Configure an OpenAI-Compatible Provider Server

Open /editor and AI configuration. Select the compatible-provider option and replace any loopback preset with your provider's documented HTTPS API base. For OpenAI itself, use https://api.openai.com/v1 and an authorized model, or select the dedicated OpenAI option. Vinyasa appends /chat/completions, so do not include that suffix twice. Enter the exact model identifier and credentials required by the provider server; credentials are held in tab memory. This does not install or run Vinyasa locally.

Set Model context window to the runtime's configured context, not merely the model family's advertised maximum. The default is 50,000 tokens, but smaller local deployments may need less. Save configuration stores these values in memory; it is not a network connection test. Generate proposal performs the actual request.

The endpoint must implement chat completions with system/user messages, JSON-object response format, max_tokens, and a completed choice with finish_reason set to stop. The model must follow the bounded schema/search/read/propose protocol, not return Markdown fences or an entire YAML document. A server accepting an OpenAI-like URL is not enough if these capabilities are missing.

Use an HTTPS provider endpoint. URLs containing embedded credentials, query strings, or fragments are rejected. No request is silently rerouted to a different provider.

Requests originate in the hosted application's browser tab, not a Vinyasa inference proxy. Your provider server must permit through its origin/CORS configuration. Configure narrow origins and server authentication; do not expose an unauthenticated inference server to bypass a browser error. Provider-server configuration is distinct from the hosted Vinyasa MCP server.

If a request fails, check the exact model name, JSON support, origin restrictions, browser permissions, timeout, and context setting. A 401 suggests credentials, 429 suggests rate limits, and truncated generation suggests insufficient output capacity or an oversized task. Endpoint errors leave the canvas unchanged.

An OpenAI-compatible provider may forward requests to other services. Check its retention, billing and processing policies before sending layouts. Website assets and remote images have their own network behavior. For external coding agents, connect to the hosted MCP server at /mcp using the MCP configuration guide. No local Vinyasa process is required. Review the AI data flow before using confidential layouts.