FAQ

What is loci?

An intelligence substrate for working with AI. In practice: plain-text firmware, the templates and processes that decide how memory, context, and trust work in your AI setup, regardless of which AI runs it.

Your setup is called a palace: a folder of markdown files you own. Local-first. No cloud, no accounts, no lock-in.

Does loci send data anywhere?

No. A palace is a folder of plain text on your machine. Nothing syncs, nothing phones home, and there are no accounts.

The CLI is read-only and makes no network calls. The desktop app runs on your device.

If you point a cloud-based AI (Claude, ChatGPT) at your palace, that provider processes what it reads, as with anything you give it. loci itself transmits nothing.

How do I run it?

Three ways. The substrate is the same underneath all of them.

• Plain markdown (the door). Clone the repo, copy the templates/ kit into a folder of your own, open FIRST-SESSION.md, and point any file-aware AI at your palace CLAUDE.md. No build.
• Desktop app (Tauri, v0.6.0-beta). The cockpit: a native dashboard plus a tab per instrument your palace generates for itself.
• CLI (loci-cli, Rust, v0.6.0-beta). Five commands: loci status, loci crystals, loci read, loci handover, loci init. Read-only. No network. No inference.

You do not need the desktop app or the CLI to start. The markdown kit is the whole substrate; the other two are expressions built on top of it.

What AI tools are supported?

Any AI that can read files. The methodology is engine-agnostic plain text: point Claude, Claude Code, Cursor, Codex, or a local model at your palace and it runs.

AGENT-SETUP.md in the repo includes wiring notes for Claude Code (a global identity file that stacks with your palace CLAUDE.md), Cowork, and the Claude web interface (Project Instructions). The palace itself is fully portable between them.

Does it work with local models?

Yes, in the way that matters: the substrate is plain text, and any LLM with file access can run it. Nothing in loci assumes a specific provider.

The CLI does no inference at all, local or remote. It reads and prints your palace; your AI does the thinking.

Is there an MCP server?

The backend, yes: the desktop app's Rust side ships a real, compiled JSON-RPC 2.0 server (127.0.0.1:3456, two write tools, three read resources). The button, not yet: the shipped app interface never calls it, so today it only starts by invoking its Tauri commands directly. See the MCP API reference for the full picture.

The CLI has no equivalent (loci serve doesn't exist yet). Tools that just read files (Claude Code, Cursor, any file-aware agent) don't need MCP at all: a palace is plain text, so they load it directly.

What is a crystal?

A crystal is a unit of persistent memory with a tier and a lifecycle. Three tiers: ◇ exploratory, ◈ contextual, ◆ confirmed. You can promote, expire, pin, and compost them, and time-sensitive crystals carry a valid_until date.

The lifecycle doctrine is simple: keep what surprised you, let the rest decay on its tier's own clock, and keep all of it inspectable and prunable. The mechanics live in templates/crystals-guide.md and templates/memory-lifecycle.md.

What is the difference between a room and a project?

A room is a persistent context workspace. A project is what you are building.

You might have one room per project. Or multiple projects in one room. Or one project spanning several rooms (a design room, a dev room, a research room).

Rooms map to how you think, not what you build. If you context-switch between "coding mode" and "writing mode," those are different rooms, even if they serve the same project.

Is loci the same as a RAG system?

No. RAG retrieves text chunks and injects them into prompts. loci is the layer underneath: the firmware that decides what persists, how it is structured (rooms, crystals, the garden, handovers), and what loads when.

It is process and structure, not a retrieval pipeline. And because a palace is plain text, you can point any retrieval tooling at it if you want one.

Is there a web interface?

No. loci is local-first by design. The surfaces are the plain files themselves, the desktop app, and the CLI. Nothing requires an account or a server.

Can teams share a palace?

Yes. The Team is one of the six use-case shapes in the repo's feature map. A palace is a folder of text files, so it can be shared the way any text is shared: a git repository or a shared folder.

The Trust & Governance feature set exists for exactly this: a named review gate, foreign-process quarantine, confirm-against-disk, and a read-only structural audit keep a shared palace honest.

How do I back up my data?

Your palace is a folder of markdown. Back it up like any folder: rsync, Time Machine, git, whatever you use.

If you ran loci init, the CLI keeps its own config at ~/.config/loci/config.toml.

How do I keep my palace up to date?

Ask your AI. Say "Update my palace" or "What's new in loci" and point it at the repo's PALACE-UPDATE.md. It compares your palace against the published methodology version, reports each change with why it matters, and offers the new pieces one at a time.

You decide what enters. Nothing is ever auto-applied.

What happened to the browser extension and the tiers?

Earlier versions of these docs described a browser extension, a conversation index, and three tiers (Scholar, Wizard, LLMAGE). None of that is the shipped product. See the Browser Extension page for the full correction.

The current release is v0.6.0-beta (2026-07-02), the 0.6 open beta: the markdown template kit, the desktop app, and the CLI on one version line for the first time. This page reflects that release; where older pages disagree, trust this one and the repo's CHANGELOG.md.