Post to social media from Claude Code

Anthropic describes Claude Code as "an agentic coding tool that reads your codebase, edits files, runs commands, and integrates with your development tools," available in the terminal, IDEs, a desktop app and the browser. It connects to external tools over MCP, and its docs say MCP servers configured once work across those surfaces.

That makes it a good fit for posting that starts from code: release notes from a changelog, a launch post from a merged PR, a weekly summary from commit history. The repo is already in context, so the draft is grounded in what shipped.

Claude Code's MCP docs give the form claude mcp add --transport http <name> <url> for remote servers, with --header for a bearer token. Export the key, then run:

Scope matters. The default local scope keeps the server to you in this project, stored in ~/.claude.json. --scope user loads it in all your projects. --scope project writes .mcp.json in the project root to share through version control, so don't use it with a raw key. Check the connection with claude mcp list, or /mcp inside a session.

export SOCIALS_API_KEY=sk_...
claude mcp add --transport http socials https://socialsforai.vercel.app/api/mcp \
  --header "Authorization: Bearer $SOCIALS_API_KEY"

claude mcp list

Claude Code names MCP tools mcp__<server>__<tool>, so with the server called socials you'll see mcp__socials__list_accounts, mcp__socials__create_post, mcp__socials__get_post and five more. Those full names are what permission rules match on. You rarely type them: Claude reads each tool's description when the server connects and picks the right one from a plain request. The inputs for all eight are on the hub under every tool and its inputs.

Each prompt below lists the tool calls Claude Code will make. Claude reads tool descriptions, so you don't have to name tools yourself.

• "Read CHANGELOG.md, draft a release post for LinkedIn and a shorter one for X and Bluesky, dry run all three, and show me the previews." Calls list_accounts, then create_post with dry_run: true, the long text in text and short versions in platform_options.x.text and platform_options.bluesky.text. After your yes, the same call without dry_run and with an idempotency_key.
• "Schedule that post for Tuesday at 9am Pacific, then show me what's scheduled." Calls create_post with scheduled_at: "2026-11-03T09:00:00-08:00", then list_posts with status: "scheduled". If you change your mind, cancel_post works until it starts publishing.
• "Connect my Bluesky account." Calls create_connect_link with platforms: ["bluesky"] and prints the URL. Open it, enter your handle and an app password, then ask Claude to list accounts again.

This is the first prompt's dry run, from the repo's changelog to three previews. One create_post call carries the long LinkedIn text in text and the two short versions in platform_options, so each network gets its own wording from a single request and a single idempotency key. The preview is what Claude shows you before asking to publish. Ids are illustrative.

If one version is over its limit, ok is false and the issue names the field to change, for example platform_options.x.text. Claude can edit that string and dry run again without touching the other two. All three networks are live on the hosted server, so this runs as written once each account is connected.

// Claude Code, after reading CHANGELOG.md
mcp__socials__create_post({
  "accounts": ["acc_Li4nKd8Qp2Zr7TmX", "acc_x7Qm2pLr9sTn4VwB", "acc_Bs2rTq8Lm5Wn1KdP"],
  "text": "v2.4 is out. Webhooks now retry for 24 hours, ...",
  "platform_options": {
    "x": { "text": "v2.4: webhooks retry for 24h, signed payloads." },
    "bluesky": { "text": "v2.4 is out: webhooks retry for 24 hours. Notes: https://example.com/changelog" }
  },
  "dry_run": true
})
→ { "object": "preflight", "ok": true, "issues": [],
    "preview": [ { "platform": "linkedin", "text": "v2.4 is out. ..." },
                 { "platform": "x", "text": "v2.4: webhooks retry for 24h, signed payloads." },
                 { "platform": "bluesky", "text": "v2.4 is out: webhooks retry for 24 hours. ..." } ] }

In Claude Code, permission rules decide whether a tool call prompts you. Don't add an allow rule for mcp__socials__create_post, so every publish asks first. Allowing the read-only tools (mcp__socials__list_accounts, mcp__socials__get_post, mcp__socials__list_posts) is reasonable. You can also put the review rule in CLAUDE.md: "Always dry run social posts and wait for my approval." The server-side controls that back this up, from dry runs to per-account caps, are listed under keeping a human in the loop.

Most failures in Claude Code happen before any tool runs. Errors that come back from a tool call, such as preflight_failed or account_needs_reauth, carry a fix field, and the full error list explains each code.

• claude mcp list shows the server as failed or needing authentication: the header didn't arrive. Remove and re-add it with claude mcp remove socials, making sure $SOCIALS_API_KEY is set in that shell.
• The header was empty from the start. Your shell expands $SOCIALS_API_KEY when you run claude mcp add, so if the variable wasn't exported yet, the stored header is Bearer with nothing after it and every connection gets a 401.
• The tools exist in one project but not another: the server was added at the default local scope, which belongs to that project. Re-add it with --scope user to have it everywhere.
• A teammate cloned the repo and has no server: only --scope project writes anything into the repo (.mcp.json), and you shouldn't use it with a raw key. Each person runs the add command with their own key.