How to connect MCP servers to a local AI on your Mac
- MCP is the standard way to hand a model tools it didn't ship with.
- Outlier connects to MCP servers running on your machine and to hosted ones over Streamable HTTP.
- Hosted servers can authenticate through your browser, so you are not pasting long-lived tokens into a config file.
- The model stays local — MCP extends what it can do, not where it runs.
A model on its own is a very well-read colleague with no hands. MCP is how it gets hands: a common protocol for exposing tools, so every app doesn't need a bespoke integration for every service.
What MCP actually is
The Model Context Protocol is a shared contract between an application and a tool server. The server advertises what it can do; the app makes those capabilities available to the model; the model calls them.
Its value is combinatorial. Without a standard, every app integrates every service by hand. With one, any compliant server works with any compliant app — which is why support for it stopped being a differentiator and became table stakes.
Local servers vs hosted ones
Local (stdio) servers run as a process on your Mac. Good for filesystem access, local databases, anything already on the machine.
Hosted (Streamable HTTP) servers live on someone else's infrastructure — the service you're integrating with. These are the ones that need authentication, and the honest way to do that is a browser sign-in rather than a pasted token that never expires.
Outlier speaks both. Choose token auth or browser sign-in per connection when you add it.
Adding one
- Open the MCP section in Settings.
- Add the server — a local command, or a remote URL.
- Choose how it authenticates if it's remote: a token you hold, or sign-in through your browser.
- Use it in chat or in an agent run; the tools become available to the model.
Worth thinking about first
- A hosted server sees what you send it. Local inference keeps the model private; it does not make a remote tool private. That is a per-tool decision.
- Tools need a model that can call them. Tool-calling ability varies by tier.
- Approve what runs. Tools that touch your filesystem or shell deserve the same scrutiny as any other program.
Common questions
Can a local AI use MCP tools?
Yes. Outlier connects to MCP servers running locally over stdio and to hosted servers over Streamable HTTP, and exposes their tools to the model.
Do I need to paste an API token for a hosted MCP server?
Not necessarily — you can authenticate through your browser instead, so the app holds a scoped credential rather than a long-lived token pasted into a config file.
Does using MCP send my conversation to the cloud?
The model still runs locally. But a remote MCP server necessarily receives whatever the model sends it, so treat each hosted tool as its own privacy decision.
Which models can call MCP tools?
Tiers with tool-calling support. Core is the reliable choice for agentic work.
Try it on your own Mac
One signed, notarized download. No account, no token bill, and it keeps working with the Wi-Fi off.
Get Outlier for Mac