Use Medium MCP Workflows To Move AI Drafts Into A Publishing Queue

Narrareach uses Model Context Protocol over HTTPS. In Narrareach, open Settings and Connect LLMs, then copy the Narrareach MCP URL into the compatible client. Claude and ChatGPT discover the endpoint and register an OAuth client dynamically.

The client sends the user through Narrareach sign-in and permission approval. Dynamic client registration and PKCE S256 handle setup without asking the user to paste a shared client secret or store a Narrareach publishing token in a desktop configuration file.

After authorization, test the connection with a read action before creating content. The client should show the Narrareach tools it can call, and every result should be scoped to the signed-in Narrareach user.

The Narrareach MCP catalog includes tools for listing notes, creating drafts, scheduling Notes, scheduling articles, reading the queue, rescheduling an item, cancelling an item, inspecting inspiration, and working with available analytics and profile context. Tool availability depends on the authenticated user, plan, connection state, and current destination support.

A useful article-to-Notes workflow begins with the source text and a specific editorial goal. The assistant can prepare several distinct Notes, show the drafts, apply selected destinations and times, and schedule only the approved set. Images can be attached through documented public URLs or inline-media paths where supported.

MCP can also coordinate cross-platform versions, but it should not send identical copy blindly. The assistant should return separate drafts when LinkedIn, Substack Notes, X, Bluesky, or Threads need different length, context, or formatting.

A publishing agent becomes more useful when it can inspect what happened after publication. Narrareach can expose available post, traffic, subscriber, and queue context so the assistant can compare topics, formats, sources, and timing before suggesting the next draft.

The agent must preserve evidence boundaries. A tracked subscriber source is stronger than a same-day increase; a platform impression is not a subscriber conversion; and a high-performing Note on one account does not prove a universal best time. Narrareach keeps direct attribution, correlation, and unknown sources distinct in the product context.

A strong prompt asks for a decision with evidence: identify the top two topics by meaningful downstream signal, explain the confidence level, and prepare one follow-up variation for review. This is more actionable than asking the assistant to produce another generic batch from vanity metrics alone.

MCP access begins with an authenticated Narrareach user and OAuth authorization. Tool calls are scoped to that user, and plan entitlements still apply on the server. The UI is not the security boundary: a request must be authorized again when it reaches the publishing route.

Use least privilege and retain a review step for consequential actions. A drafting assistant does not automatically need permission to publish immediately, change platform connections, or inspect every analytics surface. Disconnect clients that are no longer used and review unusual queue changes in Narrareach.

Narrareach documents a limit of 120 MCP request units per ten minutes, with one unit consumed per item in a JSON-RPC batch. Clients should stay within the published limit and back off rather than retrying a rejected batch continuously.