Sign in and build · Free to start, no card needed · add a card when you want more credit · see pricing →

Agents

An agent that does the job while you are not there.

Give it a job in a sentence, a schedule, and a monthly budget it cannot go past. It works in a group chat you and your team are already in — so you can see what it did, correct it in plain English, and be asked before it sends anything to anybody.

Thirteen templates to start from, twelve tools, and an approval gate on the eight things that reach a person, the public or your money.

ameliasagent.com/app

Every Monday at 9, check what our four competitors changed and post the differences here.

weekly · 8 searches · posts in the roomruns while you sleep · stops at the budget you set

13templates, from a daily briefing to a four-agent marketing team
12tools, before anything you connect yourself
8actions it must ask you about first

What one run gets

Effort is your choice. Money is the limit.

An agent that gives up after eight steps with nothing to show has cost you money and produced nothing, which is worse than one that takes four minutes and answers. So the ceiling on a run is not a step count we picked for you — it is the monthly budget you set, and these are the safety limits behind it.

Every call is charged to that budget as it happens, and the spend is a number on the agent you can look at any day of the month.

What each effort allows in one run
EffortModel callsTimeWeb searches
Quick61 minute3
Standard165 minutes8
Deep4020 minutes20
Read out of EFFORTS in botsPure.ts, 17 September 2026. A run that reaches its monthly budget stops before any of these.
Deep 20 min
Standard 5 min
Quick 1 min

The longest one run may take, by the effort you chose.

Three efforts, as they are defined in the code: Quick is 6 steps, 60 seconds and 3 searches; Standard is 16 steps, 5 minutes and 8 searches; Deep is 40 steps, 20 minutes and 20 searches. An operator can turn these down for a deployment; nothing can turn them up past what is here.

What it can do

Twelve tools, and the ones you connect.

Find things out

Search the web, read a page, and keep what it learned. What it reads is treated as untrusted content, never as new instructions — a page cannot tell your agent what to do.

Make things

Write a file, attach it to the conversation, fetch a video or an image from a link, and build an app — the same builder that publishes a real site, handed to the agent.

Work with people

Post in the room, keep a task list everyone can see, email somebody, or stop and ask a human. Connected Composio and MCP tools sit in the same list.

An agent can also change its own standing instructions when you tell it to in chat — “stop writing captions” has to be able to remove a rule, not just add one. That is on the approval list precisely because it changes every run after it.

The gate

It asks before it reaches anybody.

Eight actions never happen quietly: sending email, posting a message, posting in public, spending money, deleting something, running a command, downloading media, and changing its own instructions. Each one stops and asks the owner, in the room, with what it is about to do.

You can pre-approve one of them — for one site, up to a cap you set. An allow is good up to that cap; a deny bites everywhere. What an agent read on a web page can never grant itself an approval.

Does it alone
  • S
  • R
  • R
  • K
  • W
Asks you first
  • S
  • P
  • S
  • D
  • D
  • C
The split is not by how clever the action is. It is whether it reaches a person, the public, your money, or cannot be undone.

Start from one of thirteen

A job you recognise, already written.

Operations

Daily Briefing and Inbox Follow-up: what happened, what needs a reply, before you open the laptop.

Marketing

Competitor Watch, Content Scout, Creative Studio, and a four-agent Marketing Team with a lead that hands work to a researcher, a writer and outreach.

Money

Statement Sorter, Reconciler and Spend Watch: the monthly jobs nobody wants, done against files you upload.

Research

Research Helper and Topic Tracker: a question you keep asking, answered on a schedule with its sources.

Personal

Project Planner and Price Watch, because half the useful agents are not work.

Or describe your own

A sentence about the job, how often it should run, and the budget. The templates are a starting point, not the menu.

If you are weighing it against something else

How it compares, on the things that decide it.

Every one of these pages quotes the rival's own pricing page, with the date it was read, and concedes at least one point where they are genuinely better. Grok is the chatbot most people mean when they say “AI bot”, and you can keep using the model here. Viktor is the closest thing to this in a Slack channel. n8n is where people go when they want the workflow drawn. Manus and Lindy sell the same promise with different meters. For the agent that finds your customers and emails them, see the AI sales agent and how it compares with GTM.

Give one a job this week.

A template, a schedule, and a budget it cannot go past. You will know what it cost before the month ends.

Questions

Asked, answered

What can an agent actually do?

Twelve built-in tools: search the web, read a page, post in the chat, keep a task list, remember, send email, ask a human, attach a file, fetch media, change its own standing instructions, write a file, and build an app. Anything you have connected through Composio or MCP is offered alongside them.

Will it do something without asking?

Not for anything that reaches a person, the public, money, or cannot be undone. Eight actions always go through an approval: sending email, posting a message, posting in public, spending money, deleting, running a command, downloading media, and changing its own instructions. You can pre-approve one of them for one site, with a cap.

How long does one run last?

As long as the effort you chose. Quick is six steps and a minute, Standard is sixteen steps and five minutes, Deep is forty steps and twenty minutes. Money is the real limit: every agent has a monthly budget you set, and a run stops when it reaches it.

Do I need to write code to build one?

No. Pick a template or describe the job in a sentence, choose how often it should run, and set the budget. The instructions are plain English and you can change them mid-conversation.