How to run Claude-Code-style agents offline on a Mac

Claude Code lives in your terminal. It reads your files, edits them, runs shell commands, and stops to ask before it does anything risky. None of that loop actually depends on a cloud. Give it a local model and the right tooling and you get the same workflow with the network unplugged. This is how.

The agent loop, decomposed

Strip a coding agent down and you find three moving pieces.

1. A model that emits structured tool calls. Read this file. Write this diff. Run this test.
2. A tool executor that actually runs those calls against your filesystem and shell, with approval gates in the right spots.
3. A conversation manager that feeds the tool results back to the model and keeps a turn-by-turn log.

The cloud version of Claude Code uses Claude Opus 4.7 for piece one and a slick terminal UI for the other two. The local version just swaps piece one for a model that runs on your Mac. The tool executor stays local. The whole thing lives in a Mac-native UI instead of a terminal.

What you need locally

Piece one needs a model that's been fine-tuned to emit structured tool calls. Outlier's Core 27B and Code 27B both do this. They're the strongest coding tier, which means the model holds its own on real refactors instead of folding on the first hard diff. Piece two wants a sandboxed shell, a filesystem driver, and approval gates. Piece three wants a project-aware UI that remembers your chat history. Outlier ships all of it in one app, sitting behind MCP (Model Context Protocol) so the tool surface stays portable.

Setting it up

1. Grab the Outlier DMG from outlier.host and install it.
2. First launch pulls down Nano 4B as the chat default. For coding you'll want Core 27B (any paid Pro tier). The first time you use it, that's a ~16 GB download.
3. Open a project folder. The app indexes it: filenames, source structure, and optionally the file contents as embeddings.
4. Flip to "Agent" mode. Now the model has tools. File read, file write, shell, project search, plus web search if you turned it on. The first time it reaches for any tool, you get an approval prompt.

It looks a lot like Claude Code. The agent lays out a plan, works through it step by step, and tells you what it did. The conversation shape carries over, the approval gates are where you'd expect, and the project context is right there too.

What's different in practice

Dimension	Claude Code (cloud)	Outlier Agent (local)
Model	Claude Opus 4.7	Core 27B / Code 27B / Plus 397B
Speed	~80–100 tok/s	~22 tok/s (Core), ~2.1 tok/s (Plus)
API key	Yes	None
Usage limit	Per-plan token cap	None — your hardware is the limit
Offline	No	Yes — works on a plane
Code leaves machine	Yes, to Anthropic	No
Cost after install	Subscription	$0 (Free) or $20/mo Pro for everything

Where Claude Code still wins

I'm not going to pretend it's a tie. The cloud version still beats the local one in a few places.

• Raw model strength. Claude Opus 4.7 is one of the strongest cloud-only models you can hit in 2026. Throw a novel research problem at it, or a 50k+ token context, and it still pulls ahead of Core 27B.
• Throughput. A 5,000-token diff lands in about 60s on Claude. On Core 27B locally it's closer to 4 minutes. Same output at the end. The wait is real.
• Tier swap is free. Claude never has to physically swap models between requests. Outlier does, and a cold-load on a tier change runs you about 74s.

Now the other side. In head-to-head testing against Claude Opus 4.7 in mid-2026, Outlier held its own on output quality, including the hardest cases (chess engine, raft/paxos, ZK proofs, needle-in-context, that tier of problem). For the 95% of coding work where the output matters more than how fast it arrives, the local version basically catches up.

Frequently asked questions