Model Relay documentation
Setup
- Open Connections, sign in and add at least one provider API key.
- Add the connector in Muse. The endpoint is
https://relay.svyat.site/mcp. - On the consent page, keep Allow sending prompts on for full use, or turn it off for read-only access.
Authentication
OAuth 2.1 authorization code with PKCE (S256) and dynamic client registration. Metadata: https://relay.svyat.site/.well-known/oauth-authorization-server and https://relay.svyat.site/.well-known/oauth-protected-resource/mcp. Access tokens last one hour, refresh tokens rotate and last 90 days, and tokens are valid only for this MCP endpoint.
| Scope | Allows |
|---|---|
models:read | Read tools only. This is the read-only connection. |
models:invoke | Sending prompts to models. |
Tools
| Tool | Class | Scope | Inputs | Returns | Side effects |
|---|---|---|---|---|---|
list_models | Read | models:read | provider, search (both optional) | Up to 100 model references with name and access method, the total count and whether the list was cut. | None. Asks each connected provider for its model list, which providers do not bill. |
usage_report | Read | models:read | days, 1 to 365 (default 30) | Calls, tokens and estimated cost per access method. | None. |
account_status | Read | models:read | None | Plan, price, limits, connected providers and the management link. | None. |
ask_model | Write | models:invoke | model, prompt, context, system, max_tokens, via | Answer, model, stop reason and token usage. | Sends the text to the chosen provider. The provider may charge the user's account. |
compare_models | Write | models:invoke | models (2 to 4), prompt, context, system, max_tokens | One answer or error per model. | Sends the text to every chosen provider. Each may charge the user's account. |
route_task | Write | models:invoke | task_kind (code, long_document, quick, general), prompt, context | Answer, the model used and token usage. | Sends the text to the selected provider. The provider may charge the user's account. |
No tool is a sensitive write: none buys anything, publishes, or sends anything to another person.
Model references look like claude:claude-sonnet-5-5, openai:gpt-6.1-sol, gemini:gemini-4-argon or openrouter:openrouter/auto. Use list_models for the ones your keys can reach.
Status and failures
Model calls finish within the request. stop_reason says why an answer ended. A call that returns an error produced no answer. Retrying sends the prompt again and may be billed again. In compare_models, one failed model does not hide the others.
Errors
Failures are tool errors with the text code: message.
| Code | Meaning |
|---|---|
no_access_for_provider | No key is connected for that provider. Add one at /connect. |
provider_auth_failed | The provider rejected the stored key. Replace it at /connect. |
provider_limit_reached | The provider's rate or spending limit was reached. It is not worked around. |
provider_error | The provider failed or could not be reached. Nothing was returned. |
unknown_model | The model reference is not valid. Call list_models. |
rate_limited | The per-user limit was exceeded. Retry after a minute. |
plan_limit | The daily call limit of the plan was reached. |
read_only_connection | This connection was granted read-only access. |
unauthorized | The access token is missing or invalid. Reconnect. |
Limits
- 30 model calls per minute per user.
- 16,000 output tokens per call at most.
- Prompt 200,000 characters, context 1,000,000 characters, system instructions 50,000 characters.
- Provider limits also apply and are returned unchanged.
Requirements
A Google account and an API key from Anthropic, OpenAI, Google AI Studio or OpenRouter. Model and regional availability follow your provider account.
Home · Documentation · Connections · Privacy · Terms · support@svyat.site