Run your social media from Claude Code: a practical MCP playbook

What is a social media MCP server?

The Model Context Protocol is an open standard for connecting AI applications to outside systems. A server publishes tools with typed inputs; a client such as Claude Code, Cursor, Codex, or Windsurf lists them, calls what it needs, and shows you the results. The protocol specification defines the handshake, the schemas, and the annotations clients use to decide what to approve without asking.

A social media MCP server is that standard applied to a content pipeline: list the drafts, create a post, schedule it, read what performed. The argument for connecting one is not fewer clicks. It is fewer transcriptions: your assistant already holds the context you would otherwise describe a second time.

The trust boundary: what an agent may and may not do

Two things decide what your agent can do: what the key allows, and what your prompt asked for. Only the first is enforced. A model can misread an instruction; it cannot call a tool its key does not carry. Scope is the boundary to reason about.

Krevaya filters the tool list per key: a read-only key lists 21 tools, a full-scope key 51. That saves context as much as it protects you, since every listed tool costs prompt tokens.

Be precise about what review-first protects. A post created without a schedule time lands as a draft, and a draft cannot reach a network until it is approved. But approving is itself a tool, and a key with the publish scope can call it, or bypass the draft entirely by creating a post with a future timestamp. Review-first stops an accident. It does not overrule authority you chose to grant.

Connect a client in about two minutes

1. Mint a scoped key. In Krevaya, Settings → API keys. Keys are revocable and you can hold ten, so a key per agent is cheap.
2. Add the server. Krevaya's hosted endpoint is https://api.krevaya.com/mcp, authenticated with an Authorization: Bearer header.
3. Verify. Ask the agent to call whoami. It returns the account and the scopes the key really carries.
4. Read the tool list once. A missing tool means a missing scope, not a bug.

claude mcp add --transport http krevaya https://api.krevaya.com/mcp \
  --header "Authorization: Bearer krv_..."

Other clients want the same three facts in their own shape: a URL, a transport, and a header. Two settings cause most failed connections. Clients disagree on how a remote HTTP server is declared, and most default to a tool timeout near a minute, shorter than a real generation run. Raise it, and check the connection guide for your client's key names.

Krevaya also ships four sessions as protocol-level prompts, which most clients surface as slash commands: a content-ops pass, a weekly review, a repurpose run, and a themed-week plan. The review-first order travels with the connection.

A session recipe: read, triage, schedule, draft

The order is the method. Bad sessions start by generating, which leaves you triaging a bigger queue with no more information.

1. Read the state. Queue counts, thirty days of results, any standing recommendation.
2. Triage the drafts. One at a time, each with a recommendation and the evidence behind it.
3. Reject with a reason. Krevaya takes a short code beside the free text, which makes it countable.
4. Schedule what you approved. Omitting the time takes the next free slot.
5. Then draft next week, now that you can name the theme that performed.
6. Close the loop. A queued post is a plan; a published post is an outcome.

You:  Run a content-ops pass. Read results first, then show me
      each draft with a recommendation. Approve nothing yet.

  > get_queue_summary       6 drafts, 2 approved, 3 scheduled
  > get_analytics days=30   carousels outperformed plain text
  > get_audit               one fix: how-to series on a fixed day
  > list_posts status=draft 6 returned

  1  "What to bring to a first call"   approve
  2  "We are the best studio in town"  reject (too_salesy)
  3  "Pricing changes next month"      hold, offer unconfirmed

You:  Approve 1. Reject 2 with that code. Hold 3 until Friday.

  > approve_post 1 · reject_post 2 code=too_salesy
  > schedule_post 1
  Approved 1 / rejected 1 / scheduled 1 / held 1.

That transcript shows the shape of a session, not a recording of one. Your drafts will differ. The sequence should not: the agent reads, you decide, the agent acts.

Prompt patterns that keep a session honest

• Name the order first. "Read results before you show me drafts" removes most of the variance between sessions.
• Withhold approval explicitly. "Approve nothing until I say yes, per draft." Pair it with a read-only key to have that enforced rather than requested.
• Require a structured reason. Ask for a rejection code the tool accepts, not a paragraph.
• Tell it to relay refusals. A refusal, such as a spent cap, should reach you. Blind retries turn one into twenty.
• Ask for a closing summary. Approved, rejected, scheduled, held, and what to watch next.

Notice what none of these ask for. None asks the agent to judge what is on brand, or to publish for you. It is fast hands on a pipeline you still own.

Failure modes worth knowing in advance

The repeat is the one to take seriously. An agent retries in milliseconds, without the hesitation a person feels after a call that may already have worked. Krevaya annotates every tool as read-only, destructive, or idempotent, so a client knows which calls it can make without asking you first. Truncation deserves the same care: a result is capped at 40 KB and marked as cut.

How to evaluate any social media MCP server

This checklist is vendor-neutral. Most of it is answered by public documentation; the rest by connecting a trial key and watching.

One test is easy to skip: can the agent read outcomes, or only write content? A server that exposes creation without measurement fills a queue but cannot tell you whether that helped.

What Krevaya's MCP server exposes today

The server is hosted: one endpoint, the same scoped keys as the REST API, nothing to install. A full-scope key lists 51 tools covering triage, creation, design, media, scheduling, and measurement; a read-only key lists 21. Six further tools curate the global trend rails and are listed only for platform administrators.

Everything an agent submits runs the pipeline's own rules: the same safety screening, plan limits, and review queue as work done by hand. Auto-publish is a setting you turn on per pipeline, under daily caps, for a workflow that has earned it. Approved posts reach Facebook Pages, Instagram, Threads, LinkedIn, and Bluesky; see the publishing page for current formats.

Limits are published rather than discovered: 120 requests per minute per account, shared across every programmatic surface so no client can round-robin past the budget, plus a slower hourly ceiling on generation. Webhooks run the other way, with signed deliveries for published posts, failures, and competitor breakouts.

Two honest limits. Authentication is a key header rather than OAuth, so a client that accepts only OAuth cannot connect yet, and there is no per-call audit endpoint: your record is the named key and the post history.

MCP and API access is included from the Plus plan up. New accounts join through the waitlist.