@vibeframe/mcp-server

The MCP (Model Context Protocol) server for VibeFrame. This package is only the MCP adapter — it exposes VibeFrame's operations as typed MCP tools so an MCP-capable host can call them by natural language.

VibeFrame is CLI-first, not terminal-only. The CLI is the stable runtime; MCP is the app-host surface for clients that prefer typed tools over shell commands.

Confirmed MCP hosts today: Claude Desktop, Cursor, and Claude Code. Codex can drive vibe natively via shell + AGENTS.md, and can also load a project-scoped MCP server through .codex/config.toml. For other shell-capable hosts, use @vibeframe/cli directly — same operations.

For long-running video builds, use the host's native goal mode as the outer loop: Codex /goal, Claude Code /goal, Cursor's agent loop, or the host equivalent. VibeFrame's MCP tools provide the inner video runtime: JSON-style results, dry-run/cost gates, build-report.json, review-report.json, retryWith, fixOwner, deterministic repair, and render inspection.

Just want a CLI? Use @vibeframe/cli instead — same operations, invoked directly in your shell as vibe <command>. This package and the CLI wrap the same underlying engine; pick whichever fits your workflow. Many users install both.

| Surface | Package | How you call it | | ------------------------------------------------------------------- | -------------------------------- | --------------------------------------------------------------------------------------- | | MCP host (Claude Desktop / Cursor / OpenCode / Claude Code) | @vibeframe/mcp-server (this) | host calls tool by name, for example mcp__vibeframe__build({...}) | | Shell / scripts (any agent host: Codex / Aider / Gemini CLI / etc.) | @vibeframe/cli | vibe init my-video && vibe build my-video && vibe render my-video | | Optional standalone agent REPL | @vibeframe/cli (vibe agent) | natural language -> CLI calls when you do not already use Claude Code/Codex/Cursor/etc. |

Native goal stop rules should be explicit: final MP4 path exists, duration and aspect ratio match the brief, render inspection has no errors, any AI review score meets the goal threshold when AI review is requested, and every fixOwner:"host-agent" issue is fixed, accepted with rationale, or reported as blocked.

The tool list below is what the MCP host sees. The same operations exist as vibe <verb> <noun> subcommands in the CLI — see vibe --help.

Quick Setup

The easiest way to generate the right snippet is:

vibe host setup codex
vibe host setup claude
vibe host setup cursor
vibe host doctor all

vibe host setup prints snippets by default. Add --write to merge the config into the relevant project/app file.

Codex

Add to .codex/config.toml:

[mcp_servers.vibeframe]
command = "npx"
args = ["-y", "@vibeframe/mcp-server"]
enabled = true

Claude Desktop (recommended: extension)

Install the prebuilt Desktop Extension — no Node, npx, or JSON editing needed:

1. Download the extension: vibeframe.mcpb (always latest) — or pick a specific version from the releases page.
2. In Claude Desktop open Settings → Extensions and drag the .mcpb file in (or double-click the file).
3. In the extension's settings pick your Workspace folder (projects are created there; a .env file in that folder is loaded for API keys) and optionally paste provider keys.

Rendering still needs Google Chrome and ffmpeg on the machine. Free local Kokoro narration is bundled inside the extension — no install step; the first use downloads a ~90 MB voice model to ~/.cache/vibeframe/models. To update, download the new release's .mcpb and install it over the old one.

To build the bundle from source: pnpm -F @vibeframe/mcp-server build:mcpb (output in packages/mcp-server/dist-mcpb/).

Claude Desktop (manual JSON config)

Add to ~/Library/Application Support/Claude/claude_desktop_config.json:

{
  "mcpServers": {
    "vibeframe": {
      "command": "bash",
      "args": [
        "-lc",
        "cd '/absolute/path/to/your/vibeframe-workspace' && exec npx -y @vibeframe/mcp-server"
      ]
    }
  }
}

The shell wrapper is important for Claude Desktop because it is a global app config and may not preserve a raw cwd field. It anchors relative project paths so prompts like "create a project named launch" create launch/ under your workspace instead of a temporary directory.

Note: some Claude Desktop builds rewrite this file from memory and can mangle hand-edited entries (the entry comes back as "command": "custom"). The extension install above avoids that entirely and is the recommended path.

Cursor

Add to .cursor/mcp.json:

{
  "mcpServers": {
    "vibeframe": {
      "command": "npx",
      "args": ["-y", "@vibeframe/mcp-server"]
    }
  }
}

OpenCode

Add to .opencode/mcp.json (or your global config per opencode.ai/docs/config):

{
  "mcpServers": {
    "vibeframe": {
      "command": "npx",
      "args": ["-y", "@vibeframe/mcp-server"]
    }
  }
}

Claude Code

Claude Code drives vibe natively via shell + the scaffolded AGENTS.md / CLAUDE.md — MCP isn't required. If you'd like the typed-tool surface anyway:

claude mcp add vibeframe --scope project -- npx -y @vibeframe/mcp-server

Human-in-the-Loop Choices

The server prefers asking over silently picking defaults:

• On hosts that support MCP elicitation (Claude Code 2.1.76+), the build tool opens a native form before the build starts for any choice the call left unspecified: narration provider (Kokoro free/local, OpenAI's gpt-4o-mini-tts, or ElevenLabs), backdrop image generation (skip vs paid OpenAI), and a max cost cap. Declining cancels the build before any provider spend. Set VIBE_MCP_ELICIT=off in the server env to disable the form (headless and automation setups).
• On hosts without elicitation (Claude Desktop today), the server instructions tell the agent to present the same choices in chat and wait for your answer.

What You Can Do

Once connected, your MCP host can resolve prompts like these into typed tool calls:

"Scaffold a 12-second Swiss-Pulse promo project, three beats, and render it" → init + 3× scene_add + render

"Generate a cinematic backdrop image, animate it for 5 seconds, add narration" → generate_image + generate_video + generate_speech

"Remove silent segments and add captions to my interview" → edit_silence_cut + edit_caption

Available Tools

Tool names are MCP-side. Your host typically prefixes them (e.g. Claude shows them as mcp__vibeframe__init). Each one wraps the same engine call as the matching vibe CLI subcommand.

Project flow (top-level)

| Tool | Description | | -------------- | ------------------------------------------------------------------------------- | | project_list | List workspace video projects with build/render status (_archive/ skipped) | | init | Scaffold a video project with STORYBOARD.md + DESIGN.md | | build | Build a storyboard project: narration TTS, image assets, scene HTML composition | | render | Deterministic Hyperframes render → MP4/WebM/MOV |

Scene authoring (lower-level)

| Tool | Description | | ----------------------- | ------------------------------------------------------------------------------- | | scene_list_styles | List the 8 vendored visual identities (Swiss Pulse, Data Drift, …) or fetch one | | scene_add | Append a beat (narration + backdrop + composed HTML) | | scene_install_skill | Install the Hyperframes skill bundle into a scene project | | scene_lint | Validate composition HTML against the visual identity | | scene_compose_prompts | Emit the per-beat compose plan without making an LLM call |

Generation (13)

| Tool | Description | Providers | | ----------------------------------------------- | --------------------------------------------- | ---------------------------------------------------- | | generate_image | Text-to-image | OpenAI, Google, Stability | | generate_background | Cinematic backdrop image (video-tuned prompt) | OpenAI | | generate_video | Text/image-to-video (long-running) | Seedance via fal.ai, Grok, Kling, Runway, Google Veo | | generate_video_status / _cancel / _extend | Manage long-running video jobs | (provider-specific) | | generate_motion | Generate standalone designed motion graphics | Claude or Gemini + Remotion | | generate_speech | Text-to-speech | ElevenLabs | | generate_music | AI background music | Suno, ElevenLabs, Replicate MusicGen | | generate_music_status | Poll Replicate music task | Replicate | | generate_sound_effect | SFX from prompt | ElevenLabs | | generate_thumbnail | AI thumbnail composition | OpenAI, Google | | generate_storyboard | Multi-beat storyboard frames | OpenAI, Google |

Editing (16)

| Tool | Description | | ---------------------------------------- | -------------------------------------------------------------------------------- | | edit_silence_cut | Remove silent segments (FFmpeg or Gemini) | | edit_jump_cut | Remove filler words (Whisper) | | edit_caption / edit_animated_caption | Burn styled / animated captions | | edit_text_overlay | Simple static text burn-in | | edit_motion_overlay | Designed animated overlays or user-provided Lottie overlays | | edit_fade | Fade in/out | | edit_grade | Color grading | | edit_speed_ramp | Variable-speed segments | | edit_reframe | Aspect-ratio reframe (e.g. 16:9 → 9:16) | | edit_interpolate | Frame interpolation / slow-mo | | edit_upscale | AI upscaling | | edit_image | Image editing (gpt-image-2, Gemini) | | edit_noise_reduce | Audio/video denoise | | edit_translate_srt | Translate SRT subtitles | | edit_fill_gaps | Detect & fill missing video segments via TTS narration timing (Plan G — Phase 4) |

Audio (5)

| Tool | Description | | ------------------- | ------------------------------------------- | | audio_dub | AI voice dubbing (ElevenLabs) | | audio_clone_voice | Voice clone from sample | | audio_isolate | Vocal / background isolation | | audio_duck | Auto-duck BGM under speech | | audio_transcribe | Transcript with word-level timing (Whisper) |

Detection (3)

| Tool | Description | | ---------------- | -------------------- | | detect_silence | Find silent segments | | detect_scenes | Find shot boundaries | | detect_beats | Find music beats |

Inspection (4)

| Tool | Description | | ----------------- | ----------------------------------------------------------------------- | | inspect_media | Unified image / video / YouTube analysis (Gemini) | | inspect_video | Temporal video understanding (Gemini) | | inspect_review | AI video review + auto-fix suggestions | | inspect_suggest | Natural-language project edit suggestions (Gemini); optional auto-apply |

Timeline (10)

| Tool | Description | | --------------------------------------------------------- | ----------------------------------------------- | | timeline_create / timeline_info | Create or inspect low-level timeline JSON state | | timeline_add_source | Import media (video/audio/image) | | timeline_add_clip / _split_clip / _trim_clip | Build & shape clips | | timeline_move_clip / _duplicate_clip / _delete_clip | Arrange clips | | timeline_add_track | Add video/audio track | | timeline_add_effect | Apply effect (fade, blur, …) | | timeline_list | List all project contents |

Compatibility & Export

| Tool | Description | | --------------------------------- | -------------------------------------------------------- | | project_create / project_info | Deprecated compatibility aliases for timeline JSON state | | export_video | Export timeline JSON to MP4/WebM/MOV via FFmpeg |

Remix & pipelines (4)

| Tool | Description | | ------------------------ | ------------------------------------------------------------------- | | run | Execute a multi-stage YAML pipeline (vibe run pipeline.yaml) | | remix_highlights | Long-form → highlight clips | | remix_auto_shorts | Long-form → vertical shorts | | remix_regenerate_scene | Re-render a single scene against an existing storyboard.{yaml,json} |

Guides (1)

| Tool | Description | | ------- | ------------------------------------------------------------------------- | | guide | Cross-host guides for motion, scene, pipeline, and architecture workflows |

CLI ↔ MCP sync: packages/mcp-server/src/tools/cli-sync.test.ts is a vitest hook that fails CI when a CLI subcommand is added/removed/renamed without the matching MCP change. Open the test file to see the live mapping table — null rows mark CLI-only commands (e.g. vibe audio list-voices, vibe timeline set) that are intentionally not exposed via MCP.

Resources

| URI | Description | | ------------------------- | ------------------ | | vibe://project/current | Full project state | | vibe://project/clips | All clips | | vibe://project/sources | Media sources | | vibe://project/tracks | Track list | | vibe://project/settings | Project settings |

Prompts

| Prompt | Description | | -------------------- | ------------------------------------- | | edit_video | Natural-language editing instructions | | create_montage | Montage with automatic pacing | | add_transitions | Add transitions between clips | | color_grade | Apply color grading | | generate_subtitles | Subtitles via AI transcription | | create_shorts | Short-form from longer video | | sync_to_music | Cut to music beats |

Environment Variables

API keys are read from the host's environment (~/.zshrc, MCP config env block, etc.). All optional — only set the ones whose providers you use.

| Variable | Used by | | ------------------------------------------------------------ | -------------------------------------------------------------- | | OPENAI_API_KEY | gpt-image-2, Whisper, GPT | | ANTHROPIC_API_KEY | Claude (translate-srt, highlights, build compose pipeline) | | GOOGLE_API_KEY | Gemini (analyze, review, silence-cut, narrate) | | ELEVENLABS_API_KEY | TTS, voice-clone, dubbing, SFX | | XAI_API_KEY | Grok | | FAL_API_KEY | Seedance image-to-video | | RUNWAY_API_SECRET | Runway video | | KLING_API_KEY | Kling video | | IMGBB_API_KEY | Default temporary image host for Seedance/Kling image-to-video | | VIBE_UPLOAD_PROVIDER | imgbb (default) or s3 for temporary image uploads | | VIBE_UPLOAD_S3_BUCKET | S3 bucket when VIBE_UPLOAD_PROVIDER=s3 | | VIBE_UPLOAD_S3_PREFIX | Optional S3 key prefix for temporary image uploads | | VIBE_UPLOAD_TTL_SECONDS | Optional TTL hint for temporary upload URLs | | AWS_ACCESS_KEY_ID / AWS_SECRET_ACCESS_KEY / AWS_REGION | S3 upload host credentials | | VIBE_PROJECT_PATH | Default timeline JSON path for resources |

Requirements

• Node.js 20+
• FFmpeg on PATH (export, editing, pipelines)

Privacy Policy

VibeFrame is local-first: no telemetry, no VibeFrame-operated servers in the data path. Projects and generated media stay in the workspace folder you choose; content is sent only to the AI providers you configure (Anthropic, OpenAI, Google, ElevenLabs, …) when you explicitly invoke them, and API keys live only on your machine. Tools that reach external services are marked with openWorldHint in their MCP annotations; everything else (timeline edits, ffmpeg operations, linting, rendering, local Kokoro TTS) runs fully locally.

Full policy: https://vibeframe.ai/privacy (source: PRIVACY.md).

Installing the .mcpb file directly shows a standard "not verified by Anthropic" sideload notice in Claude Desktop — that is expected for extensions distributed outside the official directory.

License

MIT