Oceans LLM

Appearance

Client Harness Configuration ​

See also: Model Routing and API Behavior, Budgets, Budgets and Spending

Model Config Page

Oceans generates client configuration snippets from the live model catalog so users can point local agent harnesses at the gateway without hand-writing model metadata.

Open /admin/models, select one or more configurable models, then click Generate config. You can also use the row-level client config action to generate a one-model snippet.

Generated snippets are available for:

The snippets use the gateway model ids shown in the Models table. API keys are still created and governed in Oceans; the local client config only tells the harness which gateway URL, key variable, and model ids to use.

By default, snippets use the local development gateway base URL http://127.0.0.1:3000/v1. Production deployments should set GATEWAY_CLIENT_CONFIG_BASE_URL on the gateway process, or gateway.clientConfigGatewayBaseUrl in the Helm chart, to the public gateway API base URL users can reach, for example https://gateway.example.com/v1.

OpenCode and Pi can include multiple selected models in one generated file. When the selection mixes Anthropic Messages models and OpenAI-compatible models, Oceans emits separate provider entries so each provider keeps the correct client adapter. Claude Code only includes selected models that use Anthropic Messages; non-Anthropic selections are ignored for the Claude Code tab instead of generating invalid overrides. Codex snippets require a single Responses-capable model selection.

OpenCode ​

The OpenCode tab emits opencode.json content for the user-level OpenCode configuration at ~/.config/opencode/opencode.json. The dialog shows the configuration path, gateway API key environment variable, and OpenCode configuration docs before the copied JSON block.

Pi ​

The Pi tab emits models.json content for Pi custom provider/model configuration. Pi settings are separate configuration: use ~/.pi/agent/settings.json for global settings and .pi/settings.json for project overrides. The dialog shows those paths together with the generated provider configuration path and links to the Pi settings docs.

Claude Code ​

The Claude Code tab emits .claude/settings.json content with the SchemaStore Claude Code schema URL. The gateway settings block includes:

  • ANTHROPIC_AUTH_TOKEN, set to a replaceable gateway API token placeholder
  • ANTHROPIC_BASE_URL, set to the Claude-compatible gateway base URL
  • CLAUDE_CODE_ENABLE_GATEWAY_MODEL_DISCOVERY, so Claude Code can discover gateway-routed models
  • ANTHROPIC_MODEL, ANTHROPIC_SMALL_FAST_MODEL, and the matching default model class variable
  • modelOverrides, mapping Claude Code's Anthropic model ids to the selected gateway model ids

Claude Code appends Anthropic endpoints such as /v1/messages and /v1/models to ANTHROPIC_BASE_URL. Do not include /v1/messages in the configured gateway URL. OpenCode and Pi still use the OpenAI-compatible /v1 base URL, grouped by client API style when needed.

The same dialog also shows a second settings.json block for a smaller local Claude Code experience. It disables telemetry, experimental betas, 1M context, and related UI/reporting behavior, and sets a lower auto-compact window.

Codex ​

The Codex tab emits config.toml content for the user-level Codex configuration at ~/.codex/config.toml. Codex ignores provider, auth, and telemetry keys from project-local .codex/config.toml files, so copy this block into the user-level config unless the Codex docs say otherwise.

Oceans includes Codex only when the selection contains one Responses-capable model. The generated TOML uses the selected gateway model id, configured gateway base URL, Oceans provider id/name, and gateway API-key environment variable. It also sets wire_api = "responses", disables OpenAI auth for the proxy provider, disables analytics, disables OTEL prompt logging, and sets model_reasoning_effort = "medium" for a predictable default.

Responses-compatible Bedrock routes can still support only a subset of OpenAI-hosted tools. For example, Bedrock Mantle GPT-5.5 is usable through Codex over the Responses API but does not support the OpenAI-hosted image_generation tool; see Provider API Compatibility.

Budgets And Access ​

Client snippets do not grant access by themselves. A request is accepted only when the gateway API key is active, the caller has model access, and any applicable hard budget still has room.

Budget scopes are independent of the client harness:

  • human users can have an overall user budget
  • service accounts must have an active service-account budget before active service-account API keys can be used
  • human users can also have user model budgets for one gateway model or one upstream model name

Use /admin/spend-controls to configure those budgets. For the full taxonomy and setup workflow, see Budgets.