Svitrio

AI & MCP API

The AI module is core (always mounted). Provider keys are sealed at rest; reads mask them.

Providers

  • GET /admin/p/{ref}/ai/providers — list with masked api_key_hint (ai:read).

  • POST /admin/p/{ref}/ai/providers — { name, kind, model?, base_url?, api_key?, capabilities?, is_default? } (ai:write).

  • GET /ai/providers/{id} returns the unmasked key for the edit form; PATCH / DELETE /{id} as expected.

Provider kind: openai, anthropic (claude), deepseek, mistral, qwen, zai — or any OpenAI-compatible endpoint via base_url.

Inline translate

POST /admin/p/{ref}/ai/translate

BYOK translation (ai:read). Body { provider_id?, from, to, text, field? }. Dispatches on the provider kind (native Anthropic /v1/messages, or the OpenAI-compatible driver).

{ "data": { "translation": "…", "provider": "…", "model": "…" } }

MCP

Give any MCP client (Claude Desktop, Cursor, claude.ai) an agent-ready endpoint.

  • POST/GET/DELETE /admin/p/{ref}/ai/mcp/tokens — manage MCP tokens (POST returns the plaintext once).

  • POST /mcp — the MCP JSON-RPC endpoint. Auth: Bearer <mcp-token>; the token resolves the project scope on every call.

  • GET /.well-known/oauth-protected-resource, /.well-known/oauth-authorization-server and /oauth/{register,authorize,token} — OAuth 2.1 discovery + flow (RFC 9728) so a client can connect without a pre-shared token.