Every other AI client and harness
Prism is a standard remote MCP server, so the model never decides this: the client does. Whichever you pick, it stays yours: your subscription, your history, your interface. Prism takes nothing for the model that drives it and never asks for an API key. Anything below takes the same address:
https://live.oneprism.io/after-effectsWhat a client has to support
| Requirement | Why |
|---|---|
| Remote MCP servers | Prism is reached at an address, not launched from your disk |
| Streamable HTTP | Keeps the connection open |
| Browser sign-in | Prism issues no keys, so a fixed header has nothing to carry |
| Tool calling | The AI invokes tools, not just text |
Popular harnesses
| Client | Connects | How you add Prism |
|---|---|---|
| OpenCode | Yes | Detects the 401 and runs OAuth itself, or opencode mcp auth prism |
| Cline | Yes | MCP Servers ▸ Remote, then approve in the browser |
| Cursor | Yes | mcp.json, a url entry |
| Windsurf | Yes | mcp_config.json, a serverUrl entry |
| VS Code · Copilot | Yes | mcp.json in the workspace, type http |
| Zed | Yes | Context servers, remote |
| Qwen Code | Yes | qwen mcp add, using httpUrl |
| Kimi Code | Yes | kimi mcp add --auth oauth |
| Codex | Yes | Its own MCP config, same address as ChatGPT |
| DeepSeek Harness | Partly | Needs an OAuth plugin |
| Antigravity | Yes | serverUrl, not httpUrl |
| Any chat app with no MCP client | No | Nothing to add a server to |
Running an open-weight model
Prism never sees which model is driving. Qwen, DeepSeek, Kimi, Llama or anything you host yourself all work identically, because the client speaks MCP and the model just calls the tools it is offered.
The generic setup
Find where your client adds one. The field is called url, serverUrl or httpUrl depending on the client, and the transport is called Streamable HTTP, http, or just Remote.
{
"mcpServers": {
"prism": {
"url": "https://live.oneprism.io/after-effects"
}
}
}opencode mcp auth prism or kimi mcp auth prism.If the client was already open, quit it and open it again
Save the address, then run the client's sign-in step. If your AI client was open before the address existed and does not show Prism, quit it and open it again: most clients load their connections at launch. Then open a composition and ask for it by name.
Common questions
No. The tools, the allowance and the results are identical whichever model drives them. Stronger models plan better, which shows up as fewer wasted calls rather than different capability.
Yes. Run it in any client that speaks MCP and signs in through a browser. Prism authenticates the person, never the model, so where the weights live is not something it can see.
Yes. Add the same address to as many as you like. Each asks you to approve it once, and you can switch freely afterwards.
Almost always one of two things: the remote transport was not selected, so the client is trying to launch a program that does not exist, or the browser approval was never completed. Check both, then see troubleshooting.
Related
DeepSeek
Connect DeepSeek Harness to your live After Effects project. Install an OAuth MCP plugin, sign in through the browser, and get real editable keyframes back.
First prompt
Example Prism prompts for After Effects and exactly what comes back: eased keyframes, new comps, SVG import, subtitles and beat markers.