AutoKeren MCP

AutoKeren MCP — AI Video Studio for Agents

Give any AI agent a complete, autonomous video production studio.

AutoKeren MCP turns Claude Code, opencode, Cursor, Codex, or any MCP client into an end-to-end video producer: write the script, generate images, voice the narration, source real YouTube footage, QA the result with Gemini, and ship a finished 1080p MP4 — all from a single MCP server.

• Endpoint: https://mcp.autokeren.com (Streamable HTTP)
• Auth: self-service API keys — studio.autokeren.com/settings
• Agent skill (one URL, everything): mcp.autokeren.com/skill.md
• Landing: mcp.autokeren.com

Quickstart

Claude Code:

claude mcp add --transport http autokeren https://mcp.autokeren.com \
  --header "Authorization: Bearer ak_YOUR_KEY"

opencode (~/.config/opencode/opencode.json):

{
  "mcp": {
    "autokeren": {
      "type": "remote",
      "url": "https://mcp.autokeren.com",
      "headers": { "Authorization": "Bearer ak_YOUR_KEY" }
    }
  }
}

Cursor / Windsurf (MCP Settings → JSON):

{
  "mcpServers": {
    "autokeren": {
      "url": "https://mcp.autokeren.com",
      "headers": { "Authorization": "Bearer ak_YOUR_KEY" }
    }
  }
}

Any HTTP client:

curl -X POST https://mcp.autokeren.com \
  -H "Authorization: Bearer ak_YOUR_KEY" \
  -H "Content-Type: application/json" \
  -d '{"jsonrpc":"2.0","id":1,"method":"tools/call",
       "params":{"name":"get_workflow","arguments":{}}}'

Then just tell your agent: "Read https://mcp.autokeren.com/skill.md and follow it."

The 54 Tools

Category	What it does
Pipeline (19)	get_workflow, estimate_cost, create_project, load_scenes, generate_assets, render_video, set_scene_media, attach_scene_media_batch (50 scenes/call), verify_render, get_render_errors, get_usage, list_assets (browse your asset library for revisions)... — full production + billing telemetry
Media Generation (7)	generate_image, generate_speech, generate_video_clip (Veo/Seedance), ElevenLabs image/video flows with async polling
Research & Sourcing (6)	youtube_search, yt_dlp_download, index_shots (Gemini watches & timestamps every shot), cut_video (per-scene clips via separate:true), extract_frames, fetch_to_r2
AI Vision & QA (4)	check_vision (free), analyze_media, check_video — full-video audit that classifies real footage vs AI stills per time range, review_video (production score /10)
Packaging & Upload (3)	suggest_packaging (free A/B titles + thumbnail + SEO), get_presigned_upload (bypass 100MB limits), prepare_remotion
Premium Render (4)	render_remotion (one-shot async) + render_remotion_result (poll), or the incremental path: render_remotion_slide (ONE scene, full motion graphics, 0.5 CR, 9:16 Shorts supported) → QA per slide → stitch_remotion_slides (1 CR)

Why agents love it

• Verified uploads everywhere — every media URL the server returns is HEAD-verified in R2 (verified: true); set_scene_media hard-rejects dead URLs (404/410). No more silent AI-image fallbacks.
• Auto-refund on failure — credits never leak on failed generations; the ledger records every refund with a reason.
• Honest QA — check_video reports media_type_report: which time ranges contain real moving footage vs still images with camera motion. Your agent can't fool itself.
• 0-CR free tier — Cloudflare models (script + image) + Aura TTS produce a complete video for zero credits.
• Async premium renders — submit returns poll_id instantly; poll render_remotion_result like a pro.
• Per-slide premium renders — render one Remotion slide at a time (seconds per slide), QA each before committing, stitch at the end. Failure isolation + per-slide QA.
• Attached media is sacred — generate_assets never overwrites media you attached (official photos, sourced footage); it only fills what's missing.
• Asset Library (both ways) — every verified MCP output is auto-indexed into the dashboard library, and list_assets lets agents browse + reuse existing assets for revisions.
• All the formats — create_project({aspect_ratio: "9:16"}) for Shorts (portrait canvas + portrait images), render_video({resolution: "4k"}) for 3840x2160 masters.
• Battle-tested — hardened by live agent audits; every hard-won rule is baked into skill.md.

Pricing snapshot

Stage	Free (0 CR)	Paid
Script	@cf/zai-org/glm-5.3-flash	gemini 0.05–0.1 CR
Image	@cf/black-forest-labs/flux-2-klein-9b	gemini-3-pro-image 8 CR, runway 5–170 CR
TTS	@cf/deepgram/aura-2-en	ElevenLabs 2–10 CR, Gemini TTS 1 CR
Video clip	—	Runway 60–740 CR per 5s (duration-scaled)

Self-service top-up (QRIS/VA for Indonesia, card for international) via topup_credits or the studio. Live prices: list_models.

✍️ Script Generation — AI writes the script

create_script (Gemini) writes + auto-saves scenes. Models (billed per script):

• gemini-4-argon — PREMIUM flagship reasoning, 1M context — 10 CR
• gemini-3-flash-preview — default, fast — 0.1 CR | @cf/qwen, @cf/glm-5.3-flash — FREE

Research feeding: send distilled notes 3-10K chars (facts/numbers/quotes — not raw page dumps), hard cap 50K. Or research: true (default) for auto Google-Search grounding. load_scenes still available for hand-written scripts.

🎨 Motion Graphics Templates — data-driven slides

11 data-driven templates: stat-counter (count-up prices/stats), comparison-chart (racing bars), line-chart, progress-bars, cinematic-title-intro, end-card, subscribe-reminder, gallery-grid (2-6 photos), lower-third, quote-card, glitch-text.

list_templates() → see schemas
render_template_slide({template_id: "stat-counter", data: {value: 19990, prefix: "€", label: "..."}, output_r2key: "..."})
stitch_remotion_slides({slides: [...]})  ← mix template + regular slides in ONE video

Adapted from reactvideoeditor/remotion-templates (free). ~0.5 CR per slide.

🎬 Film Mode (cinematic_film)

Consistent-character short films, end to end:

• generate_character_sheet — multi-angle identity lock (one-time per character)
• extract_last_frame — frame chaining: scene N's last frame → scene N+1's first frame (continuous motion)
• cinematic_film script gen — three-act structure, character bible, shot/lighting language, dialogue
• Full MCP playbook in skill.md (FILM MODE section)

🎙️ Voice Routing — 45 verified voices

Per-language voice catalog; unknown names are rejected (no silent Adam fallback):

Language	Voices
English (US)	Adam, Antoni, Brian, Dave, Callum, Harry, James, Jeremy, Liam, Michael, Sam, Sarah, Rachel, Nicole, Charlie
English (UK)	Daniel, Freya, George, Charlotte, Dorothy, Jessie
Deutsch (native)	Otto, Leon Stern, Leonie, Christian, Lea, Joha
Multilingual (incl. Indonesian)	Clyde, Arnold, Domi, Ethan, Fin, Mimi, Matilda, Grace, Giovanni, Serena, Thomas, Drew, Elli, Emily, Gigi, Glinda, Matthew, Patrick

Persona → Voice auto-mapping: automotive aggressive + German → Leon Stern, nuanced + German → Otto (EN: Callum/George, ID: Arnold/Matilda). No manual voice needed — createScript stamps the persona automatically.

Links

• 🤖 Agent skill: mcp.autokeren.com/skill.md — the complete playbook in one URL
• ⚙️ Get an API key: studio.autokeren.com/settings
• 🎬 Studio (web app): studio.autokeren.com
• 🏠 WebMCP: autokeren.com/mcp — browser agents auto-discover via document.modelContext

Status & reliability

• MCP protocol versions 2024-11-05 / 2025-03-26 / 2025-06-18 (echoed on initialize)
• JSON-RPC notifications per spec (202 + empty body), strict-client compatible (opencode verified)
• Optional-field discipline: structuredContent only when present
• Runs on Cloudflare Workers + Containers (Region: Earth)

---

AutoKeren MCP is a hosted service. This repo is the public documentation, onboarding, and registry home — the live server is operated at mcp.autokeren.com.