Local LLMs with LM Studio#

LM Studio can run a local model server that Sero treats like an OpenAI-compatible provider. Use this when you want to test local models, work with private local endpoints, or avoid sending a specific task to a hosted model provider.

Local models still need enough memory and GPU/CPU capacity on your machine. They may be slower or less capable than hosted models, and they are not automatically available inside every workspace container unless the URL is reachable from the process that needs it.

Quick path#

1. In LM Studio, download a chat/instruct model.
2. Load the model.
3. Start LM Studio's local server with OpenAI-compatible API enabled.
4. In Sero, open Settings/Admin → Models → Local models.
5. Add a provider and choose the LM Studio preset.
6. Confirm the base URL is http://localhost:1234/v1.
7. Use API key lm-studio or none unless your server requires a specific key.
8. Test the connection, fetch models, save the provider, then assign LOW/MED/HIGH tiers.

Set up LM Studio#

In LM Studio:

1. Install or open LM Studio.
2. Download a model that supports chat.
3. Load the model into the runtime.
4. Open the local server panel.
5. Start the server.
6. Keep the server running while Sero uses it.

The Sero preset expects:

Field	Value
Provider name	lm-studio
Base URL	http://localhost:1234/v1
API shape	openai-completions
API key	lm-studio
Compatibility	developer role off, reasoning effort off

If your LM Studio server uses a different port, update the base URL before testing.

Add the provider in Sero#

1. Open the model management surface from Settings/Admin or the model manager.
2. Open Local models.
3. Click Add Local Provider.
4. Choose LM Studio in Quick Setup.
5. Click Test connection.
6. Click Fetch from server to import model IDs from LM Studio's /models endpoint.
7. Save the provider.

Sero writes local provider configuration to <SERO_HOME>/agent/models.json and refreshes model availability after saving.

Assign tiers#

After saving the provider, open Admin → Model and choose LOW, MED, and HIGH defaults.

A practical local setup is:

Tier	Suggested use
LOW	Small/fast local model for quick edits or summaries
MED	Stronger local model for everyday development work
HIGH	Best local model you can run comfortably, or a hosted fallback

Thinking levels only appear when Sero believes the selected model supports them. Many OpenAI-compatible local servers do not support reasoning-effort controls, so the LM Studio preset disables that compatibility flag.

Host and container reachability#

localhost means “this process's machine or network namespace.” That is usually fine for Sero desktop talking to LM Studio on your Mac. If a tool inside a workspace container must call the same local server directly, localhost from inside the container may point at the container, not the host.

If a containerized command cannot reach LM Studio:

• use a host-reachable URL instead of localhost when your setup supports it
• verify the LM Studio server binds to an address reachable from the container
• test from the same place that will make the request
• avoid exposing the server on untrusted networks

For most model selection and chat usage, configure the provider through Sero and let Sero manage model calls from the desktop process.

Troubleshooting#

Problem	What to check
Test connection fails	LM Studio server is running, base URL includes /v1, port is correct, firewall is not blocking loopback.
Fetch returns no models	A model is loaded in LM Studio and the server's /models endpoint returns data.
HTTP 404	Base URL may be missing /v1 or using the wrong port.
Auth error	Use lm-studio or none unless your server requires a real API key.
Model appears but tier save warns	The model may have been unloaded, renamed, or removed from the local server. Fetch again or choose another model.
Container tool cannot reach the server	See the host/container reachability caveat above.

Related docs#

• Models and Providers
• models.json Reference
• State and Folders