Run your LinkedIn content pipeline from Claude

Most "AI for LinkedIn" tooling has the same problem: the model has no idea what you have been reading. It writes competent, generic posts because generic is all it has. The fix is not a better prompt, it is giving the model access to your own library.

LinkedMash ships an MCP server, so any MCP client, Claude Desktop, Claude Code, Cursor, VS Code, Windsurf, Codex, ChatGPT, can search your saved posts, draft from them, and put the result in your publishing queue. Here is the setup and the workflow that actually earns its keep.

Connect it once

The server is hosted, so there is nothing to run locally. Grab your API token from settings, then point your client at it. In Claude Code that is one line:

claude mcp add --transport http linkedmash "https://mcp.linkedmash.com/api/mcp?token=<your token>"

For clients that take a JSON config, the same URL goes in as an HTTP server. You can also pass the token as an Authorization: Bearer header instead of a query parameter if your client supports it. Copy-paste configs for each client are in Connect Your Agent.

One thing worth knowing up front: the token scopes to your account. The model sees your saves and your drafts, nothing else, and nobody else's library is reachable through it.

What the model can actually do

The tool surface maps to the product rather than to a chat gimmick. The ones that carry the workflow:

• search_bookmarks — text search or semantic search over everything you saved. This is the one that matters, because "find the posts about pricing objections" works even when nobody wrote the word "pricing".
• list_bookmarks — filtered listing by author, label, media type, or date range, with pagination.
• add_labels — the model can tidy up as it goes, so a research pass leaves your library better organised than it found it.
• create_draft — writes into Content Studio, where you edit it like anything else. Nothing publishes itself.
• schedule_post and get_queue — put the approved draft in a slot, and see what is already queued.
• get_analytics — pull performance back in, so the next research pass knows what worked.

Full parameter reference for each is in MCP Tools.

The workflow that is worth automating

Do not ask the model to "write me a LinkedIn post". That is where generic comes from. Ask it to work from the library, in three passes.

Pass 1: research

Something like: search my saves for posts about hiring, group them by the argument they make, and show me which angles I have collected most. You are asking for a map of your own reading. Nine times out of ten there is a cluster you did not know you had.

Pass 2: draft

Pick a cluster and hand over your own material: draft a post using the structure of the second one, but the example is my own, here is what happened. The saved post supplies the shape. You supply the substance. That division is what keeps the output sounding like you.

Pass 3: queue

When a draft survives your edit, ask for it to be scheduled into the next open slot. The queue is the same one the web app shows, so you can move it later without touching the chat again.

The model is not the writer. It is the research assistant that already read everything you saved and never forgot any of it.

Where it falls down

Two honest limits. First, if your library is thin, the output is thin. MCP does not manufacture taste, it retrieves it, so the first month of saving matters more than the prompt.

Second, resist full autonomy. It is technically possible to wire draft-and-schedule into a single instruction and never look at it. Do not. The failure mode of an unattended content pipeline is not a bad post, it is thirty bad posts under your name before you notice.

If you want scheduled, hands-off automation for the mechanical parts, filing saves into a database, tagging, exports, use webhooks or the n8n integration for that and keep the human in the loop where the writing happens.