cereal Open console

You build the agent.
We run the loop.

Cereal runs your agents: model calls, tool coordination, and durable sessions. Your tools and hooks shape what happens along the way.

See how it runs See the API

Your models. Your keys.Bring the providers you already use.

Your tools stay with you.Connect your own tools over MCP.

Your logic, in the loop.Shape execution with hooks.

We run the agent.
You shape its behavior.

Start a session. Cereal drives the model-and-tool loop, calling your application where you want control.

Inside a sessionIllustrative execution

Running in Cereal

Completed
  1. Start a sessionYour instructions and context

    Your application sends a task over HTTP. The session pins an agent version and records the conversation.

  2. Call the modelYour provider and model

    Cereal builds the prompt from session history, calls your configured model, and records its response and usage.

  3. Run the requested toolsYour policies and hook decisions

    For a gated tool call, Cereal asks your hook what to do. Allowed calls go to the connected tool; its result returns to the session.

  4. Continue the loopHistory and progress kept together

    The next model call uses the tool results. Cereal continues until the turn finishes, waits for a decision, or reaches a configured limit.

  5. Stream results and deliver eventsBack to your application

    Stream the response to your UI and subscribe to durable events with hooks. Inspect model calls, tool results, and usage in the console.

The session keeps its history, ready for the next message.

Expand a step to inspect it.

An illustration of execution, with no live model or tool calls.

Your tool and hook handlers stay in your application. Cereal runs the session.

HTTP API · MCP tools · Durable sessions · Event streams

Use the HTTP API from your stack, or the TypeScript SDK for application helpers. Explore the architecture

01

Keep your stack.

Connect your models and MCP tools. Call your agents over HTTP from the product you already have.

02

Stay in control.

Version your instructions, set limits, and use hooks to apply your own rules during execution.

03

See the whole loop.

Inspect model calls, tool results, and usage in one session trace.

Good agents deserve
a better runtime.

Open the console

Bring your model key. Make your first agent.