Create your first agent

This tutorial walks through building an auxilia agent end to end — from naming it to chatting with it. We’ll build a concrete example: a Sales Research Assistant that looks up CRM records and drafts outreach. The same steps apply to any agent you want to build.

You don’t need to be a developer to follow along, but an admin may need to register MCP servers for you first (see MCP Servers).

What you’ll build

An agent that can:

• Look up contacts, companies, and deals in HubSpot
• Research a prospect’s company on the web
• Draft a short, personalized outreach email for a human to review

By the end you’ll understand the four things that define every agent: identity, tools, permissions, and instructions.

Before you start

• You need the editor or admin role. The person who creates an agent becomes its owner.
• At least one MCP server should be registered in your workspace. For this example we’ll use HubSpot (see the HubSpot setup guide) and a web-search server.

Step 1 — Create the agent and give it an identity

Navigate to Agents in the sidebar and click Create Agent.

Fill in the identity fields:

• Name — what users see in the agent list. Keep it short and descriptive: Sales Research Assistant.
• Emoji — pick one from the emoji picker. It becomes the agent’s avatar in the sidebar and chat header. A 🔭 or 💼 works well here.
• Avatar background color — choose a hex color for the badge behind the emoji. Pick a color that helps users tell this agent apart from others at a glance — give each team’s agents a consistent color, for example.
• Description — one sentence shown on the agent card. Users pick agents by skimming this, so make it concrete: “Researches prospects and drafts outreach using HubSpot and the web.”

The name, emoji, color, and description are all editable later — none of them affect how the agent behaves, only how people find and recognize it.

Save the agent. You’ll land on its configuration page, where the real work happens.

Step 2 — Add MCP servers

Tools are what turn a chatbot into an agent. They come from MCP servers bound to the agent.

On the configuration page, click Add MCP Server and pick a server from the workspace catalog. For our example, add two:

• HubSpot — for CRM lookups
• A web-search server — for company research

When you add a server, auxilia fetches its tool list and attaches it to the agent.

Connect OAuth servers once

Some MCP servers — HubSpot, Notion, Linear, Google services, and most other account-based integrations — require OAuth. Registering the server at the workspace level is not enough: each user must connect their own account once before the agent can use those tools on their behalf.

If you see a Connect prompt on the server card:

1. Click Connect
2. You’re redirected to the provider’s consent screen
3. Approve the requested permissions
4. You’re redirected back to auxilia, and the card shows as connected

Your tokens are stored privately and scoped to you — a teammate using the same agent connects with their own account. API-key servers (and bearer-token servers like n8n) skip this step because they share one workspace credential. See MCP Server Authentication for the full picture.

Step 3 — Select and configure tools

Just because a server exposes 30 tools doesn’t mean your agent should have all 30. A focused toolset makes the agent faster, cheaper, and more predictable.

Expand each bound server to see its tools. Every tool has one of three states:

For the Sales Research Assistant, a sensible setup:

• HubSpot read tools (get contact, search deals, get company) → Always allow
• HubSpot write tools (create note, update contact) → Needs approval — we want a human to see exactly what gets written to the CRM
• HubSpot tools the agent doesn’t need (bulk delete, admin settings) → Disabled
• Web-search tool → Always allow

When you first bind a server, every tool defaults to Always allow. Always review the list and tighten writes before sharing the agent — this is the single most important safety step. See Tools & Permissions for the full model.

Tool settings live on the agent–server binding, so the same HubSpot server can be read-only on this agent and fully writable on an automation agent. You’re not configuring HubSpot globally — only what this agent can do with it.

Step 4 — Write the instructions

The instructions are the system prompt sent on every message. This is where you turn a generic model with tools into your assistant. Good instructions cover four things: who the agent is, which tools to use when, how to format output, and what to do when something fails.

Here’s a complete example for our agent:

You are a Sales Research Assistant for the revenue team. Your job is to research
prospects and draft outreach — not to send anything yourself.

Tools:
- Use the HubSpot tools to look up contacts, companies, and deals. Always start
  from HubSpot before searching the web, so you don't contradict the CRM.
- Use web search to enrich what HubSpot doesn't have: recent company news,
  funding, headcount, product launches.

When researching a prospect, gather: the contact's role, the company's industry
and size, the current deal stage (if any), and one recent, specific hook from
the web (a launch, a hire, a press mention).

When drafting outreach:
- Keep it under 120 words, friendly and specific. Reference the hook you found.
- Never invent facts. If you can't verify something, leave it out.
- Present the draft for review. Do not create or send anything until a human
  approves the write.

Formatting: always cite the HubSpot contact ID you used. Format deal values with
a currency symbol.

If a HubSpot record can't be found, say so plainly and ask for the correct name
or ID — don't guess.

Notice what this does:

• Sets boundaries — “research and draft, don’t send” keeps the agent in a safe lane, reinforced by the needs approval setting on writes.
• Orders the tools — HubSpot first, web second, so the agent doesn’t drift.
• Defines done — exactly what a good research result contains.
• Handles failure — what to do when a lookup comes up empty, so the model doesn’t hallucinate.

Save your instructions. Changes take effect on the next message; existing chat threads aren’t recomputed. See Configuration for more prompt-writing tips.

Step 5 — Test the agent in Chat

Open a chat with the agent (the Chat button on the agent, or pick it from the Agents list) and try a realistic request:

Research Maria Gomez at Northwind Logistics and draft a follow-up email.

Watch what happens:

• Tool calls appear inline as expandable blocks showing the tool name, the arguments the agent passed, and the result it got back. This is how you verify the agent is reaching for the right tools with the right inputs.
• Approval cards appear for any needs approval tool. When the agent tries to write a note back to HubSpot, you’ll get an Approve / Reject card showing the exact payload — approve it and the run continues, reject it and the agent adapts.
• Pick a model for the thread from the model selector. Different models trade off cost, speed, and reasoning depth.

Iterate

Testing in Chat is where you tune the agent. Common adjustments:

• The agent used a tool you didn’t expect → set it to needs approval or disabled in Step 3.
• The agent guessed instead of asking → tighten the failure-handling line in the instructions.
• The drafts are too long or off-tone → add a concrete formatting rule and re-test.
• The agent ignored HubSpot and went straight to the web → make the tool ordering explicit in the prompt.

Each change takes effect on your next message, so the loop is fast: chat, observe the tool calls, adjust, chat again.

You’re done

You now have a working agent. From here you can:

• Share it with your workspace — grant teammates user, editor, or admin access
• Trigger it over HTTP — call it from scripts, webhooks, or spreadsheets
• Trigger it from Slack — let the team talk to it in threads
• Run it on a schedule — daily digests and recurring reports
• Add subagents or the sandbox for more advanced agents