Architecture

Rebyte separates Agent management from Agent execution. The OpenAI SDK is the execution client. Rebyte owns the Agent definition and the managed execution behind the Responses-compatible endpoint.

Control plane                           Execution

agent.toml ── Rebyte CLI ── Agent API   OpenAI SDK ── /v1/responses
                         │                         │
                         └──── managed Agent ──────┘
                               Skills + MCP
                               cloud runtime

Agent API

The Agent API owns reusable Agent configuration and durable interaction state.

Resource	Responsibility
Agent	Instructions, model, maximum steps, tools, MCP servers, and Skills.
Conversation	One durable execution context and transcript.
Response	One input and the Agent execution it produces.

Most integrations start here. Create an Agent, use its ID as model, and run it through the OpenAI-compatible Responses endpoint.

Runtime API

The Runtime API owns direct infrastructure control: isolated environments, commands, files, secrets, network egress, checkpoints, and resource limits.

An Agent can use Runtime capabilities through configured tools without your application managing the Runtime directly. Use the Runtime API when the application itself must control those infrastructure primitives.

Client boundary

Organization API keys stay on a trusted server. Browser and mobile clients call your application, which authenticates its user, adds the Agent ID and organization key, and forwards the Response stream.

Interfaces

Interface	Use it for
OpenAI SDK	Responses-compatible Agent execution.
React package	Optional headless browser chat state. It calls your application server.
Rebyte UI	Optional Agent chat and execution inspector components.
Rebyte CLI	Validate and apply local agent.toml definitions.
Agent REST API	Direct Agent configuration and resource lifecycle.

There is no separate Rebyte Responses client. Server code uses the official OpenAI SDK.