Self-hosted memory, over MCP.
multirys is an open-source memory server any AI tool can share over the Model Context Protocol. It returns ranked snippets instead of whole documents, runs a small model on-device, and reports the tokens it avoided on every call. Run it as a Mac app, a CLI, or a Docker server for your team.
Three ways to run it
The same engine ships as a one-click desktop app, a CLI, and a Docker image. All self-hosted, all open source.
Desktop app
One binary that is both the GUI and the MCP server. Claude launches it on demand — no separate setup. Build the macOS app from source (needs Rust + the Tauri CLI):
# build the .dmg from source
cargo install tauri-cli --version "^2.0"
cd app/desktop/src-tauri
cargo tauri build
# -> target/release/bundle/dmg/multirys_*.dmg
CLI
Mirror a folder, register multirys with your AI clients, and run the MCP server — all from the command line.
cargo build --release
# mirror a folder of docs
./target/release/multirys sync \
~/company-docs --name docs
# register with Claude Desktop + Code
./target/release/multirys connect
# run the MCP server
./target/release/multirys mcp \
--client claude-desktop
Docker (team)
A shared team server with an admin UI and token-savings dashboard on port 8090.
cd app
MULTIRYS_TOKEN=pick-a-secret \
docker compose up --build
# connect a client
claude mcp add --transport http \
multirys \
http://your-host:8090/mcp \
--header "Authorization: Bearer \
pick-a-secret"
Without MULTIRYS_TOKEN the server runs open — fine on localhost, not on a shared network. The desktop app is unsigned for now, so on first open: right-click the app → Open.
Works with every MCP‑compatible tool
multirys speaks the Model Context Protocol. If your AI tool supports MCP, it works with multirys—no plugins or adapters.
Claude
Cursor
Copilot
ChatGPT
Hermes
OpenClaw
Manus
Windsurf
Local stdio clients (Claude, Claude Code, Cursor) attach directly. HTTPS-only clients (ChatGPT Dev Mode, Copilot/M365) connect to the hosted/Docker server.
The saving is measured, not claimed
Everyone in this category markets a savings percentage. multirys is the only one that shows the model and the user their own, live.
Every call records what a naive full-chunk RAG load would have cost, minus what multirys actually returned:
savings = naive full‑chunk RAG cost − tokens actually returned
A per-client savings_report MCP tool returns this client's call count, tokens returned vs. avoided, and the percentage reduction—with the baseline stated in-band. The Docker server aggregates the same data into an admin dashboard.
How it works
A lifecycle that forgets and supersedes, so it injects fewer tokens—and a local model doing the grunt work.
On-device model (MR0)
A small model runs entirely on your Mac to extract, dedup, rerank and structure your context. No API key, no BYO Ollama—nothing leaves your machine.
Memory lifecycle
Extraction, dedup, supersession, decay and provenance on facts. Old context is forgotten and replaced—the cause of the lower token footprint.
Human-in-the-loop
Folder-level permissions per assistant. Every memory the AI writes waits in a review queue until a human approves it.
Reads and writes notes
Mirror a folder or your Obsidian vault; let your assistant search it and file new notes back in the right place—wikilinks and frontmatter intact.
Explore the graph
An interactive graph of your notes and how they link, right in the desktop app.
Local-first & portable
Runs on your infrastructure. The memory you own, that spans every AI you use—and that the vendors won't let talk to each other.
Open source, BSL 1.1
Free to self-host. A Rust workspace plus a Tauri desktop app.
app/ Rust workspace
crates/
multirys-core Storage, search, ACLs, memory queue, note placement, graph
multirys-connectors Connector trait + filesystem connector + vault writer
multirys-clients Register multirys with Claude Desktop / Claude Code
multirys-llm UtilityModel trait (local / cheap-cloud routing)
multirys-mcp MCP protocol + token-budgeted toolbox (search, memory, notes)
multirys-server Team server: HTTP MCP, admin API, embedded UI
multirys-cli `multirys` binary
desktop/ Tauri 2 desktop app (GUI + bundled MCP server)
See the README for the full build & CLI reference, and ARCHITECTURE.md for the design decisions. Licensed under the Business Source License 1.1.