ModelRein
How to use

ModelRein: how to use it today

ModelRein (formerly LMCP) runs on a GPU server you control. You reach it from the Rescot desktop app, which connects to that server over SSH, forwards the ModelRein port to your computer and gives the agent the lmcp_ tools in Task mode.

Availability, checked 2026-11-04: there is no pip package (pypi.org/pypi/lmcp returns 404) and no public code repository. The Rescot app does not download or install the ModelRein service for you: it connects to a server that already has the ModelRein service directory (by default ~/lmcp) and can start it there.

1Install Rescot

Rescot is a free Windows 11 desktop agent, also on the Microsoft Store. Download it from learningfind.com. The ModelRein settings live inside the app; nothing else needs to be installed on your computer.

2Open LMCP connections

In Rescot go to Settings › Connections › LMCP connections (the app still labels this section with the former name LMCP). You see two cards:

3Connect to your GPU server over SSH

  1. Paste the SSH login command your cloud provider gives you into SSH connection command, for example ssh -p 37436 root@connect.example.com. The form ssh://user@host:port also works.
  2. Type the server password. Leave it empty to use the keys in your ~/.ssh folder. The password is used for this connection only: it stays in memory and is never written to disk or sent to us.
  3. Optional: under Advanced settings change LMCP port on the server (default 8030).
  4. Press Connect to server. Rescot's built-in SSH client parses the command, logs in, checks the GPU with nvidia-smi and the ModelRein health endpoint on the server, then opens an encrypted tunnel to the first free local port between 8030 and 8035.

The first time you connect, Rescot remembers the server's host-key fingerprint. If the fingerprint changes later, Rescot stops and asks you to confirm before it trusts the new key.

4Start the ModelRein service on the server

If the server row says the ModelRein service is not running, open Start LMCP on the server. Rescot fills in this start command, which you can edit before pressing Run on server:

cd ~/lmcp 2>/dev/null || cd ~; PY=$(command -v python3 || command -v python); nohup "$PY" -m uvicorn server:app --host 127.0.0.1 --port 8030 > ~/lmcp.log 2>&1 & sleep 4; tail -n 20 ~/lmcp.log

Rescot runs it over the SSH connection, shows the last lines of ~/lmcp.log and checks the service again. Server-side requirements: Python 3.10+, at least 8 GB RAM; a GPU is optional and its size depends on the model you load (a 0.5B model runs on CPU, 1.5B fits about 6 GB VRAM, 70B needs 1–2 H100s).

5Use it in Task mode

Turn on Used in Task mode for the server. In Task mode the agent can then call 24 lmcp_ tools, such as lmcp_introspect, lmcp_generate, lmcp_compute_vector_auto, lmcp_head_mask, lmcp_layer_route and lmcp_pipeline. The tunnel forwards to the server, so a quick check from your computer is:

curl http://localhost:8030/healthz   # use the local port Rescot shows for the server

Other ways to connect

Troubleshooting

Password or key rejected

Check the user name in the command (user@host) and the password. Without a password Rescot tries the keys in ~/.ssh.

Ports 8030–8035 are all in use

Close the programs that use these local ports, then connect again.

Connected, but the ModelRein service does not respond

The tunnel is up but nothing answers on the server port. Start the service (step 4) and read the command output Rescot shows.

Not connected after restarting Rescot

The password is never saved, so after a restart the server shows as not connected. Press Reconnect and enter the password again.

Read next

ModelRein · how to use · checked against Rescot a890707 source on 2026-11-04 · home · answers · learningfind.com