Skip to main content
SSetodDocsOpen app →

Start here

An agent is a set of written instructions, a few accounts it is allowed to use, and a schedule that decides when it wakes up. Everything else on this platform exists to serve those three things.

How it fits together

You connect an account once - a Gmail inbox, a Telegram bot, a phone number. That account then becomes available to every agent in your workspace. When you build an agent, you pick which of those accounts it may use, write instructions telling it what to do, and choose how often it should run. When it runs, an AI model reads your instructions, looks at whatever the accounts show it, and takes actions on your behalf.

The five words that matter

  • Connector - one connected account. Shared across the whole workspace, set up once.
  • Tool - one thing an agent can do, like send_email or read_unread_emails. Attaching a connector grants that account's tools. A code skill is a tool you write yourself.
  • Agent - instructions plus connectors plus a trigger.
  • Trigger - what wakes the agent up: a schedule, an inbound message (Telegram, WhatsApp, Instagram, SMS, webhook), a phone call, another agent, or you pressing the button yourself.
  • Run - one execution, recorded in full so you can read back exactly what the agent thought and did.

Getting your first agent working

  1. Pick a brain. On Pro and Business, agents run on Setod-managed models out of the box, paid from your plan's AI credit. On Free (or if you prefer your own account) add an OpenAI or Anthropic API key on the Connectors page.
  2. Connect something for it to act on. Gmail, a Telegram bot, Twilio, or an MCP server. An agent with a brain but no accounts can think and say things, but cannot do anything.
  3. Create an agent. The Templates page has ready-made agents for common jobs, grouped by what they are for - pick one, read its instructions, and press Use template. Everything is editable before it is created. Or start from scratch on the Agents page and write what you want it to do in plain English.
  4. Test it. Press Test run and read the transcript line by line.
  5. Publish it. Until you publish, the agent never runs on its own.

A test run is a real run

Pressing Test run does not simulate anything. It uses your connected accounts and genuinely sends whatever the agent decides to send. Write cautious instructions before the first test, not after.

Two things worth knowing early

  • Editing a live agent is safe. Your edits go to a draft. What runs on the schedule keeps being the last version you published, until you publish again.
  • Agents do not re-do work. An agent polling your inbox every 30 minutes sees the same unread mail every time, but the platform tracks what it has already handled so it does not reply twice.

Code you can publish

A prompt skill tells an agent how to behave. A code skill is a Python function you write, publish, and attach. The agent then calls it the same way it calls send_email. The function defines main(input, context). Network access stays off until you turn it on, and secret values are never shown again after you save them. Code skills are on Pro and Business plans. The limits are on the Skills page.

Coding Agents

Claude Code and Cursor can build and diagnose agents in a workspace once you create a personal access token. An MCP connector is the other direction: it lets one of your agents call a server such as GitHub or Slack. Coding Agents.

NextConnecting accounts →