Configure Claude Code, Codex and other model providers

How providers connect

AiOrch supports several provider implementations. Claude Code, Codex and Kimi use CLI integrations; OpenAI uses an API integration; Ollama connects to a configurable inference endpoint. Each has its own authentication and availability requirements.

The installer offers provider setup assistance. In the dashboard, use Settings → Providers and the provider’s connection test. Choose from the models your runtime actually reports as available, rather than assuming a model name in an example will work for every account.

Use Claude Code and Codex in one workflow

Configure and test both CLI providers. In a new session, Model Configuration lets you choose separate models for Agent Work, Planning and Review. For example, you can choose an available Codex model for implementation and an available Claude model for review, or reverse those roles.

For different implementation models within a session, enable Smart Model Routing and select models for the hard, medium and easy task tiers. The planner classifies task difficulty; the configured tier determines the implementation model. Planning and review selections remain separate.

Before execution, review the proposed plan and model choices. A mixed-provider workflow still needs clear scopes, working credentials, adequate rate limits and a repository-specific validation policy.

Use an API provider

For the OpenAI API provider, configure your key in Settings and use the base URL appropriate for your deployment. The runtime also supports documented ORCH_OPENAI_API_KEY and ORCH_OPENAI_BASE_URL configuration. Keep credentials out of task descriptions, repository files and shared logs.

An OpenAI-compatible gateway must provide the API behavior and model capabilities the selected provider expects. Endpoint compatibility alone does not establish reliable agent tool use; test the connection and a small task before relying on it.

Use the provider’s current account/model documentation to check availability. AiOrch does not add a token markup; inference billing and limits come from your provider.

Use Ollama for local inference

Configure the Ollama server URL in Settings or through ORCH_OLLAMA_BASE_URL. The URL must be reachable from the AiOrch container. Inside a container, localhost refers to that container, not automatically to an Ollama process on your host.

Install the intended model on your Ollama server and test its availability. Coding work requires a model that supports tools; not every local model does. Planning or review requirements can differ from execution requirements.

Local inference keeps that inference traffic on your configured infrastructure. License activation and validation still contact the configured license service, and GitHub delivery contacts GitHub when enabled. Local inference does not by itself make the whole installation air-gapped.

Choose a useful first configuration

1. Begin with one working implementation model and a tested reviewer.
2. Keep plan approval enabled and use a small concurrency limit.
3. Check the task outcome and revision history before changing model routing.
4. Measure provider usage, elapsed time and test outcomes for the same task before making a cost or quality comparison.

For scheduling and common-file conflicts, read parallel coding agents with git worktrees. For data handling, read architecture and data flow.