System requirements
- Windows 10 or Windows 11 on an x64 computer.
- At least 1 GB free for the application, extraction and temporary setup files, plus space for every GGUF model you download.
- Enough system memory for the model, quantisation and context size you choose. Loco's fit indicator provides guidance before download.
- An internet connection for the application and model downloads. Local inference itself does not require a cloud AI account.
Download and install Loco for Windows
Guided Windows installer
Run the downloaded installer and follow the setup wizard. You can install Loco on any drive (C:, D:, or external storage) and choose custom model directories.
Download Loco 2.0.2 Installer →The official Loco 2.0.2 installer is digitally code-signed by Marais Software. Confirm that you downloaded the file directly from this site.
Your first local conversation
- Check your hardware.Open Loco and review the CPU, memory and GPU profile shown in the interface.
- Choose a GGUF model.Open Models, select a curated option or provide a direct GGUF download URL.
- Wait for the download.Keep Loco open until the model is listed as installed.
- Start the runtime.Select the installed model and start the local model server from Runtime or the top bar.
- Open Chat.Send a prompt. Responses are generated by the model running on your PC.
Where Loco stores its files
The installed app uses Electron's per-user application-data directory. On a typical Windows installation this is under %APPDATA%\Loco. Inside it, Loco maintains:
models\- Downloaded GGUF model files.
data\- Settings, agents, users, conversations, workflows, API keys and credentials used by the app.
These folders are not automatically a secure vault. Protect your Windows account, disk and backups, particularly if you save credentials or confidential documents.
Manage local models
Choosing a model
Use model size, quantisation and Loco's hardware-fit guidance as a starting point. A model publisher's licence and acceptable-use terms still apply to that model.
Starting the runtime
Select one installed model, choose your context size and GPU-layer settings, then start the server. Loco exposes the model on 127.0.0.1:5319 while it is running. Stop the runtime before changing models or moving model files.
Custom model URLs
A direct URL must point to a GGUF file. Only download models from sources you trust, and verify the publisher and model licence independently.
Agents, documents and users
Create an agent
Choose an installed model, add a system prompt, set temperature and maximum tokens, and optionally attach supporting documents. Loco accepts PDF, TXT, Markdown and CSV files for this feature.
Control local access
Create staff users and assign only the agents they should use. When the LAN portal is enabled, its address follows the pattern http://YOUR-PC-IP:5318/chat.
Review conversations
The Audit area shows conversations created through the staff-agent experience. Treat this history as sensitive business data when backing up or sharing the data directory.
Workflows, credentials and API keys
Loco can create event-driven workflows, incoming webhooks and local API keys. Credentials stored for SMTP or other configured services are used only when the related action runs.
- Give each API key a clear name and remove keys no longer in use.
- Use separate, limited credentials for integrations wherever the service supports them.
- Assume that content sent to SMTP, webhooks or external workflow services leaves the local privacy boundary.
- Do not expose port 5318 directly to the public internet.
What is local and what uses the network
Local by default
- GGUF inference
- Desktop chat with the running model
- Models and application data
- Management and model endpoints on your PC
Network activity you initiate
- Loco, model and runtime downloads
- LAN portal connections
- SMTP and webhooks
- External workflow destinations
Read the full privacy explanation for both the desktop app and this website.
Local service endpoints
Management service
http://127.0.0.1:5318
Used by the desktop app for models, settings, agents, users, workflows and the local portal.
Model service
http://127.0.0.1:5319
Runs while a model is loaded and provides an OpenAI-compatible chat-completions route.
AI IDE Connectors
Connect Zed, OpenCode, Kilo Code, Roo Code, Continue, Aider & more.
Read the AI IDE Connectors guide →Chat-completions example
With the runtime running, this PowerShell request sends a short prompt to the local model:
$body = @{
messages = @(
@{ role = "user"; content = "Write a one-line welcome message." }
)
stream = $false
} | ConvertTo-Json -Depth 4
Invoke-RestMethod -Method Post `
-Uri "http://127.0.0.1:5319/v1/chat/completions" `
-ContentType "application/json" `
-Body $bodyThe endpoint is available only while the model runtime is active. Keep local services behind your workstation and trusted-network controls.
Back up or move your Loco data
- Stop the model runtime and close Loco.
- Copy the Loco per-user application-data folder to an encrypted backup location.
- Restore it only while Loco is closed and only into the matching Windows user's application-data location.
- Restart Loco and confirm models, agents and settings before deleting any older copy.
Model files can be large. Make sure the backup destination has enough capacity and is allowed to contain any confidential conversations, documents or credentials stored in Loco.
Common problems
The model will not start
Confirm the download completed, choose the installed model, reduce context size or GPU layers, and check Runtime logs. Close any process already using port 5319.
A model is too slow or runs out of memory
Stop the runtime and try a smaller or more heavily quantised model. Close memory-intensive applications and reduce the context size.
The download fails
Check your internet connection, free disk space and whether security software is blocking the source. Retry only from a model source you trust.
The LAN portal does not open
Confirm both devices are on the same trusted network, use the PC's current LAN IP address, and allow Loco through Windows Firewall for the appropriate network profile.
Port 5318 or 5319 is already in use
Close any older Loco or llama.cpp process, then restart Loco. A reboot can clear an orphaned local process if needed.
Windows shows Unknown publisher
Loco 2.0.2 is digitally code-signed by Marais Software. Download only from this website and verify the SHA-256 checksum before running it.
Verify release installer
In PowerShell, run Get-FileHash .\Loco-2.0.2-x64.exe -Algorithm SHA256 to verify file integrity.
Installer
02A80454635AB90DD768EE8B7C6210F4624585A5EEB4B377041488E477E51CD6