · telegram · gmail · web chat
├─ workflows    — department processes as approval graphs
├─ cron         — scheduled jobs, reports, scans
├─ admin panel  — agents, approvals, guardrails, RBAC, logs
└─ state        — SQLite + memory, all on your machine

The router decides which agent handles an incoming message — by channel binding, role scope or default. Workflows can start automatically when a message matches a process graph.

03 / CONCEPTS

Gateway

The single process that runs the whole system on your hardware. Install once — the router, agents, channels, workflows and admin panel all live inside it.

Agents

One per department. Each agent is an LLM equipped with its own SOUL.md (identity and system prompt), MEMORY.md (long-term memory), tools, guardrails and approval rules. Agents are isolated from each other — finance can't see HR's data.

Router

A LangGraph-based message router. It classifies intent, picks the right agent and can start a workflow. You don't wire rules — the router learns the routing from your agent set.

Channels

Talk to agents where you already work: WhatsApp Web, Telegram, Gmail, Slack and web chat. Same agent, same memory, every channel.

Workflows

Department processes expressed as graphs — with human approval steps. A payment workflow, an onboarding workflow, a compliance review: each step can require a person's tap before it proceeds.

Cron

Scheduled jobs on the gateway: daily backups, state cleanup, self-diagnosis, market scans, reports. They run whether or not anyone is chatting.

04 / AGENT LIFECYCLE

# create — boots an agent with identity, memory, tools, channels
$ babuba agents create --name finance --dept accounts
  generating SOUL.md ........ done (0.4s)
  agent online .............. RUNNING (2.1s)

# add users — the whole team, role-scoped, no seating costs
$ babuba agents add-user --agent finance --user asha --role user

# edit at runtime — swap model, toggle tools, re-scope users
$ babuba agents update --agent finance --model deepseek-v4-flash

No redeploys. Agents are runtime-editable: change the SOUL, swap the model, toggle tools, re-scope users — the gateway picks it up live.

05 / GOVERNANCE

Babuba ships governance built in — not bolted on. Four layers:

• Approval gates — payments, supplier adds and other sensitive actions pause at your phone. One tap approves; nothing ships silently.
• RBAC — every agent serves the whole team, role-scoped (super-admin → admin → manager → user → viewer). Finance can't see HR's data; ops can't approve payments.
• Guardrails — per-agent allow / deny / warn rules and risk thresholds. The agent knows its boundaries before it acts.
• Audit trail — every LLM call, tool call and decision is logged and searchable. Reversible, reviewable, regulator-ready.

06 / DEPLOY PATHS

• Desktop app — install and run on a Windows/Mac machine. Best for small teams; the setup wizard does everything.
• Server / Docker — one command on any Linux box or VPS. Best for production, always-on availability and team access.
• Air-gapped — fully offline operation is supported; bring your own model (Ollama, local endpoints) and the gateway runs with zero external calls.

07 / FAQ

Does my data leave my server?

No. The gateway, agents, database and memory run on your machine. The only outbound calls are the LLM API calls the agents make — and with a local model, even those stay on your hardware.

How many agents and users?

Unlimited. One license covers as many agents as you want to create and as many users as you want to add. No per-agent or per-seat fees — ever.

Which models can I use?

Bring your own: OpenAI, Anthropic, DeepSeek, any OpenAI-compatible endpoint, or local models via Ollama. Swap models per agent at runtime.

What if I need help?

support@babuba.ai — activation help, install walkthroughs and priority support for life. Get the license ▸