Skip to main content

​
claude.ai / Claude Desktop (OAuth)

Add a custom connector with the URL:
The client discovers everything itself: the 401 challenge points at the RFC 9728 resource metadata, which points at this instance as the authorization server; the client self-registers (RFC 7591), the member signs in and approves the consent screen, and the connector holds a 15-minute access token with a rotating refresh token. Deactivating the member kills the connector instantly (tokens are re-checked against the live membership on every call). The consent screen names the workspace being granted, and the grant is bound to exactly that workspace for its whole life — refreshes included. Members of several workspaces pick one at consent; connecting the same client to another workspace is a second consent (send prompt=consent to force the picker past an existing grant).

​
Claude Code / CLIs (API key)

/mcp also accepts the instance’s API keys directly — no OAuth dance:
Mint keys in the dashboard (API keys → Create key). Scopes gate what tools can do: presentations:read for reads, presentations:write for mutations.

​
Verify an instance

get_me returning your identity proves discovery, registration, login, consent, token exchange, JWKS verification, and the live membership check in one call.

​
The tool set

All product tools are prefixed slideless_ and act as the connected user — identity always comes from the verified credential (OAuth token or API key), never from a tool parameter. Reads require presentations:read, writes presentations:write; the API’s fail-closed allowlist and per-deck read privacy apply unchanged (a tool can never read a deck the caller can’t). The /mcp transport caps request bodies at 1 MiB, so inline uploads are bounded at 768 KiB of decoded content (base64 inflation means anything larger cannot fit the JSON-RPC envelope anyway) — the tools answer a clean error pointing at slideless push / slideless pull for bigger decks.

​
For products extending the template

Tools live in apps/server/src/mcp/. Conventions (ported from a proven predecessor MCP template): reads declare readOnlyHint and check presentations:read; writes describe themselves as confirm-first and check presentations:write (tool-level checks are UX — the API’s fail-closed allowlist in middleware/scopes.ts is the enforcement point); tools call the instance’s own API in-process forwarding the caller’s bearer (MCP is just another API client — never a privileged path); the acting user is NEVER a tool parameter (identity comes from the verified credential); map domain error codes to model-readable hints in mcp/errors.ts. New behavior lands in the API first — the MCP tool is a thin projection of it.