Skip to main content
Endpoint: https://mcp.memories.ai/mcp · Covers: the Video Datalake Works with: Claude, Claude Code, Codex, Cursor — any MCP client Auth: OAuth — sign in to memories.ai in the browser (no key to paste)
The Memories.ai MCP Server is our official Model Context Protocol server for the Video Datalake, hosted at https://mcp.memories.ai/mcp. Point an MCP-compatible agent at it and it can operate the lake — index and search your video, read any moment, manage the face library, and watch for safety events — all in natural language, with no glue code. You don’t build or run this server — we host it. On connect, a browser opens to sign in to your memories.ai account; the server handles auth via OAuth and never sees a pasted key.

​
Prerequisites

  • A memories.ai account (free tier works).
  • An MCP-compatible client: Claude Code, Codex, Claude Desktop, Cursor, or any other.

​
Connect

The easiest path — the Memories plugin bundles the MCP server plus commands:
On first use a browser opens to sign in and authorize. See the Plugin page for Codex and the commands.

​
Verify

Ask your agent:
“List the tools available from Memories.ai.”
You should see the tools below. If nothing appears, jump to Troubleshooting.

​
What your agent can do

An intent-first tool set for the Datalake (not a 1:1 mirror of the REST endpoints): Tools chain naturally — search_moments returns a ref that feeds straight into get_moment; upload_video returns an operation that wait_for_operation blocks on until indexing is done.

​
Example prompts

Start simple, then compose:
  • “List my collections and how many videos each has.”
  • “Upload this video to store-cams and tell me when someone opens a door.”
  • “Search matches for every goal and give me the clips.”
  • “Who appears in vid_abc, and when?”
  • “Watch this RTMP stream and alert me on any safety event.”

​
Troubleshooting

​
Safety

The server enforces the same guardrails as the API:
  • Destructive tools (delete_video, delete_collection) require an explicit confirmation argument — an agent can’t wipe data on a stray instruction.
  • Face-library tools require a compliance token; the server never fabricates consent.
  • Signed media URLs are fetched fresh per request, never cached past expiry.