How it works

From an API
to an agent people use.

Connect the services a job depends on, build an agent around them, and decide who can reach it and where. Each step happens in the same app, and every run can be opened afterwards.

The Site Weather Brief agent's setup: work mode and the model that does the work.
An agent's setup in the staging demo workspace.

One agent, start to finish

Watch a weather brief come together.

In under a minute: connect a public forecast service from its API description, create the agent, pick its model and tool, configure an answer check, ask it a question and open the run behind the answer.

This is a single recorded run on public data. The run page in the video counts that one run only, so its figures say nothing about how the agent does over time.

Recorded on staging on 18 November 2026 using public weather data. This agent has an optional second-model answer check configured in report-only mode.

Three steps

Connect, build, share.

The same three steps apply whether the agent serves your team, your customers or your own application.

Connect

Point almyty at a service you use.

Add the link to a service's API description, or upload the file. almyty reads OpenAPI, GraphQL, SOAP and gRPC descriptions, and npm packages, and lists what the service can do as tools, such as "get the forecast" or "search issues". You choose which of them an agent may use.

For your own systems you need that description and a key if the service asks for one. Someone who knows the service, or your IT team, can usually provide both.

Six connected APIs in the staging demo workspace, among them weather, exchange rates and GitHub, with how many tools each one has.
Connected APIs in the staging demo workspace.

Build

Describe the job and choose the models.

Write what the agent should do in plain words and pick its tools. An autonomous agent works out the steps itself. When the steps must always be the same, lay them out as a workflow instead.

Choose a model from any provider you have connected, including a model you run yourself. You can add an optional answer check: a second model reads the answer against what the tools returned and reports what it finds.

An agent's model settings: one model writes the answer and a model from another provider checks it.
Model and answer check settings in the staging demo workspace.

Share

Decide where people reach it.

Publish a chat page with its own address, or add a chat bubble to a website. Other supported channels include Slack, Microsoft Teams, WhatsApp, email, SMS and iMessage, each with the account setup that platform requires. You can also package a desktop or terminal app.

Your applications can call the same agent through an OpenAI-compatible API. You set the name, colours and greeting, and choose whether anyone with the link can use it or only people who sign in.

The Dependency Check agent's web chat channel, live at its own address.
A published web chat in the staging demo workspace.

Part of every agent

More than a prompt and a model.

These come with each agent. Turn on the ones its job needs.

Files and pictures
People can attach a photo, a PDF or a document to a conversation, and the agent can read it.
Memory
Keep useful context between runs and conversations, and choose how widely the agent remembers it.
Schedules
Not every job starts with a question. Run an agent at a set time, such as 8:00 every weekday, and keep each result in its runs.
Runs you can open
Every run keeps its steps, the tool calls it made, what came back and what it cost.

Runners

When the job needs a particular machine.

Some work has to happen on a specific computer: a build server, an office machine, a laptop. Start a runner there and agents can send it work. You decide who may use it, which folders it may touch and what it may never run.

Read the runner guide
Starting a runner: install the command line tool, log in and start it.
Runner setup in the staging demo workspace.

Try it on one job you know well.

Connect one service, build one agent and check its first result against what you would have done yourself.