Quickstart

You need an Eddie account, and you need a project to work on. Import your footage in the Eddie web application first. Every tool call names a project, and a project with no media has nothing to answer with.

Server name

Eddie AI

Server URL

https://mcp.heyeddie.ai/api/mcp

Client id / secret

Leave blank. The client registers itself during the sign-in.

Token

None to paste. Your client obtains and refreshes it.

1. Open Claude settings and go to connectors.
2. Add a custom connector.
3. Enter the name Eddie AI and the URL https://mcp.heyeddie.ai/api/mcp. Leave the OAuth client id and client secret fields empty.
4. Save, then connect. Claude opens the Eddie sign-in the first time.
5. Start a chat and ask for something concrete: “list the sources in this project”, with your share link.

The install page carries the same steps with links to a walkthrough that has screenshots.

1. Open ChatGPT settings and find connectors. The desktop application is the easier place to do this.
2. Add a custom MCP server.
3. Enter the name Eddie AI and the URL https://mcp.heyeddie.ai/api/mcp. Leave the OAuth client id and client secret fields empty.
4. Save, then connect and sign in when prompted.

One command adds the server to the current project:

claude mcp add --transport http eddie https://mcp.heyeddie.ai/api/mcp

Add --scope project to write the server into a .mcp.json file your repository can share. The file it writes looks like this:

{
  "mcpServers": {
    "eddie": {
      "type": "http",
      "url": "https://mcp.heyeddie.ai/api/mcp"
    }
  }
}

Run /mcp inside Claude Code to sign in and to check that the server reports as connected.

The server speaks the Streamable HTTP transport. Point your client at the endpoint and let it run the OAuth flow. Two settings matter:

• Transport: http, sometimes labelled “streamable HTTP”. Not stdio, and not the older SSE transport.
• Authentication: OAuth with discovery. Your client reads the challenge on the first unauthorised call and takes it from there.

You can confirm the endpoint is reachable, and see the challenge, with one request:

curl -i -X POST https://mcp.heyeddie.ai/api/mcp \
  -H 'Content-Type: application/json' \
  -H 'Accept: application/json, text/event-stream' \
  -d '{"jsonrpc":"2.0","id":1,"method":"initialize","params":{}}'

A healthy server answers 401 with a WWW-Authenticate header naming the metadata document. That is the response your client is meant to act on. Authentication walks the rest of the chain.

Start with a read. It proves the connection, the sign-in and the project reference in one step, and it changes nothing.

List the sources in https://app.heyeddie.ai/share/YOUR_SHARE_ID

The project reference accepts either form: the bare share id, or the whole share URL. Copy the link from the project in the Eddie web application.

When that answers, move on to a build. Say what you want and say how long it should be — the length is an instruction, not a detail: “build a six-minute rough cut from this project”.