Every article on local LLM MCP will walk you through editing a JSON file and toggling a server on. That's the easy part.
What they won't tell you is why your tool calls start failing halfway through a session, why your model ignores a perfectly valid tool, or why connecting 10 MCP servers makes your local LLM worse — not better.
This article covers what the tutorials skip. We're connecting LM Studio to MCP Express, and we're doing it in a way that actually holds up in practice.
What You're Actually Building
By the end of this, LM Studio will have direct, controlled access to your MCP Express tools.
Examples of what becomes possible:
- "Summarize the open GitHub issues assigned to me this sprint."
- "Pull the last 10 CloudWatch error logs and tell me what's failing."
- "Find the Notion doc on our API rate limits and give me the key points."
The copy-paste loop between your tools and your AI disappears.
Before You Start
You'll need:
- LM Studio
- MCP Express account — free tier, no credit card required
- An MCP Express server with at least one tool integrated
Refer to this documentation on how to create an MCP Express server. It barely takes 2-3 minutes.
Estimated time: 5 minutes.
Step 1: Create a Client in MCP Express
With an MCP Express server already prepared, click on the Get Started button in this section of Connect Your AI.
Choose LM Studio from the available list. Here, you'll find instructions along with a snippet as follows:
{
"mcpServers": {
"mcp-express": {
"url": "https://api.mcp-express.com/gateway/mcp"
}
}
}
Copy the snippet you receive and keep it handy.
Step 2: Navigate to LM Studio's mcp.json file
Open LM Studio. In the right sidebar, click this button.
Here, click on the +Install button below Integrations and choose Edit mcp.json.
Click it, paste the snippet from MCP Express, and hit Save. Your local LLM is now running with your MCP tools.
You'll see this dialog box. Click on Configure or you can click on the Authenticate button in the Authentication Required section.
Then, click on Trust and Open in Browser.
You'll then be redirected to your MCP Express app, where you can select the server with your tools to give LM Studio access to them.
Upon selecting, you can open LM Studio again and view that your MCP tools are now accessible to LM Studio's models.
Pick the Right Model
Not all local models handle tool calling equally. A model that's great at conversation or coding may completely fall apart when asked to select and invoke tools reliably.
For reliable tool use, you want a model explicitly trained for function calling — ideally at least 7–8B parameters. Thinking variants tend to handle tool calls more effectively. Models worth considering: the Qwen3 series, DeepSeek R1, and the Gemma3 series. In LM Studio, look for the 🔨 icon next to the model name — that's your confirmation it was trained with tool use in mind.
One more thing most articles miss entirely: context length configuration. LM Studio defaults to a 4,096-token context window for newly downloaded models. That's often too low for tool execution — you'll hit errors as soon as tool outputs start populating the context. Before you run anything, click the gear icon next to your model, increase the context length, and reload.
A practical starting point is 16,000–32,000 tokens. Don't max it out — there's a real reason to leave headroom, which we'll get to next.
What You Can Do With This Setup
Once LM Studio is connected to MCP Express, here's what works well with local models — and where the ceiling is.
Development workflows — GitHub issues, pull request summaries, commit lookups. Local models are solid at structured retrieval tasks that don't require heavy reasoning chains.
Internal knowledge access — Notion, Nuclino, Confluence. Pull up a doc, summarize a page, find a reference.
Ops and monitoring — CloudWatch logs, Kubernetes status, Supabase queries. Good for "what happened" questions that need real data, not model guesses.
Communication tools — Slack message drafting, channel lookups. Keep these simple — local models can struggle with multi-step communication workflows that require holding earlier context.
What works less well: long multi-tool chains where each step feeds into the next. Local models at 7–14B parameters don't have the reasoning depth to reliably orchestrate five sequential tool calls. Design your workflows with that ceiling in mind.
Try It Yourself
MCP Express connects your local LLM to the tools it needs — GitHub, Notion, Slack, CloudWatch, and more — without the overhead of managing individual servers.
Get started free → No credit card required.
Further Resources
- Documentation — See every supported integration, configuration option, and code example in one place.
- Contact Us — Got a question before signing up, or just want to talk through your setup? Drop us an email.
- Open a Support Ticket — Already inside the app and something's not working? Open a ticket directly from your dashboard and we'll get back to you as soon as possible.
