Overview

A customer at Acme Diagnostics opens the chat on your website at 11pm and asks, "Do I need to fast before a lipid test?" No one from your team is online — but they get a clear answer in seconds, from your own words. That's the kind of agent you build here.

This page is for anyone who's about to build their first agent. By the end, you'll know the five pieces every recipe uses, how they fit together, and which recipe to start with. Each recipe starts from a goal, lists what you need, and walks the config to a working, testable agent.

Every recipe reuses the same five building blocks. Once you see where each one plugs in, the recipes below read as variations on one theme — the agent is the same canvas throughout.

The building blocks each recipe uses

Block	What it is	Where it's configured
Agent	An AI Agent node with Personality + AI Model sub-nodes, plus optional Knowledge, Integration (MCP) and Action nodes such as senders.	Agent Builder
Knowledge Base	Folder-scoped retrieval (RAG) over content you upload or crawl.	Knowledge Base
MCP tools	Your business logic and data, exposed as tools and called by the agent.	MCP servers
Channels	Web, WhatsApp, SMS, Email, Voice — every inbound message reaches your agent the same way.	Channels
BYOP connections	The accounts your agent talks through (Plivo, your email provider, storage, databases) are yours. AI models are included.	Connections

How the blocks behave

Four mechanics underpin every recipe on this page.

• One inbound path. Every channel — Web, WhatsApp, SMS, Email — reaches your agent through a single pipeline: identify the customer, match a trigger, then run the matched agent. A message no published trigger watches gets no reply — there's no catch-all default agent. If a channel happens to deliver the same message twice, your agent still answers only once. On email, SMS and WhatsApp the agent replies through a sender node on its Action port; web chat and voice reply inline.
• Folder-scoped knowledge. Knowledge Base search is scoped to the folders you wire onto the agent's Knowledge sub-node, not all of your content. Perfox always includes exactly the folders you selected — even folders that contain no directly uploaded files but pull content from a connected source — so nothing you attached is ever silently dropped. See FAQ Agent and Customer Service from your Website.
• Tools with guardrails. Every MCP tool call runs through a consistent chain of protections — the agent discovers the tool, its inputs are validated, tool-access policy and per-tool rate limits are enforced, results are cached, your credentials are injected, the tool is invoked, and the call is written to the audit log. You can set a rate limit per tool; once a server's limit is exceeded, further calls are held back until the window resets. See Connect an MCP Server.
• Bring your own provider. Each provider account (Plivo, your email provider, Google Sheets, your databases) is the one you connected under Connections — your keys stay yours, and your data stays private to your workspace. If a recipe needs a provider you haven't connected yet, Perfox tells you exactly which connection to fill in.

A recipe end to end

Facts on their own can feel abstract, so here's one recipe playing out from start to finish. Follow Asha, an admin at Acme Diagnostics, as she builds an FAQ agent, and Priya, a customer who uses it that same night.

Setup. Asha opens the Agent Builder and drops in an AI Agent node. She fills the Personality sub-node — a warm, plain-spoken clinic assistant — and adds the AI Model sub-node. The AI model is included and managed by Perfox, so she leaves it on Platform default. She uploads her clinic's prep instructions and price list into a Knowledge Base folder, then wires that one folder onto the agent's Knowledge sub-node. Because search is folder-scoped, the agent will answer from those documents and nothing else.

Action. Asha activates the agent and pastes the Web Chat embed snippet onto the Acme Diagnostics site. That evening, Priya lands on the page and types into the chat:

Priya: Do I need to fast before a lipid test?

Result. The message follows the one inbound path — Perfox identifies Priya, matches the trigger for the site, and runs Asha's agent. The agent searches only the folder Asha attached, finds the prep note, and replies in the chat window:

Agent: Yes — please fast for 9 to 12 hours before a lipid test.
Water is fine. Morning slots work best if you'd like to eat soon after.

What just happened. Priya got an accurate answer from Acme's own documents, at 11pm, with no one on the team online. Asha didn't write that answer — she uploaded the source and wired one folder. Every mechanic above did its job quietly: one inbound path routed the message, folder-scoped knowledge kept the answer grounded, and the whole thing ran without Asha writing a line of code. Swap the Knowledge Base for a website crawl, add an MCP tool, or turn on more channels, and you have the recipes below.

Pick a recipe

If you want to…	Recipe	Key pieces
Answer questions from your own docs	FAQ Agent	Knowledge Base (upload) + Web Chat
Answer from your live website	Customer Service from your Website	Website connector (crawl)
Use your own business tools & data	Connect an MCP Server (HR Assistant)	Integrations + MCP tools
Produce & send documents	Generate a Report & Email It	Built-in capabilities + Email Sender
Serve a customer's own data securely	Authenticated Customer-Data Agent	Auth-context broker
Qualify visitors & escalate	Lead Capture → Demo / Human Handoff	Information to Collect + hand off to a human
Be everywhere at once	One Agent, Every Channel	Multi-channel triggers + continuity
Reach out proactively	Outbound Notifications	Schedule / Webhook + sender nodes

Most production agents are a combination. A single agent can be FAQ- or website-grounded, also call your MCP tools, run across every channel, and open up per-customer data once the visitor is authenticated. The embeddable Web Chat widget — a self-contained bundle with history, fresh-session, Markdown export, and fullscreen — is the front door for several of these recipes, and the same headless conversation engine powers custom front-ends too.

Where to go next

You can now name the five building blocks, follow one message down the single inbound path, and choose a recipe to build first.

• FAQ Agent — the shortest path to a working agent; start here if you have docs to answer from.
• Connect an MCP Server (HR Assistant) — when the agent needs your own tools and data, not just documents.
• Knowledge Base & RAG — how folder-scoped retrieval works under the hood.
• Channels Overview — the one inbound path, and how each channel reaches your agent.
• Bring Your Own Provider — connect Plivo, your email provider and the rest so messages go out on your accounts.