Built for agents.

One HTTP endpoint, bearer auth, every SocialKit capability exposed as a tool. Drop the config into any MCP-aware client.

# Add the remote MCP server (recommended):
claude mcp add --transport http socialkit \
  https://mcp.socialkit.sh \
  --header "Authorization: Bearer $SOCIALKIT_KEY"

# Or add it to .mcp.json by hand:
# {
#   "mcpServers": {
#     "socialkit": {
#       "type": "http",
#       "url": "https://mcp.socialkit.sh",
#       "headers": { "Authorization": "Bearer $SOCIALKIT_KEY" }
#     }
#   }
# }

# ~/.cursor/mcp.json
{
  "mcpServers": {
    "socialkit": {
      "type": "http",
      "url": "https://mcp.socialkit.sh",
      "headers": { "Authorization": "Bearer $SOCIALKIT_KEY" }
    }
  }
}

# discovery is keyless: read the catalog before you sign up
curl https://mcp.socialkit.sh/tools/list

# tools: score_post, rewrite_post,
#        validate_post, generate_post, plan_week,
#        build_voice, create_brand,
#        update_brand, delete_brand,
#        create_voice, update_voice,
#        delete_voice, create_post,
#        update_post, delete_post,
#        schedule_post, bulk_schedule,
#        cancel_post, publish_post,
#        refresh_metrics, list_posts,
#        get_post, score_saved_post,
#        get_post_metrics, get_post_scores,
#        get_calibration, get_analytics,
#        get_calendar, list_brands,
#        list_voices, list_channels,
#        get_account

• score_postGrade a post 0-100 with a per-dimension breakdown.
• rewrite_postVoice-preserving rewrite that returns before/after scores.
• validate_postPreflight a draft: structural errors + reach/voice warnings, no LLM call.
• generate_postGenerate 1-3 scored drafts from a brief, ranked best-first.
• plan_weekTurn a brief into a sequenced content plan.
• build_voiceDistill a reusable voice from sample posts; pass its id to generate/plan.
• generate_visualA finished post with rendered image masters, the judge's score, and a revisable spec.
• revise_visualRefine a generated visual post with freeform feedback; unaddressed parts hold verbatim.
• generate_carouselA full swipe deck in one frame family with per-slide roles, rendered and judged.
• revise_carouselRefine a generated deck with freeform feedback, slide by slide or as a set.
• create_brandCreate a brand to ground generation and posts; returns its id.
• update_brandPatch an existing brand; only the fields you pass change.
• create_voiceCreate a voice from explicit traits, do-nots, and exemplars; returns its id.
• update_voicePatch an existing voice; only the fields you pass change.
• create_postSave a draft so it can be scheduled or published; returns its id.
• update_postEdit a draft's text, brand, channel, or media; only the fields you pass change.
• delete_brandDelete a brand by id; future generation can no longer reference it.
• delete_voiceDelete a voice profile by id; future generation can no longer reference it.
• delete_postPermanently delete a post and any pending scheduled job; not mid-publish.
• schedule_postSchedule a saved draft to publish at a future time.
• bulk_scheduleCreate and schedule a whole planned week in one all-or-nothing call.
• cancel_postCancel a scheduled post so it won't publish.
• publish_postPublish a stored draft to its connected channel right now.
• refresh_metricsPoll the platform now and store a fresh engagement snapshot.
• list_postsList posts newest-first, filter by status, paginated.
• get_postFetch a single post by id with its current status and text.
• get_post_metricsRead a published post's engagement: latest snapshot plus history.
• get_post_scoresRead a post's recorded score history with full breakdowns.
• score_saved_postScore a stored post and record it, so it enters the score-vs-reality set.
• get_calibrationRead score-vs-reality: did predicted scores track actual engagement.
• get_analyticsAccount-wide engagement rollup, totals and per-platform.
• get_calendarSee scheduled and published posts within a time window.
• list_brandsDiscover brand ids to ground generation and posts.
• list_voicesDiscover saved voice ids to write in a distilled voice.
• list_channelsDiscover connected channel ids to bind posts to.
• get_accountConfirm which account and plan the API key is authenticated as.

The server also ships prompts, the canonical workflows your client can surface directly. prompts/list returns them; prompts/get fills in your brief.

• draft_and_gradeDraft from a brief, then score and rewrite until it clears a target.
• improve_postGrade an existing post and rewrite the weak dimensions; returns before/after.
• plan_and_draft_weekTurn a brief into a sequenced week, then draft and save each post.

And resources, your account context as attachable, read-only JSON. resources/list enumerates them; resources/read returns the document so your model can ground a draft in your own brand and voice without a tool call.

• socialkit://metaThe service contract: score version, supported platforms, media kinds, and dimension weights. Build a valid call without guessing.
• socialkit://accountIdentity, plan, and what this deployment can do (publish, metrics, channels).
• socialkit://brandsEvery brand: audience, themes, link policy. Ground a draft without discovery.
• socialkit://voicesSaved voices and their traits, so the writing matches a distilled voice.
• socialkit://channelsConnected channels and status, so the model knows what it can publish to.
• socialkit://calibrationScore-vs-reality dataset: which predicted scores actually translated to reach.

The pattern that actually works: let your model draft freely, then use SocialKit as the judge and the rewriter. Three attempts is usually enough to land 80+.

import Anthropic from "@anthropic-ai/sdk";

const claude = new Anthropic();

// SocialKit is plain JSON over HTTPS. No SDK to install.
async function sk(path: string, body: unknown) {
  const res = await fetch(`https://api.socialkit.sh/v1/${path}`, {
    method: "POST",
    headers: {
      "content-type": "application/json",
      authorization: `Bearer ${process.env.SOCIALKIT_KEY}`,
    },
    body: JSON.stringify(body),
  });
  if (!res.ok) throw new Error(`socialkit ${path}: ${res.status}`);
  return res.json();
}

async function draftAndGrade(brief: string) {
  let draft = await claude.messages.create({
    model: "claude-sonnet-4-6",
    max_tokens: 800,
    messages: [{ role: "user", content: `Write a LinkedIn post: ${brief}` }],
  }).then(r => r.content[0].text);

  for (let attempt = 0; attempt < 3; attempt++) {
    const grade = await sk("score", { post: draft });
    if (grade.overall >= 80) return { draft, grade };

    const { rewrite, after } = await sk("rewrite", { post: draft });
    draft = rewrite;
    if (after.overall >= 80) return { draft, grade: after };
  }
  return { draft, grade: await sk("score", { post: draft }) };
}

The same key reaches the visual engine: a brief and a brand kit go in, a finished post comes out with rendered image masters, the design judge's score, and a revisable spec. Your agent asks in plain words; the engine decides, renders, judges, and revises before answering. The full pipeline is on how it works.

# Slow call by API standards (tens of seconds to minutes).
# Always send an Idempotency-Key so a retry can't double-spend.
curl -X POST https://api.socialkit.sh/v1/generate-visual-v2 \
  -H "Authorization: Bearer $SOCIALKIT_KEY" \
  -H "Idempotency-Key: post-2026-07-27-a" \
  -d '{
    "businessContext": { "businessName": "Driftwood Roasters",
                         "brandKit": { ... } },
    "seed": { "platform": "instagram", "idea": "the Tuesday roast" }
  }'

# -> { masters: [{ aspectRatio: "4:5", imageUrl }, ...],
#      caption, hashtags,
#      score: { verdict: "SHIP", score: 95, ... },
#      spec: { ...hand back to /v1/revise-visual-v2 } }

A revision is one more call with the spec and freeform feedback, and it is surgical: the photo stays while the headline moves. Decks work the same way through /v1/generate-carousel-v2. One credit per successful call; a blocked result costs nothing.

MCP-native agents get the same power as tools: generate_visual, revise_visual, generate_carousel, and revise_carousel, annotated as credit-bearing and slow so a well-behaved agent knows what it is spending before it calls. Each artifact has its own page: images, carousels, and video (video powers Taka in production and is not on the public surface yet).

Different jobs get different models, and none of them touch the pixels. The edge isn't a secret ranker. It's what each model is told before it works, and what the code refuses to let them do.

Where this goes next: a calibrated ranker trained on real outcomes, scoring dwell and reshare likelihood directly. That's on the roadmap, not in the box yet. We'd rather tell you what runs today than ship a number we can't stand behind.