Let your own AI agent edit your videos

Note: Using a cloud bot such as ChatGPT, Grok or Meta AI, or no install at all? Use video by link and the remote MCP instead: the bot writes a link and the video is made in your browser.

1. Add the bridge to your agent (MCP)

Needs Node.js 18+. The bridge listens on 127.0.0.1 only. The npm package indiadigital-video-mcp goes live with its first release — until then the commands below will report that the package is not found.

Claude Code

claude mcp add indiadigital-video -- npx -y indiadigital-video-mcp

Claude Desktop

Settings → Developer → Edit config (claude_desktop_config.json):

{
  "mcpServers": {
    "indiadigital-video": {
      "command": "npx",
      "args": [
        "-y",
        "indiadigital-video-mcp"
      ]
    }
  }
}

Cursor (.cursor/mcp.json) and Windsurf

{
  "mcpServers": {
    "indiadigital-video": {
      "command": "npx",
      "args": [
        "-y",
        "indiadigital-video-mcp"
      ]
    }
  }
}

VS Code (.vscode/mcp.json)

{
  "servers": {
    "indiadigital-video": {
      "type": "stdio",
      "command": "npx",
      "args": [
        "-y",
        "indiadigital-video-mcp"
      ]
    }
  }
}

Gemini CLI (~/.gemini/settings.json)

{
  "mcpServers": {
    "indiadigital-video": {
      "command": "npx",
      "args": [
        "-y",
        "indiadigital-video-mcp"
      ]
    }
  }
}

2. Connect the editor tab

1. Open the video editor → Director → Agents → Connect.
2. The tab shows a one-time pairing code such as K7Q-2MX.
3. Tell your agent: “Pair the IndiaDigital editor with code K7Q-2MX.” It calls pair_editor. Next time the tab reconnects automatically; Forget pairing revokes it.
4. Ask for the video: “Import the clips in ~/Videos/Before and ~/Videos/After and make a 25-second before/after reel for Aether Spaces.”

Every change the agent makes appears live on the timeline and is a single undo step. Imports and exports ask you first, in the tab.

3. Or chat with an agent inside the editor (ACP)

With the Agent Client Protocol the bridge starts your agent and you chat with it in the Director panel. Tool calls, plans and permission prompts show up there.

npx indiadigital-video-mcp serve --agent claude --agent gemini --agent codex

• Claude Code: uses @agentclientprotocol/claude-agent-acp (fetched by npx); sign in once with claude.
• Gemini CLI: install @google/gemini-cli and sign in; the bridge runs gemini --acp.
• Codex: uses @agentclientprotocol/codex-acp; sign in with Codex first.
• Any other ACP agent: --agent-cmd myagent="node ./agent.mjs". Commands are set on your computer only — the website can only pick one by name.

What the agent can do

• pair_editor: Pair with the editor tab using the code it shows.
• import_media: Add files from your computer (you confirm in the tab).
• get_project / get_brief / set_brief: Read the project and the creative brief.
• set_asset: Fix sideways clips and tag before/after, logo, music.
• analyse_media: Shots, motion, brightness, colour, sharpness, shake and best moments — with optional local CLIP / Florence-2 vision.
• get_shot_list / get_contact_sheet: Shot table and contact-sheet images so the agent can see the footage.
• list_styles / apply_style_template: 20 edit recipes: before/after, product launch, podcast clips, wedding, ads, explainers…
• get_edit_plan / apply_edit_plan / refine_plan: Read and write the full edit plan: shots, motion, grades, transitions, kinetic text, end card, music, beat-sync.
• render_preview_sheet: A contact sheet of the rendered output for self-review.
• export: MP4 or WebM, optionally saved straight to your Downloads or Videos folder (or a folder you allow with --save-dir). Never into system, settings or startup folders; an existing file is only replaced if you choose Replace. The result says whether the full GPU renderer was used or a fallback that may miss effects.
• transcribe: Speech to text with Whisper on your device (word timestamps, long audio in chunks with progress).

Plans follow the edit plan JSON schema (also an MCP resource). The bridge includes adirect_video prompt that walks the agent through analyse → plan → preview sheet → critique → revise → export.

Stock and AI media

• search_openverse / add_openverse_to_timeline: Creative Commons music (Jamendo), sound effects (Freesound) and images (Flickr, Wikimedia, rawpixel, NASA) from Openverse, no key. Commercial-safe licences only; music lands on a ducked Music track that the director cuts to. Title, author, source and licence are kept for credits.
• search_stock / add_stock_to_timeline: Free Pixabay photos, illustrations and clips with your free Pixabay key; credits are kept (get_stock_credits). Agents get 12 downloads per 10 minutes, then you are asked.
• stock_audio_links: Pixabay Music and sound-effect searches to download yourself; the editor imports them with credits.
• list_generators / estimate_generation: Your AI generators (local ComfyUI or your own keys) and the ₹ estimate before anything runs.
• generate_image / generate_video / generate_broll_for_shot: Images and clips from text or from a photo, or B-roll for a shot. Every paid call asks you first with the ₹ figure; agents cannot approve spending.
• prompt_to_video: A shot list, generated shots, music and edit from one prompt, with one ₹ confirmation.
• list_gen_workflows / run_gen_workflow: Multi-step generation pipelines (prompt → image → video → upscale → timeline, or variations per shot), confirmed once with the whole-run ₹ cost.
• improve_prompt / get_gen_history: Rewrite a prompt for a specific model (nothing is generated); list recent generations with seed, cost and result.
• mark_ai_generated / run_media_steps: Label AI files made elsewhere; run recipe media steps.

Generators: ComfyUI on your PC (free); “Use my ChatGPT subscription” — your own official Codex CLI run by the local relay on your PC (node indiadigital-relay.mjs --codex), billed to your ChatGPT plan limits; your own keys for fal.ai, kie.ai, Together or Google Gemini; Replicate or Higgsfield through the local relay. Paid calls always ask you first in the editor tab. Keys stay in your browser. Guide: stock and AI media.

Voiceovers and dubbing

The same connection drives the Voiceover Studio (the editor opens its Voice tab when an agent needs it):

• voice_list_voices: Voices by language and tier: on your device (free), your own local server, or a provider with your key.
• voice_get_script / voice_set_script: Read or write the voiceover script (emphasis, pauses, scene anchors).
• voice_generate: Voice the script and place it on a Voiceover track with captions; music ducks under it.
• voice_regenerate_sentence: Re-voice one line and swap it in place.
• voice_add_pronunciations: Teach it how to say names, brands and abbreviations.
• voice_dub: Dub into another language. Your agent translates the transcript lines itself and passes them in.
• voice_status: What the Voiceover Studio is doing.

• Agents cannot create or enrol voices. Only you can, with your microphone and a spoken consent sentence.
• Using "My voice" always asks you first, in the editor tab, even when the agent is connected through the bridge.
• For translation and script drafts the editor asks, in order: your connected agent (ACP), then Ollama or LM Studio on your computer, then a provider with your own key. There is no good on-device translator for Indian languages yet, so a connected agent is the best option.

What leaves your device

• Your video files never leave your computer. They are read by the bridge (only when you approve an import) and decoded in your browser.
• Your agent receives what the tools return: project facts, shot statistics, contact-sheet images (small keyframe grids) and transcripts if you made one. Where the agent sends that is governed by your agent's provider (e.g. Anthropic, Google, OpenAI) under your account.
• IndiaDigital's servers receive nothing: no keys, no footage, no prompts.

Security

• The bridge binds to 127.0.0.1, checks the Host header and only accepts the editor from https://indiadigital.org or local development origins.
• The tab only obeys a bridge that presents the one-time code it displays; pairing tokens are stored in ~/.indiadigital-video/ and in your browser, and can be forgotten from the tab.
• ACP agent commands come from your command line, never from the web page.

Troubleshooting

• “No bridge found”: make sure your MCP client started it (or run npx indiadigital-video-mcp serve) and that the port matches (default 47821).
• Chrome may ask to allow “local network access” the first time — allow it for indiadigital.org.
• Safari can block connections from https pages to local servers; use Chrome, Edge or Firefox for agent control.