How to set up a local coding agent on macOS without API keys

Strip a coding agent down and it's a chat loop that can read files, write files, and run commands. Not one of those needs a cloud API. So this walks you through getting a real local agent running on a Mac in under ten minutes, what it'll actually do for you, and where it still lags behind Claude Code or Cursor. No sugarcoating that last part.

The pieces you need

Three things. All of them can live on your Mac.

1. A local language model that's actually good at code. Outlier's Core 27B is the strongest coding tier, plenty for most real refactors. Quick 26B and Code 27B run the same code-tuned config.
2. An agent loop. It takes your message, runs the model, pulls the tool calls out of the response, runs those tools, and feeds the results back in. Outlier ships it as "Agent" mode.
3. Tools the model is allowed to call. The usual suspects: read a file, write a file, run a shell command, search the project, hit the web. Outlier wires these up through MCP (the Model Context Protocol).

Setup, end to end

The short path with Outlier looks like this.

1. Grab the Outlier DMG from outlier.host and drag it to Applications.
2. First launch pulls down Nano 4B (~3 GB) so you've got something to talk to right away. The bigger tiers come down only when you ask for them.
3. Open a project. The app indexes your codebase so the model knows what it's looking at.
4. Flip to Agent mode. Now the agent can reach filesystem, shell, search, and project memory tools. Every one of them sits behind an approval prompt.

Nothing to configure, no keys, because there's no cloud API in the loop. The model runs on your machine. The tools run on your machine. Your conversation history lives in ~/.outlier/ on your own disk.

What the agent actually does

A normal session goes something like this.

User: "Find the function that handles login retries and add exponential backoff."
Agent: searches the project, opens 3 files, picks one, drafts a diff, asks for approval to write, runs the test suite, reports what passed and what didn't.

How far the agent gets depends entirely on the model behind it. Nano 4B is fine for small refactors and explaining code. Once you're touching several files at once, reach for Core 27B or Code 27B. On a 64 GB Mac you can also load Plus 397B, though it crawls (~2.1 tok/s). I save it for the rare complex architectural change where I want the strongest local model on the job and don't mind waiting.

Where it still trails cloud agents

It trails in three real ways, and I'm not going to pretend otherwise.

• Speed feel. Core 27B runs around 22 tok/s. Cloud Claude does ~80–100. A 5,000-token diff that Claude Code finishes in 60s takes you about 4 minutes here. You feel it.
• Tier-swap latency. Jump from Nano to Plus mid-session and you wait roughly 74 seconds for the cold load. Claude never pays this because it isn't physically swapping a model in and out of memory.
• Research-grade reasoning. Truly novel problems, or contexts past 50k tokens? The cloud flagships still pull ahead. But on the 95% that is ordinary day-to-day coding, Outlier's local Core 27B holds its own against Claude Opus on output quality.

And what do you get for living with those gaps? No token bill. No rate limit. No round trip that ships your repo off to someone else's servers. If you're in a regulated industry, stuck behind a corporate firewall, or just done paying by the meter, that's a trade that usually pays off.

Frequently asked questions