Which MCP clients does LUCI support?

LUCI ships setup snippets for Claude Code, Codex, and Cursor. Beyond those, any MCP client can connect with the same connection details from the app.

Is the local connection exposed to my network or the internet?

No. Only processes on your own computer can reach it, and every request has to be authorized with the credentials shown in the app. There is no cloud endpoint.

Does connecting an agent upload my memory anywhere?

The memory itself stays on your machine. What an agent actually pulls becomes part of that conversation. If the agent runs in someone else’s cloud, that conversation goes there. Connect only the agents you trust, and ask them for what you are willing to put in that thread.

What data do the tools return?

Text and metadata: matching text blocks or transcript segments with their app, window, source, and timestamps, plus per-capture detail on request. Agents query slices of your memory; nothing is bulk-exported.

Protocol

LUCI Is a Local MCP Memory Server

Skills are the recommended way to connect a coding agent. MCP is the fallback: an open protocol so Claude Code, Codex, Cursor, or any MCP client can question your memory — while the memory itself stays on your machine.

See all integrations

Most "memory for agents" products are cloud services with an API key. LUCI inverts that: the memory is built and stored on your own machine. Official skills teach the agent when to look it up. MCP is the optional doorway if the agent cannot load those skills.

Install the official skills first: Memories-ai-labs/Luci-skills. Use this page only if you need the protocol path. LUCI shows ready-made MCP snippets for Claude Code, Codex, and Cursor; any other MCP client uses the same URL and token.

What the protocol path gives you

Standard protocol, no lock-in

MCP is an open protocol. The memory you accumulate serves every agent you use today and whatever you switch to next year — it is not scoped to one vendor’s assistant.

Private by design

The connection serves your machine, and only your machine. There is no public endpoint, no cloud relay, and no credential that ever leaves your device.

Full coverage

Exact strings, fuzzy recall by meaning, scoping by app and time, a whole range at once, what was said out loud, and any single moment in full detail — enough for an agent to find anything you saw or heard.

The memory stays put

Queries come to the data; the data never migrates to a vendor. Storage is local and encrypted at rest, and retention is governed by your disk, not a plan tier.

Try asking

Questions that work on day one

Search my screen memory for the invoice number I saw yesterday
Semantic-search my memory for anything about GPU pricing
Aggregate what I did between 9am and 11am today
Search my meeting audio for where we agreed on the deadline

Set up MCP

01
Prefer skills when you can
If your agent can load skills, install those instead from https://github.com/Memories-ai-labs/Luci-skills. That is the recommended path.
02
Turn on Screen Memory in LUCI
Install LUCI on your Mac or Windows machine and enable Screen Memory. Leave the app running. The local connection is ready as soon as the app is.
03
Copy your connection details
LUCI’s Screen Memory panel shows the URL, token, and per-client snippets for Claude Code, Codex, and Cursor.

Point your client at it

Paste the snippet for your MCP client using the details from the panel.

# Claude Code
claude mcp add --transport http luci <url> \
  --header "Authorization: Bearer <token>"

# Codex — ~/.codex/config.toml
[mcp_servers.luci]
url = "<url>"
http_headers = { "Authorization" = "Bearer <token>" }

# Cursor — ~/.cursor/mcp.json
{ "mcpServers": { "luci": {
    "url": "<url>",
    "headers": { "Authorization": "Bearer <token>" } } } }

What the connection covers

Exact-match recall: Find exact words, names, error messages, or code you remember seeing, with the surrounding app and time context.
Deep text search: Recall screen content by meaning, even when you do not remember the exact words.
App & time scoping: Narrow your memory to the apps and periods that matter for the question at hand.
Time-range digest: Get a concise account of what you were doing during a chosen period.
Spoken-word recall: Recall what was said in meetings, calls, or videos recorded on your device.
Full-context drill-down: Open a remembered moment with the context you need to understand it fully.

Frequently asked questions

Skills. They teach the agent when and how to look you up. MCP is the fallback if the agent cannot load skills. Official repo: Memories-ai-labs/Luci-skills.