Start
Introduction
Mobius is a web service that runs AI agents. It saves their conversations, run records, and produced files; connects them to other software; and gives your team a web app for checking what happened.
It is for teams that want to use agents in their own operations or products without building the server-side system for schedules, conversations, tools, approvals, and run history. If you are building a product, your software can provide the customer interface while the Mobius API runs the agent behind it.
You can start in the Mobius app. You do not need to write code for the first example.
What you can do with an agent
An agent is an AI worker you configure with instructions and a model, such as Claude or GPT. You can also choose which tools it may use.
The selected AI model produces the response. Mobius sends it the request, handles any allowed tool calls, and saves the response. Which models are available depends on how your Mobius organization is configured.
There are two common ways to give an agent work:
- Have a conversation. Send a message, read the reply, and ask a follow-up. Mobius saves that conversation in a session so the agent can use the earlier messages.
- Run a repeatable process. Write down the steps once, then start them with a button, a schedule, or an event from another system. Mobius calls that saved process a loop. Each time it starts, Mobius creates a run that shows what happened that time.
A conversation does not create a run. A loop may use one agent, several agents, or no agent. When a loop uses an agent, Mobius also saves that exchange in a session. Use a conversation when the next message should shape the work. Use a loop when you know the steps in advance. How Mobius works shows the full relationship.
What Mobius records
Open a run to see the information saved with that run, the steps it attempted, visible agent and tool activity, the result of each step, and why the run stopped. Open a session to read the messages in one conversation.
This record does not expose a model's private chain of thought. It shows the inputs, outputs, tool calls, action results, and approvals that Mobius can actually record. That lets you trace the Mobius action that attempted to send an email or diagnose a scheduled task that failed overnight. It does not by itself prove that another service delivered the email.
An interaction is a request for information, review, or approval. Human approval is optional. If work must be approved before it changes another system, put an approval interaction before the step that makes the change. The run waits for the assigned person. Approval lets it continue; rejection or expiry stops it before the next step. Approval cannot protect an earlier tool or step that already made the change.
Files that a run keeps are called artifacts. They remain available after a temporary workspace is removed, until someone deletes them or a retention policy does.
A concrete first result
Suppose Northwind IT needs to turn rough backup notes into a client update. The notes say that a backup failed, was retried, and then completed. The guide shows you how to ask an agent for a short update and check that it did not add a new incident or an unsupported cause.
The quickstart builds that example. You will:
- create an agent named Scout;
- create a loop with one agent step;
- paste the backup notes into that step;
- start the loop by hand; and
- open the run and read Scout's draft.
The example does not connect to a live ticket system or send anything to a client. Mobius still sends the prompt to the model selected for Scout. Once the draft is right, the guides show how to connect real inputs, add a review, and send the approved result.
Where you work
Click Open app in the header to reach the Mobius app. It is where you create agents and loops, start work, read runs and conversations, and answer approval requests.
Two other surfaces are available when you need to connect or automate Mobius:
- The
mobiuscommand-line tool runs commands from a terminal. - The HTTP API lets software you operate talk to Mobius directly. HTTP is the common way software sends requests over a network.
Some advanced settings are available through the command-line tool or API but not the app. Each procedure names the surface it uses.
Connecting other systems
Integrations connect Mobius to accounts such as Slack, Google Workspace, GitHub, Jira, Linear, and HubSpot. Some use an OAuth sign-in, the same kind of connection as Sign in with Google. Others require an API key, which is a password created for software rather than a person.
A connection makes that service's actions available in the project. You can call an action at a fixed loop step or put selected actions in a toolkit for an agent to use. Connecting an account does not give every agent automatic access.
You do not need an integration for the quickstart. Add one when the loop needs to use an account in another service. Software can also start a loop with an HTTP trigger or the Mobius API.
Read next
- Build and inspect your first run.
- See how the main Mobius objects fit together.
- Set time, spending, and retry limits before a loop runs without supervision.
- Triage inbound tickets with a review step.
For a new loop, start with a manual run. Give it known input, check the result, and change one thing. Add a schedule or outside event only after that version works.