🔌MCP integration

The Model Context Protocol (MCP) is an open standard that lets AI assistants call tools in other apps. Swamp Dragon includes an MCP server, so the assistant you already use — Claude Code, Codex, Cursor or any other MCP host — can read your timeline and edit it with real editing tools: cut clips, add zooms, animate titles, place motion graphics, apply skills and more.

#How it works

1. Your assistant starts Swamp Dragon's MCP server

   A small program bundled with the app. It contains no editing logic and no AI — it forwards each tool call to the running app.

2. The app checks access and applies the edit

   Only projects you've shared are visible. Edits go through the same edit operations as the editor, so they follow the same rules and limits.

3. You see it live

   If the project is open in the editor, the edit appears immediately, a note in the inspector says which assistant did what, and ⌘Z undoes it. If it isn't open, the project file is updated directly.

Everything stays on your computer. The server reaches the app through a local connection on your computer using a private token that's regenerated every time the integration starts.

#Requirements

You also need an MCP-capable assistant installed on the same computer, and Swamp Dragon must be open while the assistant works.

#Set up from the editor

1. Install your assistant’s connector once

   Get the Claude Desktop / Claude Code bundle, Gemini Desktop / CLI package or ChatGPT Desktop / Codex plugin from the connector downloads. Other local MCP hosts can use the configuration in the editor.

2. Open a project and click Connect AI

   Choose your assistant and click Connect this project. The app enables MCP and creates a private alphanumeric code for this project.

3. Copy the connection message into your assistant

   The assistant calls connect_project with your code. It can list, read and edit only that project, even if other projects are shared in Integrations. The editor shows when the assistant uses the connection.

4. Disconnect or regenerate in place

   Use the same editor dialog to revoke the code or create a replacement. Codes expire after 12 hours or an app restart. Install the connector only once; pair again when switching projects.

In ChatGPT Desktop, you can either upload swamp-dragon-chatgpt.zip under Customize → Plugins → Add → New Plugin (and add the local MCP command under Add custom MCP server), or extract the ZIP and run Install-ChatGPT.command (Install-ChatGPT.cmd on Windows) to install both the local plugin and the MCP server automatically. In Gemini Desktop, open Connectors → + Custom → Set up a custom connected app → Local MCP, enter Swamp Dragon as the server name, and paste node "/Users/you/.local/share/swamp-dragon/launcher.cjs" as the command (running Install-Gemini.command installs the launcher and copies this command to your clipboard).

#Manual host configuration

Copy your machine’s exact command, arguments and environment from the editor’s Local MCP configuration. Integrations also retains the general host setup snippets. A project code is always required before tools can read or edit.

claude mcp add swamp-dragon \
  -e ELECTRON_RUN_AS_NODE=1 \
  -e SWAMP_DRAGON_ENDPOINT="/Users/you/Library/Application Support/Swamp Dragon/mcp-endpoint.json" \
  -- "/Applications/Swamp Dragon.app/Contents/MacOS/Swamp Dragon" "/Applications/Swamp Dragon.app/Contents/Resources/app.asar/out/main/mcp.js"

{
  "mcpServers": {
    "swamp-dragon": {
      "command": "/Applications/Swamp Dragon.app/Contents/MacOS/Swamp Dragon",
      "args": ["/Applications/Swamp Dragon.app/Contents/Resources/app.asar/out/main/mcp.js"],
      "env": {
        "ELECTRON_RUN_AS_NODE": "1",
        "SWAMP_DRAGON_ENDPOINT": "/Users/you/Library/Application Support/Swamp Dragon/mcp-endpoint.json"
      }
    }
  }
}

[mcp_servers.swamp-dragon]
command = "/Applications/Swamp Dragon.app/Contents/MacOS/Swamp Dragon"
args = ["/Applications/Swamp Dragon.app/Contents/Resources/app.asar/out/main/mcp.js"]
env = { ELECTRON_RUN_AS_NODE = "1", SWAMP_DRAGON_ENDPOINT = "/Users/you/Library/Application Support/Swamp Dragon/mcp-endpoint.json" }

#Your first edit

Open a project in the editor (so you can watch), then ask your assistant something like:

Use the connection code I copied from Swamp Dragon to call connect_project.
Read the paired project’s timeline, then:
1. remove the first 4 seconds,
2. add a "Step 1 — Create your workspace" title that comes in word by word,
3. zoom 1.8× on the Invite button around 0:42 with an overshoot ease,
4. put a spotlight and an arrow on it while the zoom holds,
5. move the webcam to a bottom-right circle with a white border.

After pairing with connect_project, the assistant calls list_projects and get_timeline to understand the video, then trim_clip, add_text, add_zoom, add_graphic and set_webcam_layout. Each change shows up on the timeline, and the Activity list in Integrations logs every call.

#Prompts that work well

• "Add chapter markers where I change topics and turn them into chapter cards."
• "Make the intro kinetic: big centered title, bounce in, float loop, then a zoom-in transition into the demo."
• "Blur the email address visible from 1:10 to 1:25." Unlimited
• "Speed up the part where the page loads to 3× and add a freeze frame on the final result."
• "Give the whole video a crisp look, a gradient frame with 6% padding and no shadow."
• "Trim silences longer than one second, then add zooms to every click." Unlimited

#Privacy and safety

• You choose what's visible. Each code grants access to one project only. Project ids for other projects are refused, even if they are shared.
• Edits, not files. Tools change the edit (the timeline), never your original recordings, and never read or write arbitrary files.
• Local only. The bridge accepts connections only from your computer, using a random port and a private token stored in mcp-endpoint.json, which is deleted when you turn MCP off. Other computers can't reach it.
• What the assistant sees. Timeline data (times, text, settings) is returned to your assistant, which may send it to its provider under that provider's terms. Video and audio frames are never sent.
• Undo. With the project open, every assistant edit is one undo step labelled with the tool that made it.
• Turn it off any time. Switch off MCP integration; the bridge stops immediately.

#Tool reference

Times are in timeline seconds (after cuts). Positions are fractions of the frame from 0 to 1, with 0,0 at the top left. Tools that add something return its new id as created_id; every editing tool returns the updated timeline.

#Reading

#Cuts and clips

#Zoom and webcam

#Text and motion graphics

Animation fields (text and graphics): animation_in (none, fade, slide-up, slide-down, slide-left, slide-right, pop, bounce, zoom, spin, blur, typewriter, words, wipe, draw), animation_out (none, fade, slide-*, pop, zoom, spin, blur, wipe, draw), animation_duration (0.05–3 s), loop (none, pulse, breathe, float, wiggle, spin, blink, shake), loop_speed (0.1–5 per second).

#Look and effects

#Full integration only Unlimited

#Activity log

Integrations → Activity lists the latest tool calls: which assistant called which tool, with what arguments, whether it worked, and the error if not. In the editor, the inspector shows the last assistant action with a reminder of how to undo it.

#Troubleshooting