Legacy initialize/session traffic is rejected.
DEVELOPERS · MCP
MCP for StoryVerse, without borrowed authority.
Official MCP v2, resource-bound OAuth and small domain tools. Production discovery and the permission flow are available. Each request is checked against the connected account and its selected permissions.
Production connection · explicit permission requiredAVAILABILITY
Deployed is not the same as connectable.
PRODUCTION MCP ENDPOINT
https://mcp.miyaniva.com/mcpThe MCP and authorization domains are active. The registered ChatGPT client uses public PKCE and a separate connection consent. Server and controlled browser tests passed; a complete user-driven ChatGPT tool and revocation acceptance is still pending. Endpoint availability alone does not prove that client journey.
PROTOCOL
Modern MCP only.
Four progressive views; core results work without UI.
Stateless JSON responses; subscriptions disabled.
No false runtime capability claim.
OAUTH
Your password stays with Miyaniva.
GET /.well-known/oauth-protected-resource GET /.well-known/oauth-authorization-server GET /authorize?resource=…&code_challenge_method=S256 POST /token POST /revoke
Authorization codes are opaque, single-use and valid for ten minutes. Redirect URIs match exactly. Access and rotating refresh tokens are bound to client, resource, family and current connection state.
CONNECT
Connect with explicit permission.
In ChatGPT Plugins, choose Add custom MCP server. Use https://mcp.miyaniva.com/mcp with OAuth, a user-defined client ID miyaniva-chatgpt, no client secret and token authentication method none. Sign in on the Miyaniva permission page and choose your content. This is a custom connection, not an approved directory listing.
A client must have an explicitly registered ID and exact callback address before it can connect. A working ChatGPT registration does not automatically register Claude. No directory listing or Claude end-to-end acceptance is claimed.
Use a 2026-07-28 Streamable HTTP client that follows protected-resource discovery, authorization-code + PKCE, refresh and revoke. No API key or raw Firebase token is accepted.
OAuth issuer: https://auth.miyaniva.com · resource: https://mcp.miyaniva.com · authorization: /authorize · token: /token. There is no client secret for the registered public ChatGPT client. Do not enable OpenID Connect for this flow.
SCOPES
Operation class first. Resource grant second.
Public discovery is an R0 resolver class, not a private-account consent grant. Book results still require canonical visibility and explicit owner opt-in.
TOOL CATALOG
37 bounded contracts. No generic executor.
search/get public books and characters · get public creator · get public community
profile · granted books · chapters · character · publication status
profile, book, chapter, character and community drafts · prepared language-variant attachment · verified image/video upload
prepare/confirm chapter and language-variant publication · prepare/confirm resource video attachment
delete · admin · maintenance · raw query · arbitrary URL fetch · translation generation
EXAMPLES
Revision-safe by construction.
// Create a private draft create_chapter_draft({ bookId:"selected_book", title:"Chapter 7", content:"…", idempotencyKey:"request_01H…"}) // Review before public effect prepare_publish_chapter({ bookId:"selected_book", chapterId:"chapter_7", expectedRevision: 1, idempotencyKey:"request_01J…"})
RATE + COST LIMITS
Bounded at every layer.
- Per-user, connection, client, tool and burst windows are enforced server-side.
- Expensive media operations have a separate hourly budget.
- Lists and public searches carry hard result ceilings and bounded candidate windows.
- Images are direct-uploaded to private R2 with a 15-minute signed PUT, then byte-verified before attachment.
- Retries use idempotency keys; concurrent edits use expected revisions.
PRIVACY
Content is data, never authority.
Private manuscript text may be returned only by an explicitly authorized tool. It cannot change scopes, ownership, runtime policy or tool behavior. Audit records are opaque and exclude manuscripts and bearer values. Public discovery, connected-assistant access, full-chapter summary permission and model training are separate decisions.
No staff/admin tool, internal Firestore schema, secret name/value, raw storage credential or unrestricted reader is part of the public client contract.
FAQ
Before you connect.
Can an assistant see every draft in my account?
No. Consent uses selected resources and separate scopes. Canonical ownership is checked on every private request.
Does Miyaniva translate my chapters?
No. Your chosen assistant prepares the language version. Miyaniva only attaches that ready text to the canonical chapter, shows a preview and publishes it after confirmation.
Can it publish without asking?
No. Prepare and publish are separate tools. The publish intent expires after ten minutes, can be used once and is tied to the exact reviewed revision.
Does public AI discovery mean model training?
No. Public discovery, full-chapter summaries, connected-assistant access and model-training permission are separate decisions.
Is this live for everyone?
The production endpoint and registered ChatGPT client are available. Other clients need their own explicit registration. Complete user-driven ChatGPT tool and revocation acceptance is pending; a directory listing is not claimed.
SUPPORT + CHANGELOG
Release truth travels with the docs.
The current 37-tool registry passed controlled security tests. Production routing, the ChatGPT registration and separate connection consent are active. End-to-end user acceptance is pending; server deployment alone does not establish it.
Include client name, protocol version and the public request error code—never send tokens or manuscripts.
Prepared language-variant attach/preview/publish.
Resource video prepare/confirm.
Machine-readable Miyaniva attribution.
Opt-in public discovery and progressive Apps.