electric-dreams/computer (fork of dirty-data/computer)

Fork of the LM Studio plugin dirty-data/computer (itself an extended version of khtsly/lms-plugin-computer, MIT). It gives the local model its own isolated Linux container. Changes in this fork:

Tool call budget per request. The budget used to be module-global and reset only by the prompt preprocessor, which LM Studio does not run for API calls (POST /api/v1/chat). Over the API it never reset: after 10 calls every tool answered "budget exhausted" and the model looped. The budget now lives in each tools-provider session (one per prediction), with the configured maximum.
Global settings. All settings are global plugin settings instead of per-chat settings. The container is one shared resource, and API calls used to get the defaults (10 tool calls, no shared folder if the container had to be recreated).
Custom image. New setting "Custom Image (overrides Base Image)": any local or registry image (it must provide bash). An image that already exists locally is never pulled, so a locally built image cannot be replaced by a registry image of the same name.

Install

From LM Studio Hub: (or ).

Settings (global, in the plugin page of LM Studio)

They apply to every chat and to API calls. The most useful ones:

Setting	Notes
Custom Image	your own image, e.g. `local/my-sandbox:1` built with `docker build`; it must provide `bash`; leave empty to use Base Image
Shared Folder	host folder mounted on `/mnt/shared` (the only data that survives a rebuild)
Internet Access	off by default; changing it recreates the container (persistent mode keeps its files)
Persistence Mode	persistent keeps the container between LM Studio restarts; ephemeral recreates it at each start
Max tool calls per turn	per request (chat message or API call)

After changing Custom Image, rebuild the container (ask the model to "rebuild the computer", or docker rm -f lms-computer-main) so it is recreated from the new image.

Develop and publish

lms-plugin-computer

Give your local model its own Linux CLI computer. It gets a real isolated container it can run shell commands in, write and read files, install packages, manage processes, and run background tasks - all without touching your host system.

This plugin is still in beta version, it may be not that stable. This plugin is highly-inspired by the Claude's computer.

Before you start

You'll need a container runtime. Install one based on your OS and make sure it's running before using the plugin.

Windows & macOS - Docker Desktop is the easiest option. Just install it and keep it open in the background.

Linux - Podman works great and doesn't require root. After installing, run these once:

Then log out and back in. The plugin handles everything else automatically.

Installation

The easiest way:

Or from source:

Getting started

Open LM Studio and start a new chat with any tool-capable model
Enable the khtsly/computer plugin from the plugin panel
Say something like "do you have a computer?" or just give it a task

On first use it pulls the Ubuntu image and sets up the container - this takes a minute. After that it's instant.

Settings

Click the gear icon next to the plugin to open settings.

The basics

Internet Access - off by default. Turn this on if you want the model to install packages with apt-get, pip, npm, etc. If you change this after the container is already running, just tell the model "rebuild the computer" and it will recreate with the new setting.

Persistence Mode - persistent by default, meaning files and installed packages survive across sessions. Switch to ephemeral if you want a clean slate every time LM Studio opens.

Base Image - Ubuntu 24.04 is recommended. Alpine is available if you want something ultra-lightweight, though some packages behave differently with its musl libc.

Resource limits

Setting	Default	Range
CPU cores	2	0–8 (0 = unlimited)
Memory	1024 MB	256–8192 MB
Disk	4096 MB	512–32768 MB

Other options

Auto-Install Packages - choose what gets pre-installed when the container is first created. Options range from nothing to a full set including Python, Node.js, build tools, and networking utilities. Requires Internet Access to be on.

Shared Folder - mount a folder from your computer into the container at /mnt/shared. Useful for giving the model access to your files without going through the upload tool every time.

Port Forwards - expose ports from the container to your host. For example 8080:80 lets you open localhost:8080 in your browser to see whatever the model is running inside.

Command Timeout - how long a single command can run before it gets killed. Default is 30 seconds. Bump this up if you're doing long installs or builds.

What the model can do

The model has access to these tools automatically - no need to ask it to use them, it figures out what's needed on its own.

Shell

Tool	Description
Execute	Run shell commands in a persistent session - `cd`, `export`, `source`, `nvm use`, `conda activate` all carry over between calls, just like a real terminal
ExecuteBackground	Start a long-running process in the background (servers, watchers, builds) and get a handle to check on it later
ReadProcessLogs	Read buffered stdout/stderr from a background process

Files

Tool	Description
WriteFile	Create or overwrite a complete file
ReadFile	Read a file, optionally a specific line range to keep context short
StrReplace	Replace an exact unique string in a file - the preferred way to edit code without rewriting the whole thing
InsertLines	Insert lines at a specific position in a file
ListDirectory	Browse the filesystem
UploadFile	Copy a file from your computer into the container
DownloadFile	Copy a file from the container to your computer

Container management

Tool	Preserves data	Speed	Use when
ComputerStatus	-	Instant	Check system info, resource usage, running processes
ResetShell	V	Instant	Shell env is broken, want clean variables without touching files
RestartComputer	V	Fast	Container is sluggish, runaway process, want a fresh start
RebuildComputer	X	Slow	Network setting changed, environment is fully broken

Examples

Run some code

"Write a Python script that generates a Fibonacci sequence up to 1000 and save it as fib.py"

Edit a file surgically

"Fix the bug in the handle_request function in server.py"

The model reads just that section of the file and replaces only the broken part instead of rewriting everything.

Set up an environment

"Set up a Node.js project with Express, create a basic REST API, and start the server"

Run a web server you can actually open Set Port Forwards to 8080:8080 in settings, then:

"Build a simple dashboard and serve it on port 8080"

Work with your files Attach a CSV and say:

"Analyse this data and give me a summary with the key trends"

Troubleshooting

"No container found" error on startup Your runtime isn't running. On Windows and Mac, make sure Docker Desktop is open. On Linux, check that docker or podman is installed.

Internet doesn't work inside the container Check that Internet Access is set to On in settings. If the container was already created before you toggled it, tell the model:

"Rebuild the computer"

It will recreate the container with internet enabled.

Container keeps failing on Linux Make sure you ran the setup commands from the requirements section. Then clear the old container and let it recreate:

Commands are timing out Increase Command Timeout in settings. The default 30 seconds isn't always enough for large package installs or slow builds.

Files disappeared after reopening LM Studio Check that Persistence Mode is set to Persistent. Ephemeral mode wipes the container on close - that's expected.

A few things worth knowing

The shell session is persistent within a conversation - cd, export, nvm use, conda activate all carry over between commands, just like a real terminal.

When something goes wrong, every tool returns a short error and an actionable hint telling the model exactly what to do next - no wasted tool calls investigating the failure.

The container is fully isolated and can't access anything on your host unless you configure a shared folder or port forward.

The disk limit is only hard-enforced on XFS filesystems. On ext4 (most Linux installs) usage is tracked and reported correctly but not enforced at the OS level.

Changing the base image only takes effect on a fresh container - tell the model to rebuild after switching.

License

MIT