Kenoverse

Kenoverse speaks MCP, so Claude can read your Kenoverse library and add to it.

Hosted MCP Server

npx add-mcp 'https://api.kenoverse.app/mcp?src=mcpservers'

Installs into Claude Code, Codex, Cursor and more

Documentation

The Kenoverse connector

Kenoverse speaks MCP, so Claude can read your library and add to it. The point is the moment mid-conversation where something turns out to be worth keeping: you say "import this to Kenoverse", and it comes back as a concept with insights, questions, and a link to train on.

Endpoint: https://api.kenoverse.app/mcp

Connecting

Claude apps and claude.ai — Settings → Connectors → Add custom connector, paste the endpoint URL. Claude discovers the authorization server, registers itself, and sends you to Kenoverse to sign in and approve.

Claude Code —

claude mcp add --transport http kenoverse https://api.kenoverse.app/mcp

then /mcp inside Claude Code to run the sign-in.

A client that only sends a fixed header — mint a personal access token at kenoverse.app/connect and send it as Authorization: Bearer <token>. Tokens are listed and revocable on the same page.

There is also a Claude Code plugin in this repository (plugins/kenoverse) that bundles the server config and a /kenoverse:import command:

/plugin marketplace add elumixor/kenoverse
/plugin install kenoverse@kenoverse

Tools

ToolWhat it does
import_conceptExtracts insights + questions from text or a url, files them under an optional category, returns view and train links. Takes up to a minute.
list_conceptsThe library, newest first, with mastery percentages. Optional query filters by title.
get_conceptOne concept in full: summary, insights, per-insight mastery, sources.
list_collectionsCollections and their concept counts.
training_statusWhat is due or never seen, and where to start.
whoamiWhich account is connected, and how big its library is.

Every tool answers in markdown so Claude can read it straight back to you. Nothing is destructive — the connector can create concepts and collections, and cannot delete anything.

Privacy and support

The connector reads and writes only the signed-in user's own library. Material sent to import_concept is processed by Google Gemini through the Vercel AI Gateway to extract insights and questions; nothing is used to train models. The full policy is at kenoverse.app/privacy.

Support: kenoverse.app@gmail.com or kenoverse.app/support. Revoke the connector's access any time at kenoverse.app/connect; deleting your account from the Profile tab removes all data and cuts the connector off on its next call.

How it is built

The server is hand-rolled JSON-RPC over one POST route, apps/backend/src/api/routes/mcp/. It holds no session state and never opens an SSE stream: each tool answers in one shot, which is what a serverless function serves reliably. Two protocol eras are handled at once — the initialize handshake used by 2025-03-26 … 2025-11-25, and the handshake-free 2026-07-28 revision, which carries client metadata in params._meta instead. GET and DELETE on the endpoint return 405, and Mcp-Session-Id is ignored.

Tool definitions live in services/mcp/tools.ts; their JSON Schemas are derived from the same zod schemas that validate the arguments, so a tool cannot advertise a shape it doesn't accept.

Authorization

Kenoverse is its own OAuth 2.1 authorization server, because the only identity it has is the app's session:

EndpointRole
/.well-known/oauth-protected-resourceRFC 9728 — points at the authorization server. Served from middleware/4.well-known.ts, since Nitro's route scanner skips dot-directories.
/.well-known/oauth-authorization-serverRFC 8414 metadata.
POST /oauth/registerRFC 7591 dynamic registration — clients register themselves.
GET /oauth/authorizeValidates the client, then redirects to kenoverse.app/connect.
POST /oauth/consentCalled by that page with the app's session JWT; returns the redirect URL carrying the code.
POST /oauth/tokenauthorization_code (PKCE required) and refresh_token.

Every client is treated as public — PKCE stands in for a client secret, which is what both Claude Code (loopback redirect on an ephemeral port) and claude.ai (hosted redirect) need. Authorization codes are single-use and die after ten minutes; refresh tokens rotate, so a stolen one stops working as soon as the real client refreshes.

Tokens are stored as sha256 hashes in McpToken, and every request re-checks that the account still exists and has an email, so deleting an account cuts off its connector on the next call, not at the next token expiry. Imports through the connector go through the same rate limits, input caps and AI budget as the app.

Install attribution

Each directory listing uses its own endpoint URL, so installs can be counted per listing:

ListingEndpoint
Claude directoryhttps://api.kenoverse.app/mcp?src=claude-dir
Smitheryhttps://api.kenoverse.app/mcp?src=smithery
mcp.sohttps://api.kenoverse.app/mcp?src=mcpso
Directhttps://api.kenoverse.app/mcp

src must match [a-z0-9-]{1,32}; anything else counts as direct, and the request is otherwise served exactly as plain /mcp. The 401 challenge and the protected-resource metadata echo src, so the metadata's resource equals the URL the client was configured with (RFC 9728 §3.3).

The first authenticated request on a token stamps source, userAgent and firstUsedAt onto that McpToken row, and the first use of any of its tokens stamps the same onto the McpClient row (whose name is the DCR client_name, e.g. Claude, Cursor). Each stamp is written once. apps/backend/scripts/test-metrics.ts groups installs by source and client, and trial starts and active users by each user's first install source.