Not out of laziness. The context switch is the problem. I'm deep in a branch, the thing works, and announcing it means leaving the terminal, opening a doc, re-explaining my product to a chat window that has never heard of it, rewriting the result for three platforms, and finding a scheduler. By the time I've done that I've lost the thread I was on, so I don't do it, and the feature ships in silence.
So we built an MCP server for it. Claude Code already knows what I just built — it helped me build it. It should be able to write the post too.
write a LinkedIn and X post about the feature I just shipped, then schedule both for Tuesday morning
The server is the MCP surface for Cadencz, which is a social workspace for founders. The important part for this post is not the app, it's the thing the app has that a raw model doesn't: a Website Brain.
You point it at your URL once. It reads your positioning, your audience, the proof points you actually have, and how you already write. That gets stored. Every draft afterwards starts from your real product rather than from a prompt where you've re-typed "we're a B2B SaaS that helps teams..." for the fortieth time.
That's the difference between an AI that writes posts and an AI that knows what you built. It's most of why AI social content reads like it could belong to anyone — the model has a memory of your prompt, not of your product.
Eight agents work off that same stored context. One reads the site, one works out the brand, one writes your voice guide, one plans, one drafts, and one exists purely to reject weak drafts before a human sees them. That last one produces nothing and improved the output more than any of the others.
Tool What it does listworkspaces The workspaces this token can see listchannels Connected accounts and whether the plan covers them listbrains The Website Brains in this workspace querybrain Retrieve product context for a question getcurrentplan The current weekly content plan and its items getplanusage Allowance usage against the plan's caps listscheduledposts Posts committed to publish in the coming days getanalytics Engagement totals per platform over a window
Tool What it does composepost One platform-native post from an idea, grounded in the Brain planweek Runs Plan My Week across the whole content calendar schedulepost Commits a draft to a time cancelscheduledpost Pulls a scheduled post back to draft approveplanitem Approves a plan item into a scheduled post publishnow Queues for immediate delivery
Two are owner-only and don't even appear in tools/list for a token that can't use them: connectchannellink and getsubscription.
composepost writes for the target platform rather than producing one blob you paste everywhere. LinkedIn and X have never rewarded the same shape, and cross-posting one paragraph is the fastest way to be ignored on both.
Connecting it bash claude mcp add --transport http cadencz https://api.cadencz.com/mcp \ --header "Authorization: Bearer cdzmcp..."
Any client that sends a static Authorization header works the same way — Claude Code, Claude Desktop, Codex CLI, Cursor. It's also listed in the official MCP registry as com.cadencz/cadencz, so anything that reads from there can find it.
Tokens are minted at Settings → MCP by an account owner. The raw string is shown exactly once; the backend stores only its sha256, so there's nothing to re-reveal and no support path that hands it back to you.
A token that can drive your social accounts is a serious thing to hand to an agent, so three properties matter.
A token acts as the person who minted it, role included. A scheduler's token can compose, schedule and approve. It cannot connect platforms or see billing — those tools don't appear in its tools/list at all, rather than appearing and then failing.
The plan gate runs per request, not at mint time. Downgrade below Pro and every existing token stops working on the next call. Remove a member and their tokens die with the seat check.
The tools that move money don't exist. No checkout, no cancel, no billing portal, nothing that mints credentials. A leaked token must not be able to buy or cancel a plan, and the way to guarantee that is to never build the tool.