Agent Access (MCP) ​

Caddis can act as an MCP server, letting AI agents — Claude Code and other MCP clients — work inside your open project: inspect compositions, build and rewire node graphs, set parameters, add keyframes, and render frames to check their own work while you watch it happen in the viewport.

Agent access is off by default and only ever listens for connections from your own machine.

Enabling it ​

Open Preferences → Agents and turn on Agent Access. Caddis starts a local MCP endpoint at http://127.0.0.1:4923/mcp (the port is configurable). The status line confirms it's running.

Connecting an agent ​

Any MCP client can connect — point it at the endpoint URL (streamable HTTP transport), copyable from the Agents tab:

http://127.0.0.1:4923/mcp

For Claude Code specifically, one command registers it:

sh
claude mcp add --transport http caddis http://127.0.0.1:4923/mcp

Other clients (Cursor, Gemini CLI, Zed, …) take the same URL in their own MCP configuration — add it as a remote/HTTP server named caddis.

Then, with a project open in Caddis, ask for what you want in plain language:

Make a new layer with a lime green rounded rectangle, put a soft glow on it, and animate it sliding in from the left over the first 20 frames.

The agent builds the graph, renders a frame to see the result, and iterates — every change appears live in the editor.

What agents can do ​

  • Read — project overview, layer node graphs with parameters and connections, the full node catalog, and per-node reference documentation (the same reference as this site's Node Reference).
  • Edit — create/delete/reorder layers, add and connect nodes, set parameters (including expose-channel ports), add keyframes, adjust clip ranges, move the playhead, and auto-arrange the node canvas. Compositions too: create, rename, resize, duplicate (for format conversions like 16:9 → 9:16), and switch between them.
  • Import — bring media files into the Assets panel (with the same image- sequence detection as drag-and-drop), organize them into folders, and put assets on the timeline with the right source setup for their type.
  • Draw — author vector art on Editable Shape nodes using SVG path syntax, and edit gradient and curve ramps directly (Gradient, Colorize, Curves, Remap…).
  • Code — write WGSL or Shadertoy-style GLSL for Shader nodes. The code is compiled before it's applied: on errors the agent gets line-mapped diagnostics and self-corrects; on success the node's ports and params are derived from its //! input / //! param pragmas.
  • See — render any frame of the composition as an image, at draft, preview, or full quality.

Staying in control ​

  • Every agent action lands as a single undo step — one Cmd+Z reverses it, exactly like your own edits.
  • The server binds to 127.0.0.1 only and validates request origins; nothing outside your machine can reach it.
  • Turning the toggle off (or quitting Caddis) closes the endpoint immediately.

Limitations (v1) ​

  • Agents edit the frontmost open project; opening/saving project files stays in your hands.
  • Tracking data, per-point shape edits, and styled text runs are readable in summary but not editable by agents yet.
  • One Caddis window at a time.

© 2026 Caddis