Usage Guide

You don't need an API key to start: a hosted model is already included, and Free-tier agents run on the platform's shared pool automatically. If your plan allows your own provider, go to Connect → Providers, pick Anthropic, OpenAI, or another supported provider, and paste in your own API key from that provider's dashboard. The key is validated with a real (non-billed) call before it's saved, and encrypted at rest - once it shows ACTIVE, your agents use it instead of the hosted pool.

After that, the natural build order is: create an Agent (Build → Agents), try it privately in the Playground (Test → Playground), then either wire it into a Workflow for multi-step automation, or connect it directly to a channel (WhatsApp, Telegram, email) so real people can talk to it.

Any field asking for a system prompt or description - an agent's system prompt, a knowledge collection's description, and more over time - has a small ✨ Ask AI button next to its label. Click it, describe what you want in a sentence (e.g. "a support agent that handles refunds and order status"), and it drafts the text using whichever LLM provider you've already activated. Nothing is applied automatically - you always review the suggestion and click Use this yourself, or Regenerate for another try.

An agent is one AI persona: a system prompt, an LLM provider/model, and optionally a knowledge base and a restricted set of tools. Start from a template (support triage, research assistant, etc.) to pre-fill a reasonable system prompt, or start from scratch and use Ask AI to draft one.

Tool access and knowledge access are both allow-lists: leave them empty for "unrestricted" (the agent can use anything currently available to your tenant), or check specific items to restrict the agent to only those. This matters for anything customer-facing - you generally don't want a public support bot to have access to every internal tool.

Upload .md, PDF, Word, images (OCR), or audio/video (transcribed) files and they're chunked and embedded into a searchable vector store. Collections are folders you can nest and scope agents to - e.g. a "Returns Policy" collection only certain agents can search, versus a "Public FAQ" collection every agent can use.

An agent only searches its knowledge base if "Attach knowledge base" is checked when creating it (or toggled on later from the Agents page), and only within whatever collections its knowledge-access allow-list permits.

The visual canvas for multi-step automation: drag nodes (Trigger, Agent, Knowledge Search, Condition, Output, Device Control, and more) and connect them. Every node calls the real service behind it - a Workflow run genuinely calls your agent, genuinely searches your knowledge base, genuinely sends the WhatsApp message.

Workflows can start three ways: manually (the "Run" button), on a schedule (cron expression, e.g. "every morning at 8am"), or via a webhook URL you can call from any external system. Every run is logged step-by-step under that workflow's Runs tab, including each individual tool call an Agent node made along the way.

A marketplace of Model Context Protocol servers and built-in tools an agent can call mid-conversation (web search, a database, the filesystem, your own custom scripts...). Enable the ones you want available, then scope which agents can actually use them from each agent's tool-access editor. Disabling a server here immediately removes it from every agent's available tools, even ones that previously had it checked.

Under Admin → Channels: connect WhatsApp either through Twilio (fast sandbox setup, small per-message fee) or directly through Meta's own WhatsApp Business Platform (no middleman fee, more setup - your own Meta App, System User token, and App Secret). If both are configured, the direct Meta connection is used automatically.

The first time a new phone number messages in, it shows up under Pending sender pairings and needs to be approved before your agents will hold a live conversation with it - this stops a stranger messaging your number from getting an unattended AI conversation by default.

Pair a phone (Android or iOS) to your tenant from the Devices page, and a workflow's Device Control node can tap, type, open apps on, or screenshot it. Android supports full automation via Android's Accessibility Service; iOS is limited by Apple's own sandboxing to opening apps only - see the Mobile SDK's CONSTRAINTS.md for the full, honest breakdown of what each platform can and can't do.

Users: invite teammates and assign roles (only a tenant-admin can change roles or credentials). Approvals is the human-in-the-loop inbox - any action your guardrail policy flags as needing sign-off (e.g. an agent about to send a WhatsApp message) waits here until approved or rejected. Audit Log is a complete, append-only record of every significant action taken in your tenant, by a person or an agent.

"No active providers yet" when creating an agent - activate one under Connect → Providers first; this is the single most common first blocker.

Ask AI button gives an error - same root cause as above: it needs at least one active LLM provider.

A pending WhatsApp/Telegram sender never gets a reply - check Admin → Channels for a pairing waiting on your approval.

A tool doesn't show up for an agent - confirm it's enabled under MCP Tools, and that the agent's tool-access list either includes it or is left empty (unrestricted).