LUCI docs
Agents

Agent Bridge

LUCI installs itself into Claude, Cursor, and Codex so they can read your local memory. Skills are the default path; MCP is the fallback.

Agent Bridge is how your coding agents read LUCI Memory. You tick the agents on this machine and LUCI writes the skills into them — there is no config file to hand-edit and no server to point at. MCP is the fallback for agents that cannot load skills.

Official skills: Memories-ai-labs/Luci-skills.

What gets installed

Three skills ship together:

  • luci — search what you previously saw, read, did, or worked on
  • distill — turn each finished day into a daily record in your Life folder
  • summarize-meeting — write a meeting up into the same folder

On the Install and try step, or later in Settings → Agent access, LUCI lists the agents it found on this computer — Claude Code, Cursor, and Codex — and installs into each one's skills folder when you click Install.

For Gemini, GitHub Copilot, Windsurf, and anything else, install from the repo:

Install this agent skill for me: https://github.com/Memories-ai-labs/Luci-skills

Or:

bash
npx skills add Memories-ai-labs/Luci-skills

Full per-agent steps: Install skills.

How the agent reads your memory

The skills do not carry a copy of anything. They teach the agent to call a small local command that LUCI puts on your PATH, and that command talks to the running app.

The agent can:

  • search the exact text that was on screen, or search by meaning
  • search what was said out loud
  • narrow to one app, or to a time range
  • pull one moment in full, including the screenshot
  • reconstruct where time went over a range

The first thing the skill does is check whether LUCI is running. If it is not, the agent falls back to the written records in your Life folder rather than failing — those are plain files and do not need the app.

What to ask

  • "What was I looking at this morning?"
  • "Search my screen memory for the error I saw this morning"
  • "What was I working on yesterday between 2pm and 4pm?"
  • "distill my life"

You do not need special commands. Once the skills are in, the agent already knows when to look something up.

Optional: connect over MCP

If an agent cannot load skills, LUCI also runs an MCP server for it. Copy the ready-made connection from LUCI's Screen Memory panel and paste it into that agent's connections. See Connect over MCP.

It is bound to loopback only — never to a network interface — and every request needs a token generated for your install and held under your operating system's key protection. Requests that arrive claiming to be from a website are rejected, so a page you have open cannot drive it.

Privacy boundary

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

Calls are logged locally so you can see the volume on Insights; the log records what was called and how much came back, not what you searched for.

Per-agent setup

Was this page helpful?