Agents
An agent is the AI worker you define. Give it instructions, choose a model, and connect the tools, skills, and memory it needs.
ProductYou build the agent.
You build the agent.
Mobius runs the backend.
Define what your agent should do and which tools it can use. Run it from the dashboard, CLI, or API, or have it start on a schedule or event. Mobius keeps the work going through failures and human approvals, with budgets and model choice under your control.
Key concepts
The building blocks behind a reliable agent.
Start with an agent: instructions, a model, and tools. Approve a routine when the same request should repeat. Then choose how the work reaches the agent and where its code and models run.
Build
Routines
Ask an agent to repeat something on a schedule, approve the timing and the spending ceilings, and each occurrence comes back to the same conversation. Only a person can approve one.
Connect
Channels
Give an agent a Slack or Telegram channel and its own email address. A message reaches it wherever the work already happens, and the reply comes back on the same thread.
Custom tools (MCP)
Connect an MCP server to give your agents tools from another system. They appear alongside Mobius integrations, so the agent can use them the same way.
Run
CLI
Start and inspect work from your terminal or CI. Every Mobius API operation is available as a command, and the CLI handles sign-in for you.
$ mobius sessions streamAPI
Create and manage agents, sessions, routines, and actions from your application. Typed Python, TypeScript, and Go clients are generated from the OpenAPI spec.
POST /v1/sessions/{session_id}/turnsRun an agent with one call
Send one message to an agent. Stream the reply on the same request, or get the session and turn IDs first and connect afterward. Retrying the same message resumes the existing turn instead of creating a duplicate.
POST /v1/agents/invokeControl
Budgets & limits
Put a time limit on every agent turn, and a per-occurrence and daily ceiling on every routine. See where the money went and stop runaway work before it becomes a surprise bill.
Private models
Use a small Mobius worker to connect agents to Ollama or another model on your own hardware. The model endpoint does not need to be public.
Reliable execution
Agent work keeps going when things go wrong.
Mobius writes down what an agent is waiting for. If a service restarts, a provider has a temporary failure, or the agent waits for a person, the turn continues instead of starting over.
Resume after a restart
A turn that is waiting on an event or a person is written down, not held in memory. If the service stops, the turn picks up where it left off when it comes back.
Retry temporary failures
A provider that times out or rate-limits is retried with backoff. A failure that persists ends the turn and stays in the transcript for review.
Pause while it waits
An agent can wait for an event or ask a person to approve something. The turn releases compute while it waits and resumes when the answer arrives.
Read the whole conversation
See what was asked, which tools and actions ran, where the agent waited, and which artifacts it produced, both live and afterward.
Actions
Let the agent decide, or run the action directly.
An action is a specific operation, such as posting a Slack message. Grant it to an agent and let the agent decide when to use it, or call it yourself when there is nothing to decide.
Let the agent decide
Give the agent the Slack action as a tool. During its work, the agent can decide whether to send a message, choose the content, and use the result in its next decision.
Run it directly
Call the same Slack action yourself, from your code or the CLI, with the values you provide. No model decides anything, and the result comes straight back.
Your infrastructure
Run private operations inside your network.
Install a Mobius worker on your own machine or in your cloud. Mobius sends it tool, action, and model requests, which it executes inside your network. The worker connects outward, so you do not expose a new public endpoint.
Outbound only
The worker opens and maintains the connection to Mobius. You do not need to accept inbound traffic or change your firewall.
Keep sensitive work local
Run private code and access internal systems on the worker. Mobius receives the result, not direct access to those systems.
Use your own models
Route model calls through the worker to Ollama or another private inference endpoint. The agent can use those models without exposing the endpoint publicly.
Build the agent. Skip the backend work.
You define the agent’s instructions, model, and tools. Mobius handles the production infrastructure around it: reliable execution, memory, budgets, human review, and a complete history of what happened. Those are the parts you would otherwise build and operate yourself.
Get started
Build your first reliable agent.
Start free in minutes, no credit card required. Or follow the docs to define an agent and run it for the first time.