Relevance AI Tutorial: Set Up Your First Agent, Step by Step

Setting up a first Relevance AI agent runs in five stages: a paid plan, a starting point, a reviewed draft, knowledge and a trigger, then a supervised test. Every organization now needs a paid plan, and Relevance AI's pricing pages describe no free trial, so the way in is Pro at $19 a month billed annually or $29 a month billed monthly, for the whole organization.^[1] You then pick one of three starts: Invent, a Marketplace clone, or a blank agent. Our advice is short. Start from a template or Invent, keep approvals on, and publish only after you have read real runs.

Before the first agent: the plan, the desk, the admin switch

The builder at app.relevanceai.com works only in a desktop browser: Chrome, Firefox, Safari, Edge, Arc, Dia or Comet. Phones are for chatting. The iOS and Android apps are Chat only, so they can talk to finished agents but cannot build or edit them. Pro suits a solo builder or a pair, since it includes 2 Build Users, no End Users and one project, with 2,500 Actions and 10,000 Vendor Credits a month. Team, at $234 a month billed annually or $349 a month billed monthly, includes 5 Build Users and 45 End Users, which matters once colleagues need to use what you build.

Two meters start ticking during setup, not after launch. An Action is one Tool run, however many steps sit inside it, and a failed run still counts. Vendor Credits cover model and tool usage at wholesale, with no markup. By our arithmetic, Pro billed annually comes to $228 a year, against $348 for twelve months billed monthly, so annual billing costs less from the 8th month on. Plan Actions reset at each renewal, and the pricing guide covers top-ups and spend controls.

One switch belongs to the admin. Invent, the built-in AI builder, is on by default for every member in projects created since November 2026. Older projects keep their existing setting, and an organization admin can turn Invent on or off project by project. Check this before the first build session, so nobody opens a project and finds Invent switched off.

Three ways to start, and which one fits your job

The documentation names three starting points. You can describe the process to Invent in plain language, clone a ready agent from the Marketplace, or build from scratch with a prompt, tools and optional knowledge. Engineers can also work from coding agents such as Claude Code, Codex or Cursor through the MCP server at mcp.relevanceai.com. Pick by how unusual the job is. A common job like inbox triage or calendar booking usually has a listing worth cloning. A process specific to your company suits Invent. A builder who already knows the prompt and tools should start blank.

Starting blank is less bare than it sounds. Invent can still write a structured prompt from a few words, even inside the manual path. For the model, you can leave the choice to Relevance AI, tuned for cost or for performance, or pick one yourself from OpenAI, Anthropic Claude, Azure OpenAI, Google Gemini or OpenRouter. Each tool input can be left to the agent, set manually, or filled from an earlier tool's output. Our agents guide covers memory, escalations and the rest of an agent's anatomy.

Walking through Invent: describe, review the draft, publish

Describe the business process the way you would brief a new hire: what arrives, what has to happen, which apps are involved, where a human must sign off. Invent then builds the Agents, Tools, Triggers and Workforces the process needs, and can create Knowledge tables too. Everything lands in draft. Its tool permissions come in three settings, Auto, Ask and Always ask, and destructive actions such as delete or publish default to Always ask.^[2] Restoring an older version always needs your confirmation. Invent can only do what your own role allows. Nothing goes live until a person publishes it.

Two habits keep this safe. First, run Invent inside a sandbox project. It is scoped to the active project, so a sandbox gives it no reach into production, even though it can use that project's own connections to apps such as Gmail. Second, expect to open the Tool Builder afterward, because the docs say tools Invent creates may still need customizing there. Invent sends your conversation to the selected model provider, and its web search stays off unless you turn it on.

Invent costs real credits. The docs themselves warn that it is expensive to run and consumes significant credits, and a counter shows what the current session has used. Your plan sets its default model. Invent can also set up Evals and pull analytics, but those depend on your plan: agent evaluations are Enterprise only, and the analytics dashboard starts on Team. On Pro, the review step is you reading the draft and its test runs.

Cloning a Marketplace template without surprises

The Marketplace groups agents, tools and Workforces under Sales, Marketing, Content Creation, Research, Operations, and HR and talent. Relevance AI staff and community builders publish them. Free listings are marked Free and can be cloned again at any time. Paid listings are priced by their creators, come on top of your plan, are tied to the project where you bought them, and can be fully modified after purchase. On a listing, the Clone Agent button copies the agent into your project. You then connect the channels it needs, such as email, social or web analytics.

Clones arrive half-empty. Cloning does not copy knowledge tables, triggers, global snippets, secrets, or API key and OAuth values; triggers come only as suggestions.^[3] Budget time to reconnect accounts, rebuild the trigger and load your own data. The docs walk through a few clonable examples: a Gmail Assistant for inbox work, a HubSpot Assistant for contacts and deals, a Notion Assistant, and a Calendar Bookings Scheduler that proposes 30-minute slots across two days and skips weekends. One boundary matters here. The named agents on Relevance AI's use-case pages are sales-led setups tuned by its experts, not templates you clone. Our integrations guide explains how each app connects.

Adding knowledge and a trigger

Knowledge can start as an empty table, uploaded files (CSV, PDF, Excel, JSON, audio), a website crawl, or a synced source such as Google Drive, SharePoint or Notion. The attach mode is the real decision. Add all to prompt includes the whole set on every run, which suits a short policy sheet or price list. Allow agent to search turns it into retrieval through a tool, which suits a large help center or document archive. Retrieval offers semantic, hybrid and metadata-filtered search. One agent can use several knowledge bases, and one base can serve many agents.

A trigger is what turns a draft into a worker. Scheduled triggers run daily, weekly, monthly or on custom dates, and advanced users can write six-field AWS EventBridge cron, which includes a year field. Integration triggers fire on app events, from a list of 1,000+ app triggers. A webhook lets any system that sends HTTP requests start the agent. Map a Unique ID there, and a repeated event with the same ID is ignored. The API trigger, the JavaScript SDK and tools used as triggers round out the options. One agent can hold several triggers and pause them one at a time or all at once. Premium triggers for WhatsApp, LinkedIn and Telegram work on every paid plan, while Salesforce, Snowflake and Zendesk triggers need Enterprise.

Test with approvals on, then loosen the reins

Every tool on an agent has its own approval mode: Auto Run, Approval Required, or Let Agent Decide. Set anything that sends or writes to Approval Required for the first runs. The Outlook guide in the docs follows the same order, with approval mode first and autopilot later. Pending items land on the Tasks page, whose tabs split Approvals, Escalated, Errors and All. From there you can approve or reject in bulk and rerun an errored task from the failing step. Remember that agents decide their own path, so read a few complete runs, not just the outputs.

Guardrails also cap cost. The Agent Autonomy Limit stops an agent after a set number of consecutive tool runs without human approval, retries included; the docs' example of Tool A run three times and Tool B seven times counts as 10. A failure policy can force retries, then end the task or escalate it to a person. Each of those tool runs, retries included, is an Action, and tests spend the same 2,500 monthly Actions on Pro as live work. Watch the counter.

Every draft save and every publish creates a version, and restoring one puts it back into draft rather than live. Deleting an agent is permanent, and support cannot recover it. Sharing needs the same care. Turning on any embed option or the cloneable template makes the agent public through a URL search engines can index.^[4] Anyone with that link can run it on your credits. Since May 2026 an admin can switch off public cloning and embedding for a project, which is worth doing before anyone shares a test agent. If the agent's steps need deeper changes, our tools guide covers the Tool Builder.

What a first setup should leave you with

A finished first setup is modest on purpose: one agent, in its own project, with a reconnected trigger, the right knowledge mode, approvals on every tool that writes, and a version you can roll back to. Only then widen the autonomy limit or switch a tool to Auto Run. Keep an eye on Actions while you do, since a busy test week spends the same meter as live work. For structured learning beyond the docs, Relevance AI Academy offers structured courses that walk you through building on the platform step by step. Once one agent runs cleanly, connecting several into a Workforce is the natural next step.