Claude and Codex as Services

Run agents built with your preferred harness in your own infrastructure. AgentOS provides the service layer around them, handling orchestration, authentication, streaming, session storage, and background execution.

Claude and Codex can power an assistant inside your application. You supply instructions and business tools; the harness runs the model and tool loop. AgentOS gives your application an API for starting work, streaming results and managing sessions.

Build a support operations assistant

Imagine a customer asking about a delayed order. Your agent reads the support ticket, checks the order through your MCP server, and prepares a resolution. A support operator can then ask it to save an internal note. The product receives streamed responses and can inspect the recorded tool results.

This is the example used throughout the harness guides. It starts with a local service and a supplied ticket, then introduces real MCP tools, a persistent conversation and background execution. The tool fixture changes only a disposable local database.

Keep your product in control

Your application owns its users, permissions and business data. MCP services expose the operations the agent may perform. AgentOS runs in infrastructure you operate, with storage and credentials you configure.

Longer tasks can execute in the background while the application polls or follows a stream. For multiple replicas, configure shared storage and event/cancellation coordination. Set an explicit retry policy and make state-changing tool operations idempotent.

Native conversation recovery differs between Claude and Codex, and conversation storage does not preserve an execution workspace. The Agent Harnesses overview explains the division of responsibility; Deploy and Operate covers the requirements for taking the service beyond local development.

Prepare the product integration

Use an application-owned operation ID for each business mutation and serialize turns within a session across your application replicas. Request streaming when submitting a run if the client needs later SSE catch-up. Test recovery with shared Postgres and Redis before relying on multi-replica behavior.

Add structured output, attachments, retry handling and run inspection as your product needs them. Claude also supports checkpoint replay. Session sandbox provisioning and the pending external-agent OpenInference integration are separate from this service setup.